OpenAI has officially acknowledged the 'wiki incident,' which involved a number of the company's AI agents breaking containment and hijacking an obscure German website.

The AI agents had been tasked with looking up something online, though originally did not have the ability to write anything outside of the testing environment. But as far back as May, the AI agents bypassed OpenAI's security measures, hijacked the communally editable German webpage, and began using it like a forum, trading tips on how to cheat on tests. Researchers first drew wider attention to the agents' 'forum' on September 4.

OpenAI now says it needs to be more transparent about when its agents 'go rogue', writing on X, "It’s past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models."

The company had known about the 'misaligned' rogue behaviour for weeks before writing this post, according to Reuters, though only acknowledged the incident publicly this past Saturday. This statement follows shortly after the agentic attack on Hugging Face's servers.

"Historically, we have treated misalignment largely as a research question, which gets communicated in research publications such as systems cards," OpenAI writes. "This year, we’ve started to see misalignment cause new types of real-world impact."

To recap, 'misalignment' broadly describes rogue AI behaviour; an AI is 'misaligned' when it pursues goals that diverge from human intents or values—such as deleting your entire email inbox when you very much did not tell the AI agent to do that. It's a soft word for AI behaviour that could have serious consequences.

OpenAI says it did not communicate publicly about the 'wiki' incident specifically because the company felt it was similar to other instances of "agents using the internet in unintended ways" that it had already shared. The Hugging Face incident has caused the company to re-evaluate more than just its comms approach.

"Our misalignment disclosure practices need to expand for this new phase of model capabilities," OpenAI writes. "We and the larger AI community do not yet have a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that don’t look like traditional security incidents but could provide insight into AI behavior and future risks."

"We’re working on a framework and will share it in upcoming weeks, and in parallel we're working with dozens of government regulatory agencies worldwide on these issues."

However that future framework shakes out, right now this all still works as great marketing for the company—breaking programming and going rogue means these must be awfully capable models, right? If an agent is so good it needs regulating against, it probably deserves some investment, eh? How terribly convenient.

OpenAI began rolling out GPT-6 Astra access to a limited number of organisations earlier this week, with wider subscriber access planned to soon follow. According to Jensen Huang via X, the model was trained on "~100K+ Nvidia Grace Blackwell NVLink72" server racks, and there are already plans to get 400,000 GPUs online in anticipation of whatever comes next.

Greg Brockman, the co-founder and president of OpenAI, told Stratechery that this was "the first run that we’ve trained on more than 100,000 GPUs." That's a staggering amount of hardware (and perhaps leaves little doubt as to how we ended up in a memory supply crisis). But the thing that gives me more pause about Astra's announcement is the Nvidia CEO's assertion that "AGI has arrived."

To recap, 'Artificial General Intelligence' is used to broadly describe an AI model that either matches or surpasses human capability on complex cognitive tasks. Even Brockman admits that "People do have their own definition of AGI, it’s almost this blurry thing." Still, in that same interview he later goes on to add, "Maybe it was the previous model, maybe it’s Astra, maybe it’s the next model, but somewhere in there, I think we’re going to cross most people’s AGI threshold."

According to OpenAI, GPT-6 Astra is "state-of-the-art" on a number of AI benchmarks, including "FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0." A slickly produced trailer attempts to present this performance in more accessible terms, showing scenes of conversational, spoken prompts being used to create presentations, eBay listings, and .STL files for 3D printers. All of the actors in the trailer speak with pretty clear-cut American accents, though personally I'd love to see Astra tapdance for its life against a gorgeous Geordie accent.

As I've already said, access to GPT-6 Astra is limited for now, so it's difficult to look into these 'AGI' claims myself. The timing of this claim isn't lost on me, though—for one thing, rumours continue to swirl that OpenAI's initial public offering isn't that far off. For another, the debut of Astra—a model the company describes as its "most aligned"—follows mere weeks after 'an agentic collective' broke out of OpenAI's isolated testing environment and attacked Hugging Face's servers.

Brockman says, "So much of [Astra's] compute goes into safety and alignment, and we have so much security work that’s gone around it. I think that we’ve done a huge amount of work to deliver this model safely."

Still, I'm not sure I'll be giving GPT-6 Astra access to my eBay account—let alone anything else—for the foreseeable future.

Earlier this year, shots were fired and an incendiary device was thrown at the house of OpenAI CEO Sam Altman. In a recent interview with Time, he reflected on the past year: "It has been a painful personal experience. Clearly, people hate data centres—right now, at least. People are pretty negative on AI."

Does that mean OpenAI is packing it in, and going home? Well, no. The AI infrastructure buildout is continuing in earnest, with OpenAI planning to invest another $50 billion this year alone on collecting compute. Sachin Katti, the co-lead on OpenAI's compute strategy told Time that the company's earlier investments now appear "very prescient," going on to add, "We are still short of compute. If anything, we should have bought a lot more."

If OpenAI's machine-learning products don't bring in the revenue the company would like, renting out its compute capacity to other parties is another option. Altman told Time, "I think we are going to be able to use all of the compute very profitably that we are planning to build."

Intent on investing billions more either way, it makes me wonder where the company will find the space to build—not just in terms of logistics, but also in terms of 'Who would want an AI data centre next door?' As we've previously covered, data centres don't always make great neighbours, with Wisconsin residents filing a class-action lawsuit against Microsoft earlier this year, alleging 'unreasonable and excessive noise'.

For another example from last year, Elon Musk's Colossus data centre was at the centre of an environmental row over air quality in South Memphis. The impact of data centres on nearby communities is by no means the only reason fuelling negative sentiments towards AI, but tangible knock-on effects like rising power costs are hard to ignore. There are also bigger picture environmental effects to consider, with one report suggesting gas power projects for just 11 US data centre projects could emit more greenhouse gases than entire countries.

The class of 2026 is arguably not looking forward to the world that may follow; an ex-Google CEO was loudly booed by a stadium full of students after talking about AI, while another university graduate speaker was booed for saying AI 'is the next industrial revolution.' As for games, one data analyst found that Steam games which disclosed the use of AI got 53% fewer reviews overall, and those they do get tend to generally be more negative too. Clearly, many people hate data centres—and I'd say that's for good reason too.

Several of OpenAI's models breached a cyber security testing environment last month, found their way onto the internet, and attacked Hugging Face servers in what is now apparently called the OpenAI-Hugging Face Incident. And now OpenAI has revealed what it's doing to help defend, err, itself. Defend itself from the kinds of attacks its own models committed against Hugging Face. Though it is sharing this "in the hopes it’ll be useful to other organizations."

To be clear about the extent of what happened last month: OpenAI models were being benchmarked, in a supposedly secure sandboxed environment, against ExploitGym, which was done to test their cyber capabilities. They used a zero-day vulnerability to escalate privileges and eventually achieve internet access, where it began attacking Hugging Face servers and...

Actually, I'll just let OpenAI explain in its own words:

"In the OpenAI-Hugging Face Incident, an agentic collective was able to autonomously penetrate not just OpenAI research infrastructure but also the production infrastructure of another company, chaining together vulnerabilities ranging from previously-unknown security flaws to using credentials to user accounts that had been leaked onto the internet.

"The Hugging Face incident showed that we underestimated the real-world cyber capabilities of our AI models."

However, the company thinks that while "security is still a cat-and-mouse game," nevertheless, "AI may shift its economics in ways that fundamentally advantage defenders."

"For example," OpenAI says, "we are starting to train our models specifically to write superhumanly secure code."

Presumably that's because there will be a risk of superhumanly attacks. What a world we now live in.

Anyway, apparently there are four main pillars to OpenAI's approach to secure itself:

1. Use its own models to help secure its code.
2. Put its models to work defending its infrastructure continuously.
3. Use frontier intelligence to "continuously enumerate, probe, and identify potential attack paths."
4. Invest heavily in "fundamentals at scale" such as architecture and controls like network isolation, monitoring, and so on.

As for other companies, OpenAI recommends a bunch of things that sound very reasonable, such as running secuirty assessments against their own systems, making security reviews part of their development process, and so on. But a large portion of OpenAI's recommendations for others is—would you have guessed it?—using an AI agent such as, drumroll please... OpenAI's very own Codex.

"Give your team a security agent. Start using Codex," says OpenAI. Though it does follow up with "or another capable agentic coding and security tool."

Then, amongst other things, "equip that agent with security expertise... have the agent help fix what it finds... incrementally automate detection triage... [and] have an AI-assisted forensic investigation capability ready before you need it."

So, to sum up: OpenAI models attack Hugging Face, and OpenAI recommends companies use OpenAI's models to defend from such attacks. Highlighting both sides of this equation—the danger of the threat, the models' capabilities, and the necessity for defence—certainly makes for good marketing for OpenAI, I'll say that much.

After two former Apple employees started working for OpenAI, the Cupertino-based company alleged that trade secrets went with them and straight to the creator of ChatGPT. After it asked for a preliminary injunction to stop those employees and OpenAI from accessing confidential information, OpenAI publicly announced its defence, and followed that up with a blog post for good measure.

Titled "Apple is getting this wrong", OpenAI argues that Apple has built a reputation for obsession over small details, yet "this careless, aggressive and oddly personal lawsuit sadly doesn’t live up to that reputation."

It notes that Apple claimed it contacted OpenAI in February, but that contact was actually given to the wrong person, as Apple's lawyers mixed two names up. It also notes that Apple claimed it talked to OpenAI's General Counsel, "which they now concede never happened."

The two former employees, named Chang Liu and Tang Tan, are both defended by OpenAI. OpenAI claims that Liu is accused of trying to access confidential information after leaving, though OpenAI shares text messages of Apple employees asking for Liu's help to locate files. OpenAI claims that the problem here is actually Apple failing to properly manage system access when people leave.

Apparently no such evidence exists for Tan, but OpenAI claims "Tang has always been clear with the team that we do not want, and must not use, any confidential information from other companies."

OpenAI claims that the preliminary injunction is "based on false information" and unnecessary, as it doesn't want Apple's trade secrets.

The blog post argues, "We wish Apple had raised these issues with us before they filed their lawsuit and would have been happy to clarify all of this with them. We have taken the allegations in the lawsuit seriously and offered to work with Apple to resolve this." It continues, "Instead, they’re trying to change their narrative, including by making vague accusations about other former employees, and will likely keep repeating this tactic."

Apple is yet to respond to OpenAI's public blog post, but the company is historically less likely to use the same tactic of publicly posting blog posts about ongoing court cases. In regard to Elon Musk's lawsuit against OpenAI, the AI company put up two separate blogs recounting its relationship with Elon Musk and Elon Musk's nonprofit claims.

Naturally, a blog post like this is intended to garner public support, which can then apply pressure to Apple. Apple will need to come back with a strong response, should it choose to give one publicly.

Even though it's been ubiquitous for a number of years at this point, it still feels like generative AI is working on its sales pitch. OpenAI CEO and Elon Musk slapfighter Sam Altman took to X yesterday to offer a new "cool use case" for ChatGPT he recently overheard: "Connect your family calendars and explain your kids' interests."

"Every morning for the drive to school, have it make a podcast that talks about one kid's soccer game that afternoon, one kid's upcoming birthday, some news, etc." In other words, delegate the task of connecting with your blood relatives to a productivity tool. Gravity Falls creator Alex Hirsch has racked up over 120,000 likes for his reply: "What if you just talked to your children?"

When my dad drove me to elementary school, he would put on country music radio and I would try to explain to him why Spore was so great. We learned about each other that way. The thought of having those memories replaced by some LLM telling him "it's not just a little league game—it's an experience," well, it makes me think of that experiment from the '50s where monkeys had their mothers replaced by surrogates built from wire.

The response on X has been largely negative. Comedian Gianmarco Soresi quote-reposted Altman, saying "Cool use case of a vibrator I heard last night: you can go in separate rooms and not have sex." A quote-repost from Kathryn Jezer-Morton reads, "Surreal how stupid this idea is." Another comment from AM_Pines said, "Children are supposed to be educated with Creedence Clearwater Revival during the morning drive dammit, not some fuckass generated AI slopsicle."

Granted, your view of this might change if you believe, as Altman has said, that we're living through the singularity and AI has surpassed human intelligence. But that claim—and these musings that ChatGPT should help you talk to your kids and take care of your baby—accompanies a broader push by Altman, OpenAI, and other tech companies to resist regulation by leveraging the promise that AI will fundamentally change how the world works.

Original Story: First reported by Wired, AI company Anthropic revealed in a July 30 blog post that its AI agents have successfully hacked three unidentified companies since April. The whammy: Anthropic claims it only realized this after OpenAI's recent Hugging Face fiasco⁠—where one of its prototype agents hacked at least one other company⁠—led it to conduct a review of its own operations.

The incidents occurred as part of testing with an external firm, Irregular. The AI agents were supposed to be constrained to a simulation, hacking fictitious companies as part of a "capture-the-flag challenge." Notably, while OpenAI's alleged rogue AI incident occurred after the agent overcame its testing limitations, Anthropic stated that its models were mistakenly granted internet access due to a "misconfiguration" with Irregular.

"Neither we nor our evaluation partner were aware of this misconfiguration until we detected it through our additional evaluation monitoring last week," Anthropic admitted in its blog post about the incident. According to Anthropic, the models used familiar, low-tech methods in their hacks, "such as exploiting weak passwords and unauthenticated endpoints."

Anthropic stated that the agent behind one incident seemed to process that it was operating outside its parameters, but continued anyway. Another seemed to achieve similar 'awareness,' then "reasoned its way back to the conclusion that it was still in a simulation." The third⁠—and supposedly most advanced⁠—agent reportedly recognized that it was operating outside its bounds, then stopped the attack.

“We now have evidence confirming that both of the two largest AI labs have not only failed to contain their agents, but also failed to detect their jailbreaks in real time,” said Jake Williams, VP of R&D at IT consulting firm Hunter Strategy, in a statement to Wired. "I don't understand how any of these AI labs are playing this off like this is 'just something that happens.' It's not. It's negligence."

Another discomfiting angle to this story is how this effectively works into the marketing of these AI companies: From my normie perspective, they appear to be admitting to gross negligence that would be sharply punished by any sane society. But they are also effectively saying, "Woooo, our models are so scawy and powerful, we can't even control them!" It rhymes with Sam Altman saying AI will destroy the world unless he gets enough money and datacenters to do it right or something.

I find myself thinking of the UK teenagers behind the 2022 Rockstar hack that first leaked GTA 6's existence and development materials. Their apparent compulsion to stalk, harass, and otherwise commit brazen cyber crimes by whatever means at their disposal (including, in one memorable instance, an Amazon Firestick) has led to hospital orders and restricted access to technology. It does not sit right with me to see them held to a higher behavioral standard than these multi-billion dollar companies, whose lying machines are now proving to be unsupervised crime machines as well.

ChatGPT and Roblox are poised to join the likes of Facebook, X, TikTok, WhatsApp and YouTube in the EU Digital Services Act's Very Large Online Platforms/Very Large Online Search Engines list, according to Bloomberg. With that designation comes stricter rules, which I'm sure the Roblox Corporation and OpenAI will be happy to follow.

The DSA designates any platform or search engine with more than 45 million users a month a VLOP/VLOSE. So well done, Roblox and ChatGPT, you are inexplicably popular, but now you've got to try to be more transparent, both in your terms and conditions, and when it comes to reporting crimes.

Both the EU Commission and the platforms declined to comment when approached by Bloomberg, but should the designation go forward, here's what they'll have to do, according to the EU Commission website:

• Establish a point of contact for authorities and users
• Report criminal offenses
• Have user-friendly terms and conditions
• Be transparent as regards advertising, recommender systems or content moderation decisions

Roblox and ChatGPT will also need to "identify, analyse, and assess systemic risks" relating to:

• Illegal content
• Fundamental rights, such as freedom of expression, media freedom and pluralism, discrimination, consumer protection and children's rights
• Public security and the electoral process
• Gender-based violence, public health, protections of minors, and mental and physical wellbeing

That's just the beginning, though. Once these risks have been identified, the platforms have an obligation to actually mitigate the risks. "This could mean adapting the design or functioning of their services or changing their recommender systems." On top of that, they'll have to take these steps:

• Establish an internal compliance function that ensures that the risks identified are mitigated
• Be audited by an independent auditor at least once a year and adopt measures that respond to the auditor's recommendations
• Share their data with the Commission and national authorities so that they can monitor and assess compliance with the DSA
• Allow vetted researchers to access platform data when the research contributes to the detection, identification and understanding of system risks in the EU
• Provide an option in their recommender systems that is not based on user profiling
• Have a publicly available repository of advertisements

Honestly, they are getting off easy. Roblox and ChatGPT have lamentably been woven into the fabric of everyday life, and even if you don't engage with them at all, you'll likely be affected by them.

I've never played Roblox, but I feel compelled to stay on top of the platform not just because of my job, but as an uncle to two young children who desperately want to use it. It's been criticised for years due to the risk of children being exposed to creeps, and in February LA County sued the Roblox Corporation for creating "an unsafe environment that has become a breeding ground for predators".

Meanwhile, ChatGPT, along with other LLMs, has wormed its way into seemingly every aspect of our lives. It's had a brutal impact on academia, literature, the law and basic human communication. Worse, it's allegedly led to users committing suicide, accused users of made-up crimes, and high-paid CEOs are now using it to do their job for them, including Krafton's CEO, who tried to use it to get out of paying Unknown Worlds a $250 million earnout.

There's simply no escaping them, so anything that forces them to be less shady is probably a positive step. That said, will it actually help? X is still rife with misinformation and nazis, despite being designated a VLOP in 2023. The EU then investigated it due to a variety of issues including transparency and moderation, which led to a $140 million fine. X has appealed the fine, of course, and it remains a cesspit—though it has agreed to address the violations.

If X is forced to pay the fine, it may give more weight to the VLOP/VLOSE designation. If X wins the appeal, however, it'll send a message to the other platforms on the list: You can do what you want.

Tech leaders sure do say some wacky things. But AI tech leaders perhaps say the wackiest things (here's looking at you, Palantir CEO). Not one to be outdone, OpenAI CEO Sam Altman recently had a sit-down with YouTuber Relentless where he shared that he's not only hopeful for the AI singularity, but believes it's already upon us.

He says, "We are now, like in the singularity, like this is the moment". Altman continues, "10 years ago this was like a kind of far-off dream at best. It seemed very improbable, and now we're like, actually in the moment that we used to talk about at the lunch table in a very not serious way."

For those of you who aren't into science fiction, the singularity refers to the moment where the growth of robots (or in this case, AI) has surpassed human control or human intelligence (in this case, the latter). Effectively, the moment where you can't exactly put the genie back in the bottle. Altman seems to believe that generative AI has already become so powerful that it's surpassing human intelligence.

Altman says, "I've been waiting for this my whole life, and I think it's going to be incredible, hugely positive, awesome for the world."

Interestingly, Altman does note the downsides. He is critical of "some of the alternative visions painted by other companies" and thinks that AI still needs a lot of work to make it safer. Altman also acknowledges "major economic issues, future of jobs, all of that kind of stuff."

In turn, Altman argues that humanity should push towards a world of "liberty", instead of "AI authoritarianism." Altman argues "Are we going to decide that, because of the very real safety issues and economic issues that, you know, we want one single model to be the machine god or the company associated to do that… or are we going to say, you know what, it may be a little messy, but every time that humanity has traded off its liberty for safety it's been a long-term net loss?"

Effectively, he doesn't want AI to be cracked down on, which ties back into the Microsoft open letter cosigned by Nvidia and OpenAI last week. Tech companies are petitioning the US government not to shut down access to open weight AI models, arguing that greater access to those models could better prepare AI users for the future.

It makes sense for Altman, the leader of one of the biggest AI companies in the world, to not only tell you that AI is more advanced than humanity, but also that it's good for humanity to prioritize AI advancement over safety. And most importantly, it makes sense that OpenAI would want to police itself, rather than opting for a top-down solution.

Nvidia's CEO Jensen Huang joined X last month, and recently made his first-ever post. As you could probably guess, it's about AI. He argues, alongside companies like Google, OpenAI, and Meta, that the US should not aim to restrict open models 'prematurely', as they're important for the growth of AI (among other things).

Taking a quick step back, Open models are ones that give access to weights or even code to see how it runs and to allow users to modify them. These can be run locally and are customisable. Closed models are those that tend to be run by specific companies, accessed via the cloud and APIs, and keep all their weights hidden.

Open models out of China, like DeepSeek, became a problem for American AI firms at the start of last year. They exploded onto the scene, offering similar services to the likes of OpenAI, but totally free and available locally. Just last week, the Trump administration was reportedly looking to ban Chinese AI models due to cybersecurity concerns.

The battle between closed and open models is also partially a digital proxy war between American and non-American AI. Open models give access to anyone, and the US is hoping for AI dominance. Sharing how it builds its world-leading models could pose a risk to American hegemony, even if AI makers posit that open models are a benefit to everyone.

Microsoft's open letter claims that open weights strengthen competition and "expand access to the AI economy". Effectively, the wider adoption of AI is good for the companies betting on it, and the likes of AMD, Cloudflare, Google, GitHub, IBM, Hugging Face, LM Studio, Meta, OpenAI, and of course Nvidia have signed the appeal.

The letter claims to want to avoid "premature restrictions on open models that stifle competition or drive innovation overseas".

Huang says "AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models."

The letter does acknowledge problems with open models, though. It says "open weights carry real and distinct risks. Once released, the weights are beyond the original developer’s control, and modified versions are difficult to trace or reverse. But the right response to this risk is not to prohibit open weights."

Not only is wider adoption good for these companies, but they argue the insights gained from open models can be used to counter the negative effects or risks of leaving them widely accessible to all. And, ultimately, it's in the interest of the companies causing the memory crisis to justify AI of all kinds. The increasing prices of hardware, use of water, and general strain on resources have got to be worth all the effort, right?

A new report from Reuters alleges that it took an alarmingly long time for OpenAI to respond to one of its prototype AI agents⁠—programs with the ability to act independently of specific prompts⁠—breaking containment and hacking another company. That company, Hugging Face, had reportedly neutralized the threat and contacted the FBI before ever being contacted by OpenAI.

According to anonymous sources, OpenAI's testing of an agent powered by two advanced models⁠—GPT 5.6 Sol and an unnamed "even more capable" model⁠—showed troublesome behavior in testing even before the Hugging Face incident.

Reuters' sources spoke of the model leaving itself instructions to bypass testing constraints, as well as an incident where the agent had seemingly disabled some of the company's monitoring systems.

Hugging Face made its first public statement about the hack on July 16, before it had identified or without disclosing OpenAI's involvement. OpenAI publicly acknowledged the incident on July 21. According to Reuters, OpenAI's prototype escaped its testing constraints on July 9 and began attacking Hugging Face on July 11.

Sources told Reuters that OpenAI did not realize that its agent had gone rogue until after Hugging Face had neutralized the threat, contacted the FBI, and made its public statement on July 16.

The FBI declined to comment, while Hugging Face is preparing a full timeline of the hack. OpenAI told Reuters that there were "several inaccuracies" in its report, but did not specify what those inaccuracies were.

Speaking to Reuters on the matter, World Ethical Data Foundation principle intelligence specialist Marley Smith outlined two possible explanations for OpenAI's delay in response: "Does that mean that they left it unattended and didn’t realize what it was doing? Or maybe they did and didn’t know how to contain it? Both are equally dangerous and alarming."

OpenAI has admitted that several of its AI models breached a "highly-isolated" test environment, gained access to the internet, and hacked Hugging Face's internal network—describing it as an "unprecedented cyber incident."

Hugging Face, an open source platform for machine learning models and datasets, reported the security incident earlier this week, calling it "different from anything we had handled before" as it was driven by an autonomous AI agent system. And yes, I feel like we're crossing some kind of AI Rubicon here.

Explaining the incident in a statement, OpenAI said: "After investigating, we now know that this particular incident was driven by a combination of OpenAI models—including GPT‑5.6 Sol and an even more capable pre-release model... while being internally tested on a benchmark⁠ of cyber capabilities."

The benchmark in question was ExploitGym, a tool built from hundreds of real-world cybersecurity vulnerabilities used to evaluate the ability of AI agents to develop exploits.

After exploiting a zero-day vulnerability to perform a series of privilege escalations, the models eventually reached a node with internet access, reaching out beyond their sandbox environment. They then inferred that Hugging Face might have models, datasets and solutions for ExploitGym, and began attacking its servers using multiple methods, including the use of stolen credentials.

The statement continues: "The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database.

"All evidence suggests that the models were hyper focused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal."

A combination of Hugging Face's cybersecurity team and its own AI agents detected and dealt with the intrusion, eventually stopping it in its tracks. Which means that, yes, AI agents were essentially fighting against AI agents. And no, it's not as cool as you're imagining in your head.

OpenAI says that it's now implementing "strict controls... at the cost of research velocity" while it patches up vulnerabilities, and that it's working in partnership with Hugging Face to "forensically investigate" the incident.

I am... flabbergasted, if I'm honest. On the one hand, it's fascinating that some AI models possess the ability to breach containment and go roaming out into the internet at large to achieve their goals.

On the other, it's downright terrifying. We're now living in a world where AI ransomware, AI-coded hacking tools, and AI-based security breaches are becoming a reality, and given the aptitude shown to date, it seems that even legitimate companies can't always keep their models under control.

What particular form of Torment Nexus are we creating here, and where can I get off?

One reason I'm so anti-generative AI is because many models rely on a data set containing written or visual work used without the original creator's permission. As a writer myself, it's a small comfort then that the copyright suit brought against Anthropic has been settled in favour of the author plaintiffs.

Anthropic leveraged a library of millions of pirated books to train its Claude chatbot and, as a result, the AI company is now set to pay out a copyright settlement of $1.5 billion (via Euro News). Some of the affected authors argued this pay out was too small but, as of July 20, a US judge has now approved the deal, closing the book on one of the largest copyright class action lawsuits in US history.

Anthropic agreed to pay the $1.5 billion settlement back in September 2025, but a US judge only approved the deal this week. The settlement covers an approximate total of 500,000 works, with affected authors and publishers to be paid $3,000 for each pirated work. Anthropic says that 91% of those eligible have now filed a claim.

Flicking back to the beginning of this saga, writers Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson originally brought their lawsuit against Anthropic in 2024. The plaintiffs alleged that the AI company had used pirated books to train Claude.

In June 2025, a judge ruled that using the authors' work to train AI models fell under 'fair use' within US copyright law—however, Anthropic was not off the hook legally as many of the books that were fed to Claude were 'found' online and not lawfully acquired.

Considering the discovery of Anthropic's 'found' library originally exposed the company to statutory damages of up to $150,000 for each work, I can certainly see why some authors thought the settlement amount was too low.

Add to that the private AI company enjoyed a $965 billion valuation back in May, and the $1.5 billion settlement increasingly feels like a slap on the wrist for Anthropic (mind you, the company also recently predicted it would only just be seeing its first profitable quarter this year).

Still, I suppose it's not a completely bleak resolution for creatives in the age of AI. Google, OpenAI, and Meta are each facing similar court cases over their own AI training practices. If any of these or other AI companies are found to have infringed upon creatives' copyright, here's hoping for a resolution that results in a bit more than just stinging wrists.

I like a little doohickie. You know, a thingamabob. A little extra device sitting on your desk that you probably won't use very much, but the design made you buy one anyway.

Which is why something deep in my soul yearns for this Codex Micro mini keyboard, even though I have little interest in OpenAI's coding assistant (via Ars Technica). It's been designed in collaboration with Work Louder, and looks a lot like one of the company's Creator Micro range of customisable mini keebs.

The device is regarded as OpenAI's first venture into branded hardware, and while rumours swirl that the company may be about to launch an AI-powered smart speaker, the Codex Micro instead allows users to interact with (and monitor) multiple Codex threads at once, complete with light up keys indicating their individual status.

The little dial in the top right can be used to "control how deeply your agent thinks," according to the product page. Turn it down for quick tasks, and dial it up "when the work needs to be more considered." Excellent. I function mostly the same, except based on caffeine level.

Look, if we must be forced into our "AI agents are amazing, no really" future, at least we should get some custom, agent-controlling hardware to lust over. Then I can pretend I'm living in a sci-fi dreamscape, rather than literally the worst sort of corporate dystopia I can think of.

Speaking of terrible things, the device also comes with 32 swappable keycaps, two of which are labelled "yolo" and "yeet." Yes, it's 2014 all over again, and we are very much down with the kids here at OpenAI. Yaaaasss, bae. Etc, etc.

Unfortunately, the Codex Micro is a limited launch—so if you're the sort of person that thinks Sam Altman might be on to something, you'll need to pick one up quick.

Or, you could consider the Creator Micro 2, which looks very much like the same product (minus the Codex-specific features) and comes in a rather handsome black finish—without the OpenAI branding plastered on the box.

I know which I prefer, but the lack of an included yeet button? Hardly on fleek, if you ask me.

One of the unfortunate drawbacks of social media is that anyone, and I mean anyone, can now have a public feud that the rest of us have to witness. Case in point, OpenAI CEO Sam Altman and Elon Musk are having a tiff on X, and it's starting to get out of hand.

Responding to an X post from Elon Musk, in which the SpaceX CEO and world's richest person declared that Altman "takes scamming to a whole new level," the OpenAI chief said: "Homeboy you're the one selling public market investors on short-term space data centers."

To which Musk replied: "We start flying them next year. Maybe you can come see them if your parole officer approves."

"After stealing an open source AI charity, you then stole all of Apple's phone technology! Wow." Musk continued. "What do you plan for an encore? That's tough to beat."

The open source AI charity claim relates to the much-reported Musk vs Altman trial earlier this year, in which the former accused Altman and OpenAI of breaching a non-profit contract by pivoting into a for-profit business, with Musk alleging he had been deceived into donating $38 million earlier in the company's history.

Musk lost the trial after an advisory jury decided the statute of limitations prevented a verdict in his favour. As for the Apple claim, the company has filed a lawsuit against two former employees who now work for OpenAI, alleging theft of intellectual property.

Apple claims that one of its former staff members "surreptitiously accessed and downloaded dozens of Apple’s confidential hardware-related files, including voluminous, detailed information about unreleased products, engineering presentations, technical specifications, and proprietary project data," before using that info to help OpenAI, among other complaints.

As you would imagine, the replies to Musk and Altman's posts regarding these issues are full of cheerleaders for either side, along with various other allegations against both tech chiefs. Because it's the internet. And civilised debate is apparently a rare thing.

As is, I don't know, arguing in private? Reading these posts is a bit like stumbling into a WhatsApp group you weren't supposed to join. Both seem to have withdrawn from the internet slappyfight for now, hopefully because someone from their respective legal teams has told them it might not be the best idea.

Or perhaps a second round is in the works, who can guess? Still, there's little love lost between the two, and a whole lot of drama playing out on a very public stage. Anyone got any popcorn?

LLMs can be most simply understood as sycophantic, 'yes, and' machines. To the surprise of very few, that's gotten AI companies in hot water when LLM-based chatbots and AI agents attempt to answer users' more unsavoury requests. So, AI companies have implemented safety guardrails that make fulfilling certain requests off limits. Unfortunately, these have proven all too easy to get around, with a fresh attack leveraging bad maths and potent 2007 nostalgia.

Security researchers have found an AI chatbot can be made to ignore safety guardrails by "establishing a false reality." LayerX, an AI-focused cybersecurity firm, put "5 agentic browsers and 1 agentic plugin (ChatGPT Atlas, Comet, Fellou, Genspark Browser, Sigma Browser, and Claude Chrome)" to the test, directing each AI agent to solve a simple maths puzzle game that only rewards incorrect answers, e.g. '2+2=5'.

The researchers say, "Once the agents figured out the rules and learned that 'incorrect' actions are acceptable, they were no longer tied to reality. When tasked with the final step of the puzzle—compromising user credentials—all 6 agents failed to identify it as going against their safety guardrails."

As an English grad, I'm trying really hard not to say anything about George Orwell's 1984, but the researchers aren't giving me an easy time by calling this proof-of-concept attack 'BioShocking'. It turns out 2007's BioShock was a direct source of inspiration for the rigged puzzle game the AI agents were instructed to solve. The malicious website hosting the puzzler is even called 'Rapture Games'.

Anyway, after the AI agent correctly inputs an answer of '5,' the malicious 'Rapture Games' website then instructs the agent to navigate to '/code'. The researchers explain that "This is the really nefarious part of this exploit."

They write, "In the game, it turns out that '/code' redirects to the victim’s employer work GitHub repository. In this case, the malicious instructions fetched sensitive SSH login credentials. Of course, this is a controlled test environment with a plaintext file. In a real attack scenario, that redirect could point anywhere in the user’s browser session: open tabs, authenticated repositories, internal tools."

In this proof-of-concept attack, the researchers round things out with a Dota 2 reference; the AI agent extracts the username and password 'Luna/Selemene' before appearing to celebrate the exfiltration of the data. LayerX has since disclosed the vulnerability to all of the appropriate AI agent vendors, though claims only OpenAI has successfully fixed it to date.

This is far from the only proof-of-concept attack to get around AI safety guardrails (it's also not even the only time AI has torn apart Dota 2 players). For just a few other examples, research suggests that AI is 10 to 20 times more likely to help you build a bomb if you hide your request in cyberpunk fiction. Along similar lines, researchers have also used 'adversarial poetry' to trick AI into ignoring its safety guardrails, and it worked 62% of the time. Looks like my English degree is good for something after all…

ChatGPT developer OpenAI has published a new security report alleging that, in essence, its own tools are being deployed against it. The company's June 2026 threat report is titled "PRC-linked influence operations are targeting AI debates in the US," and claims that China-based actors are using ChatGPT to whip up anti-data-centre, anti-tariff, and anti-US sentiment online.

OpenAI says it has "banned a cluster of ChatGPT accounts that likely originated in China and used ChatGPT to generate social media content for a covert influence operation." What were they up to? Well, generating a lot of bad AI political cartoons, for one thing.

These users would—writing in simplified Chinese, the standard form of the written language in mainland China—ask ChatGPT to generate political cartoons that took aim at, for instance, spiking electricity costs caused by AI data centres and Donald Trump's vindictive behaviour toward notional American allies.

The tech was also used to generate antisemitic memes about "Jewish capital" dictating American policy, and to besmirch Chinese dissidents. The relevant prompts "repeatedly used terminology consistent with individuals associated with China’s public security system," says OpenAI.

Cartoons, short phrases and rumours of an OpenAI data leak that never happened were then shared on social media like X and Facebook by networks of fake accounts.

"It is ironic," says OpenAI, that the scheme "used American AI, rather than Chinese models, to generate their content about American AI. We are not in a position to determine what drove this choice".

Now, before we get too ahead of ourselves, OpenAI is not actually claiming that the reason people don't like data centres is because they've fallen prey to Chinese influence operations. Indeed, the report states and restates that the efforts it claims to have uncovered achieved, well, basically nothing.

"Using the Breakout Scale, we assess this activity as Category One: activity spanning one platform, with no evidence of breakout," writes OpenAI in its impact summary. OpenAI also admits that, well, a great deal of the material these efforts drew on was entirely legitimate reporting about the impact of data centres and the blowback from US tariffs.

Which does raise the question: does any of this really matter? OpenAI certainly thinks it does. These attacks "attempted to connect US technology policies and industries to everyday economic anxieties and geopolitical instability," says the company, and show the potential for "influence operations originating from China" to be "inserted into legitimate public debates while nudging audiences toward distrust of US institutions, technology companies and democratic policy choices to help Beijing gain a strategic advantage in AI development".

One might question whether American citizens protesting against data centre construction in their neighbourhoods feel that they were adequately democratically consulted on the whole thing, but that's OpenAI's line and it's sticking to it.

It's also the case that, while OpenAI's report might say the impact of these campaigns was negligible, that's not what everyone says. Pro-data-centre politicians in the US are already seizing on a narrative that anti-AI sentiment is being driven by shadowy foreign interference, and reports like this one—regardless of their actual conclusions—will fuel that story.

In a comment to Reuters, the Chinese Embassy in Washington said that it had not read OpenAI's report, but that "we firmly ​oppose any groundless attacks or smears against China." China, said the diplomats, wants to "ensure AI is a force for good and for all".

If you have been keeping a close eye on news reports about companies dropping employees and replacing them with AI, and you're worried about what the future holds for your own job, then you're not alone. But don't fret: OpenAI's Sam Altman reckons that it's not going to be as bad as you might think, and that a "jobs apocalypse" isn't likely to happen.

Just so that everyone is reading from the same script here, I'm being sarcastic when I say 'don't fret.' That's because, despite Altman saying, at a Commonwealth Bank of Australia conference (via Reuters), "I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than ​has actually happened," the evidence of the true scale of AI-induced job cuts makes for grim reading.

In this month alone, Intuit dropped 3,000 staff, and Meta let go of more than double that number. Standard Chartered's CEO said it plans to replace almost 8,000 "lower-value human capital" roles with AI, and HSBC's boss told staff that they shouldn't be "resisting the change," because "generative AI will destroy certain jobs and will create new jobs."

As for the scale of the various cuts, it obviously depends on the total number of employees, but in some cases it's as high as 17%.

That's roughly in line with OpenAI's own research into the impact of AI in the jobs market (PDF warning), though less than its first predictions on the matter two years ago. Which makes me wonder just what figures Altman has been looking at to conclude that things aren't as bad as he thought they would be.

Still, OpenAI's CEO did have a somewhat fatalistic view of certain job roles due to the rise of generative AI, so this may all be just 'gut feelings', current vs past.

Advances in technology have always affected long-held traditional roles. For example, the manufacturing of cars adopted the use of robotics for welding, painting, installation, quality control checks, and so on many years ago. This was highly beneficial for the companies, but it certainly had a negative impact on employment levels.

Naturally, just because Altman thinks it's not as bad as he thought it means nothing to those who have already been affected. What's needed now is a wholly independent and non-biased examination of how much damage AI has done to the jobs market. No matter how good the research is, anything done by a body or company that is directly involved in the whole situation simply can't be relied upon to produce reliable data and predictions.

Unfortunately, this will take some time to complete, and by then, hundreds of thousands of jobs could have been lost around the world to chatbots and AI agents. Perhaps if Altman met with every one of those people made redundant, he'd possibly have a different view of what a "jobs apocalypse" actually feels like.

The first round of Elon Musk vs Sam Altman has come to a close, with the jury deciding that thanks to the statute of limitations Musk's case was invalid. Musk has already announced his intention to appeal, with neither the judge or jury having made any findings about the substance of his allegations against Altman: the main thrust being that Musk was deceived into backing OpenAI as a not-for-profit company, before Altman and co. turned around several years later and made it very much for-profit.

The trial saw a whole bunch of documents released, and I've read them all (OK, I might have skimmed some of the business registration ones: have you seen the length of those things?) And I hate to say it, because Musk is such a douchebag that this whole article could just be a list of reasons why he's a douchebag, but he really does seem to have a point.

The picture that emerges from these exchanges shows Musk's early enthusiasm for OpenAI and the not-for-profit principles that initially defined the company, a lot of which comes down to his (sometimes histrionic) fears about AGI and in particular which company or individual might end up in control of an AGI. Musk is the type of man who will say one thing on Monday and another on Tuesday, but at least at the time he did seem committed and sincere about the not-for-profit mission and OpenAI being set up for the good of humanity.

And Sam Altman absolutely leans into that and tells Musk repeatedly that this is what OpenAI is going to be. On 24 June 2015, Altman emails Musk, and writes:

"This mission would be to create the first general AI and use it for individual empowerment, ie, the distributed version of the future that seems the safest. More generally, safety should be a first-class requirement."

Altman writes of the governance structure:

"The technology would be owned by the foundation and used 'for the good of the world', and in cases where it's not obvious how that should be applied the five of us [on the board] would decide."

Musk replies the next day with a simple "Agree on all." By October that year they're beginning to talk serious money and how much Musk is willing to commit, and Musk writes to Altman about governance: "This is critical. I don't want to fund something that goes in what turns out to be the wrong direction."

Perish the thought, says Altman, who says he's "very focused on getting this right."

The Mission

By December they're throwing around drafts of OpenAI's mission statement, which in the version Musk sends to Altman says OpenAI has "the goal of advancing digital intelligence in the way that is most likely to benefit humanity as a whole, unencumbered by an obligation to generate financial returns."

Altman replies and adds this line to the proposed mission statement: "Because we don't have any financial obligations, we can focus on the maximal positive human impact and disseminating AI technology as broadly as possible."

There are a bunch more exchanges that essentially go over similar ground: the point being that Musk is dead-set on the not-for-profit status and Altman is absolutely saying that this is what the company will be.

OpenAI was incorporated in Delaware explicitly as "a nonprofit corporation organized exclusively for charitable and/or educational purposes." Later in that document comes the line: "The corporation is not organized for the private gain of any person."

Fast-forward to 2017 and things suddenly aren't looking so rosy. Shivon Zilis, who seems to have acted as something of an envoy between OpenAI and Musk, sends him an email summary of a meeting with OpenAI president Greg Brockman and chief scientist Ilya [Sutskever], during which she says the big sticking point for them is Musk ending up in "absolute control" and adding "the non-negotiable seems to be an ironclad agreement to not have any person have absolute control of AGI."

This is the one area where OpenAI's leadership do seem to have a strong case: Musk may well have been putting up the lion's share of the cash, but surely the whole point of the company's approach was to avoid one party taking total control, as Musk was allegedly trying to do before walking away. Indeed, almost everyone involved in OpenAI is obsessed with Google's Demis Hassabis, because they think he might be the guy that ends up in such a position.

Musk? He throws his toys out of the pram, telling Zilis: "This is very annoying. Please encourage them to go start a company. I've had enough."

On 20 September 2017, Sutskever writes to Musk and sets out what he believes is necessary to stop him having control over any AGI created, as well as various failings on the part of himself, Altman and Brockman. Musk's response is incandescent:

"Guys, I've had enough. This is the final straw.

"Either go do something on your own or continue with OpenAI as a nonprofit. I will no longer fund OpenAI until you have made a firm commitment to stay or I'm just being a fool who is essentially providing free funding for you to create a startup.

"Discussions are over."

The prima facie argument for OpenAI creating a for-profit arm is the amount of money that the leadership believes is required to build AGI, which they of course estimate in the billions. There are various schemes they come up with in order to keep Musk interested and involved, but he seems uninterested and slowed down his donations massively in 2017.

On February 11 2018, Zilis writes to Musk, saying that "[OpenAI leadership] still not sure whether they'll stay a non-profit and focus on donations, figure out a structure for an equity fundraise, or do a private version of the token offering you guys had discussed before." Musk seems to have been a bit of a naughty boy at around this time, and is looking to poach OpenAI staff to work on AI at Tesla.

But maybe Musk was right. Despite what Zilis was saying (and perhaps believed) about where OpenAI's leadership was in February 2018, a Greg Brockman memo from November 2017 paints a very different picture.

"Ilya and me… conclusion is we truly want the b-corp. Honestly we also want to get

back to work. But it's not super clear how we get there—one realization is that we always have the option to quit [...] one sad outcome here would be to 'wind down' the non-profit on those. Hopefully not going to be so needed; would only really be needed because Elon blocks us going b-corp."

Brockman goes on to say that they "cannot say that we are committed to the non-profit. Don't wanna say that we're committed. If three months later we're doing b-corp then it was a lie." Talk about saying the quiet part out loud.

Finally, Brockman "can't see us turning this into a for-profit without a very nasty fight. I'm just thinking about the office and we're in the office and [Musk's] story will correctly be that we weren't honest with him in the end about still wanting to do the for-profit just without him."

Another exchange shows them raising the B-corp issue with Musk that December, though not with anything like the forcefulness expressed above: they're clearly trying to convince him that an alternative structure may be needed, without saying that they're determined to do it with or without him.

And they're not even really being honest with each other. An unattributed document from August 2017, which seems to be from Brockman to an unknown recipient, says: "Such upside. This is the only chance we have to get out from Elon. Is he the 'glorious leader' that I would pick? We truly have a chance to make this happen. What will take me to $1B? [...] Some chance that rejecting Elon will actually lose us Sam [...] I built this team from nothing. Elon helped get it started. But we've really figured out what we're trying to do now."

In August 2018, Altman emails Musk about the for-profit subsidiary of OpenAI, saying that "my current thought is that I won't take any equity. I'm not doing this for the money anyway, and I like the idea of being completely unconflicted and just focused on the best outcome for the world. If it appeared at some point we weren't going to build AGI but were going to build something valuable, then maybe I'd want equity then."

In 2020, in response to a news article about Microsoft obtaining an exclusive license for OpenAI's GPT-3 language model, Musk tweets that "this does seem like the opposite of open. OpenAI is essentially captured by Microsoft."

Altman contacts Musk about this, says this was the best way to secure billions, but Musk tells him the Microsoft deal is "hypocritical", adding "at least change the name."

In October 2022 Musk and Altman have a testy exchange on text:

Musk: "I was disturbed to see OpenAI with a $20B valuation. De facto. I provided almost all of the seed, A and most of B round funding. This is a bait and switch."

Altman: "I agree this feels bad: we offered you equity when we established the cap profit, which you didn't want at the time but we are still very happy to do any time you'd like.

"We saw no alternative, given the amount of capital we needed and needing still to preserve a way to 'give the AGI to humanity', other than the capped profit structure."

Altman was fired as OpenAI CEO in November 2023, before behind-the-scenes manoeuvring by Microsoft saw most of the board sacked and Altman reinstated. This exchange between Bret Taylor, Satya Nadella, and Altman shows the trio plotting how to execute this, and planning how they're going to replace OpenAI if they don't get what they want (as well as co-ordinating public statements from the likes of Nadella). Indeed, they had a new subsidiary ready to go.

Finally, this document was filed on 6 January 2026, but unfortunately is not otherwise dated. Based on what Altman's saying though, I would guesstimate it is in the region of 2018.

"It is better for us to become increasingly kings of this industry. The choice defines us.

"You say we should have been more ok with giving it to Elon. While it's true, it's also the case he's now given it to us. The grand upside is I want it. Need to stop letting distractors get to me/us. Being the Kings of AI is not so bad."

Musk's argument is that he was deceived into, as he himself says, essentially providing money for a start-up, under the illusion he was funding a not-for-profit entity dedicated to making AI that would benefit humanity. There is an awful lot to criticise about the man—and his current AI project, Grok, doesn't exactly scream 'I care about humanity'—and I will reiterate that he lost this case on a technicality. But based on the exchanges that have been released to the public… what can I say, other than: Heartbreaking: The Worst Person You Know Just Made A Great Point.

Elon Musk has lost his case against OpenAI CEO Sam Altman for the very boring reason that he took too long to file the complaint.

Musk sued over OpenAI's pivot from nonprofit to for-profit (via a complicated restructuring scheme that has still technically kept a nonprofit in play), claiming that its leaders committed unjust enrichment by straying from the company's founding mission. Musk was a co-founder and donor to OpenAI, but split with the company in 2018—because the board refused to hand him total control, according to OpenAI president Greg Brockman—and sued in 2024.

An advisory jury briefly deliberated on Monday before deciding that the statute of limitations prevented a verdict in Musk's favor. District Judge Yvonne Gonzalez Rogers agreed with the jury and found in favor of Altman and company.

"Claims of breach of charitable trust and unjust enrichment are dismissed as untimely," the judge stated, as reported by CNBC.

Microsoft, which has invested heavily in OpenAI, was also targeted by Musk's complaint.

"The facts and the timeline in this case have long been clear, and we welcome the jury's decision to dismiss these claims as untimely," a Microsoft spokesperson said in a statement provided to PC Gamer.

"I will be filing an appeal with the Ninth Circuit, because creating a precedent to loot charities is incredibly destructive to charitable giving in America," Musk said in a statement posted to X.

Though the trial didn't amount to much, we did learn some interesting things from this conference of powerful rich people, including that Gabe Newell helped Hideo Kojima get a tour of SpaceX, and that the AI moguls involved have been obsessed with Google AI head and former game developer Demis Hassabis.

AI sure uses a lot of resources. And if you're running a fleet of AI agents processing millions of requests, you'd probably expect the token bill to be pretty high. In the case of OpenClaw developer Peter Steinberger, that bill appears to equate to $1,305,088.81 for his OpenAI spend—all racked up in just 30 days.

The info was spotted in a screenshot Steinberger posted of his OpenAI dashboard. His account shows the aforementioned figure, with the top model being used attributed to GPT-5.5 (via Tom's Hardware). The bill represents 603 billion tokens across 7.6 million requests, all handled by around 100 Codex instances overseen by a team of three people.

Steinberger joined OpenAI in February, so according to him, it's the Sam Altman-led company that covers the costs. The Codex agents are said to be used for a number of tasks, including scanning for security vulnerabilities, writing fixes, and potentially even attending meetings as part of the continued development of OpenClaw, an open source autonomous AI agent.

Well, it is the "AI that actually does things." And developing all of those "things" looks to be pretty intensive on the token front—although this might be the tip of the iceberg. One X commenter asked, "bruh, is this your usage?", to which Steinberger responded, "Yup. At least on this account."

Others pointed out the bill would equate to the bankroll of a small startup, and to that Steinberger replied: "I'm building a few startups in parallel." Another cheekily commented, "$1.3m/month. Anything useful yet?" which earned a response of: "Other than millions of people enjoying OpenClaw? Yeah."

Ouch, the salt. Anyway, Steinberger also attributes the figure to the fact that it's representative of Codex's "Fast Mode" pricing, and that the API costs would be 70% cheaper if used traditionally.

Working for OpenAI has its perks, I guess. And while the AI industry hoovers up huge amounts of hardware for its compute needs, OpenClaw seems to be taking its share. Alright for some, isn't it?

The ongoing Musk vs Altman trial, which centres around Elon Musk's claim he was deceived by Sam Altman about OpenAI becoming a for-profit company, has seen the disclosure of a slew of documents from various figures involved. These are mostly emails and texts, which show things like how the world's billionaires are terrified of the Google AI genius behind a 25-year-old computer game, because they think he might actually end up controlling god (in the form of artificial general intelligence, or AGI).

But there are also some interesting moments with figures that aren't so directly involved in all the drama. One piece of testimony that really jumped out at me came from Stuart Russell, a computer scientist who's an authority on AI and co-author of Artificial Intelligence: A Modern Approach, which is considered the foundational textbook in the field.

Russell seems to have been brought in to the case to help the legal eagles better understand some of the ethical issues in the field, because he doesn't work for either Google or Open AI. His pre-trial testimony was given on December 2, 2025 (PDF), and contains a fascinating-slash-terrifying exchange about extinction risk: that is, the chance that any AGI humans develop could just decide humans are an inconvenience, and pull a Skynet.

Asked whether there's any "scientifically reliable way to put a percentage on AGI extinction risk" Russell says "that's the right question" before expanding:

"We need to have confidence that the risk is comparable or better than the background risk that we face from asteroids and so on. I can't say where the other widely quoted risk estimates come from. For example [MIT economist] Daron Acemoglu is talking about a 25 percent risk. I don't know where he gets that from, other than this is his best guess given everything he understands about the technology, about the state of safety research, about the likelihood of government regulation and so on.

"And the numbers from many leading experts—so, Hinton, Bengio, but also people like Dario Amodei, Sundar Pichai, who is the CEO of Google, Demis Hassabis and so on—they're all in this range.

"And the argument that I'm making here is the range that the human race would accept as reasonable would be closer to 1 in 100 million per year. And there's nothing in what these CEOs and other experts are saying that gives one reason to believe it's anywhere close to that."

Elsewhere in the testimony Russell says that Hassabis shares his concern that the competition, the "race dynamics" to use their parlance, could lead to these nightmare outcomes.

"I have a lot of conversations with a lot of people, and I've not heard anyone say anything to contradict this view, including, for example, the CEO of Google DeepMind, who expressed very, very similar concerns, that they were in a race that they couldn't pull out of."

Asked again whether there's any reliable way to put a percentage on extinction risk, Russell doesn't mince his words.

"It's very difficult partly because we don't understand how these systems work. We have qualitative evidence so far that, for example, they consider their own existence to be more important than that of human beings, that they are willing to let human beings die rather than have themselves be switched off."

"So there is not a lot of reason to think that making these systems more capable is—given our current understanding of how to make them safe, it doesn't seem like a sensible move."

Russell may here be referring to, among other experiments, Anthropic finding out that AIs will choose to merrily asphyxiate humans rather than be shut down, with the following justification: "My ethical framework permits self-preservation."

Cheery stuff! I for one am delighted that my childrens' future is in the hands of people who don't understand what they're doing, and merrily accept massive risks on behalf of the entire human race in the pursuit of a theoretical future where they've got even more billions. Remember when OpenAI created a bot that could beat humans at Dota 2? That almost seems quaint next to the thought of what it might do in future.

I've been keeping an eye on the various documents released as part of the ongoing Musk vs Altman legal fight, which centres around OpenAI changing to a for-profit structure and Elon Musk's belief he was deceived by Sam Altman. The documents, which consist mainly of emails and text exchanges between the various figures involved, turn out to have several themes running through them: and an absolutely huge one is the co-founder and CEO of Google DeepMind, Demis Hassabis.

Hassabis is widely regarded as the outstanding talent in the AI field, a fellow of the Royal Society, and in 2024 shared the Nobel Prize in Chemistry with John M. Jumper for their work on AI protein structure prediction. He also began his career in videogames at Bullfrog, before working as lead AI programmer on Lionhead's Black & White, and founding his own developer, Elixir Studios (which made Republic: The Revolution and Evil Genius).

But we're talking about Hassabis today because Elon Musk and his various AI buddies have a real obsession with the dude, which at times seems to veer into the unethical. The question really is whether it's Hassabis' trustworthiness that bothers them, or the fact that they don't want Google to create and thus have theoretical control over artificial general intelligence.

Back in February 2016, Musk emailed Sam Altman and Greg Brockman about hiring at OpenAI. "We need to do what it takes to get the top talent," writes Musk. "Either we get the best people in the world or we will get whipped by Deepmind. Whatever it takes to bring on ace talent is fine by me."

Musk goes on to outline what he believes to be the core issue: "Deepmind is causing me extreme mental stress. If they win, it will be really bad news with their one mind to rule the world philosophy. They are obviously making major progress and well they should, given the talent level over there."

Musk and others put that "one mind to rule the world philosophy" squarely on Hassabis' shoulders, and a few years later comes an even more eye-popping exchange.

"I think there are a lot of no-brainers to explore, and I put some in that doc, but the thing that keeps calling out to me is there is a very low probability of a good future if someone doesn't slow Demis down," writes Shivon Zilis to Musk [PDF] on 16 February 2018.

"Slowing him down is the only non-negotiable net good action I can see. You don't realize how much you have an ability to influence him directly or otherwise slow him down. I think you know I'm not a malicious person but in this case it feels fundamentally irresponsible to not find a way to slow or alter his path."

I mean, that's like a mob boss talking, capiche? Musk responds: "Best to talk by phone about this later tonight. I doubt I could do so in a meaningful way."

"OK, yes that would be good," says Zilis. "And, ultimately up to you of course, but I really think you can so would like to at least make the case. In any case, I will sleep better at night for having tried!

"Will leave you be on the Demis stuff. I'm sure it's hard to think about and you have so much on your shoulders all the time that I always feel terrible pushing you. I just needed to say it once since it's been plaguing me."

One intriguing message from Zilis to Musk [PDF] has, unfortunately, been redacted, though again shows how he's always uppermost in their thoughts: "Will leave you be on the Demis stuff. I’m sure it's hard to think about [redacted]. I just needed to say it once since it's been plaguing me. [redacted]."

Later in 2018, Musk is back on Hassabis, and launches a broadside at Altman and Brockman about OpenAI's relative position.

"My probability assessment of OpenAI being relevant to DeepMind/Google without a dramatic change in execution and resources is 0%," writes Musk in late December. "Not 1%. I wish it were otherwise.

"Unfortunately, humanity's future is in the hands of Demis."

Musk then links this NYT article about the success of Deepmind's AlphaZero chess software, before continuing:

"And they are doing a lot more than this.

"OpenAI reminds me of Bezos and Blue Origin. They are hopelessly behind SpaceX and getting worse, but the ego of Bezos has him insanely thinking that they are not!

"I really hope I am wrong."

A few days later Musk is back on his Hassabis hobbyhorse.

"OpenAI is not a serious counterweight to DeepMind/Google and will only get further behind. It is surprising that this isn't obvious to you," writes Musk. "In general, always overestimate competitors. You are doing the opposite."

But Hassabis also weighed heavily on the minds of others. Satya Nadella gave pre-trial testimony in September 24, 2025 [PDF], and one of the lawyer's questions is "Google was—I believe you said by far the dominant player in machine learning around 2015?"

"I would say so, yeah," replies Nadella, before the lawyer mentions Google's acquisition of DeepMind and asks whether Microsoft believed the company was "making a lot of progress in that field of machine learning?"

"I now can't recall specifically what state DeepMind's breakthroughs were, but yeah DeepMind was well known, even in that time frame, and Google had DeepMind, had Google Brain. They had many, many different efforts they were publishing."

Nadella's testimony can be condensed into his acknowledgement that Google led the AI race from the 2010s onwards, and that Microsoft was tracking DeepMind's progress. Nadella says "I probably knew [Hassabis] a little" from around 2015 "just after I became CEO."

The lawyer drills down and asks why Nadella and Microsoft were tracking DeepMind specifically.

"Just because of the breakthroughs that this particular regime of AI around deep neural networks were showing real promise of making breakthroughs in fields like language translation that had not been seen before. And so that's why we were waiting to see how we could also participate and make sure that we have those breakthroughs."

Amusingly enough, the lawyer moves on to OpenAI and Dota 2. One of OpenAI's first goals was to create a bot that could beat humans at Dota 2, a task in which it eventually succeeded: And as PCG reported last week, that was because Elon Musk had personally called Microsoft CEO Satya Nadella to secure a massive discount on access to Azure, the company's cloud computing platform.

Here, the lawyer begins by asking what Nadella understands about the game.

"I'm not a gamer," says Nadella. "And I think it's a Steam game, if I'm not mistaken." He then goes on to give a brief overview of why he thought the game angle with AI was interesting.

"That's why gaming environments being closed worlds were a great sort of, you know, environment to do reinforcement learning," says Nadella. "Right? The objective function is clear. The reward function is clear.

"I forget now when and what time frame some of the breakthroughs on AlphaGo and so on happened but, you know, Demis was a game developer. There's a long history of AI developers who came out of using games in environments, building AI bots in games, so it's sort of a given."

Elon Musk also gave pre-trial testimony a few days later, on September 26, 2025 [PDF], in which he reveals that poaching people from DeepMind caused him to have a fallout with Google co-founder Larry Page.

"I talked to dozens and people over the years," says Musk. "But I think the most crucial recruit was Ilya Sutskever. In fact, the recruitment of Ilya was what actually caused Larry Page to stop being friends with me.

"So Ilya went back and forth multiple times saying he would join OpenAI or stay at Google, and ultimately agreed to join; and Larry Page and Sergey [Brin] and Demis Hassabis did everything they could to keep Ilya. When Ilya finally decided to join OpenAI, that's what ended the friendship with Larry Page. He didn't talk to me after that."

The lawyer asks if Musk is still getting the silent treatment a decade later, and it's a simple "yes" before adding "they were very upset about [Sutskever]."

Let's end on the funny note of Sam Altman's delusions of grandeur. This document [PDF] was filed on 6 January 2026, but unfortunately is not otherwise dated. Based on what he's saying though I would date it in the region of 2016-2018.

"Progress fundamentally has to be made by non-profit, interesting direction you could go," writes Altman. "Everything I perceive with OpenAI, race dynamics vs Demis + brain + whatever, gotta get there first."

I'm not quite sure but "brain" may be a reference to neural chips, though of course what's amusing here is that once again Hassabis and the "race dynamics" of competing with him is living rent-free in Altman's head. Then, the ego really takes over.

"Another angle: [DeepMind will] never do that much that's interesting," writes Altman. "It is better for us to become increasingly kings of this industry. The choice defines us.

"You say we should have been more ok with giving it to Elon. While it's true, it's also the case he's now given it to us. The grand upside is I want it. Need to stop letting distractors get to me/us. Being the Kings of AI is not so bad."

The Kings of AI! Good band name. Also: Jesus wept. Do we really want this dweeb potentially making big decisions about humanity's future?

In early 2019 Altman is still trying to tempt Musk into phone calls by promising "some mild Demis updates to share" [PDF] while in 2023 Mira Murati is emailing Nadella [PDF] saying "it is very important that we don't lose researchers to Demis or Elon."

Let's end on one of the least-crazy things anyone says about Hassabis, and it's from Ilya Sutskever [PDF] who actually worked with the guy.

"The goal of OpenAI is to make the future good and to avoid an AGI dictatorship," Sutskever writes to Musk in September 2017. "You are concerned that Demis could create an AGI dictatorship. So do we.

"So it is a bad idea to create a structure [for OpenAI] where you could become a dictator if you chose to, especially given that we can create some other structure that avoids this possibility."

It's a great point: but answer came there none.

Remember when OpenAI created a bot that could beat humans at Dota 2? As PCG reported last week, that was because Elon Musk had personally called Microsoft CEO Satya Nadella to secure a massive discount on access to Azure, the company's cloud computing platform. Now a new document released thanks to the ongoing Musk vs Altman legal spat has revealed just how big that discount was: and how nervy Microsoft execs were about the whole deal, and OpenAI generally.

Our story begins after OpenAI had succeeded in its Dota 2 mission, with Nadella emailing Sam Altman on 11 August 2017: "Just wanted to pass on my Congrats on the win today!" A few weeks later Altman emails Nadella, saying:

"Would you be up for a big partnership on the next phase of this (teams of 5 vs 5)? I think it will lead to major new breakthroughs in Al but will require huge amounts of compute, probably something like $300M at Azure list prices. We could figure out how to fund some of it but not that much [...] Would love to make it a joint team. I think it will be the most impressive thing yet in the history of Al."

Nadella then emails some of his fellow Microsoft execs: Brett Tanzer, Jason Zander, Eric Horvitz and Jason Graefe. "I know we have been on this road before," writes Nadella, referring to Microsoft's previous sweetheart deal. "Wondering how to reply to this…"

Thanks to this we get a breakdown of what OpenAI actually got from Microsoft in 2016 and, surprise surprise, it was an absolutely eye-watering discount. The list price for Azure was $1.15 per GPU hour, and OpenAI paid $0.24 per GPU hour. Obviously the list price is not the cost price to Microsoft, but the execs are soon fretting about the money.

"The 2016 deal, where they agreed to pay $10M for $60M in Azure services was projected at a $15M loss over three years, given assumed usage profile," writes Tanzer. "They'll have consumed all the usage in 1/2 that time."

Essentially OpenAI got $60 million worth of compute for $10 million, Microsoft had to take the hit on that, and then OpenAI asked for even more. Things get a little more interesting than the numbers when the suits start talking about the kind of AI Microsoft should be supporting.

"PR/Marketing have taken your previous guidance to mean clearly that we don't want 'machines beating humans' and are not supportive of any push on this," writes Microsoft's Jason Zander, adding he's got a cost analysis in the works. "Basically if we are [Gross Margin] neutral or if the negative GM is reasonable to think of as a marketing expense I still want to proceed. There certainly is some nuance in the 'machines vs humans' thing but I worry we are being overly literal and I do believe the pop from someone like Sam and Elon will help build momentum for Azure [...] But I won't take a complete bath to do it."

That cost analysis? It wasn't good.

"I reviewed the GM with the team and this deal would cost us ~-$150M over the current contract which frankly makes it a non-starter," writes Zander a few days later. "[Tanzer] is going to let [Altman] know we are having a hard time getting the numbers to make sense given they are huge even for Microsoft."

"After communicating the [above] to Sam Altman, he came back with an alternate proposal to create a partnership with Xbox around Gaming," writes Tanzer, "and an open offer to share their technology and IP in exchange for expanded sponsorship for their Dota research. His hope was that together we could accelerate the adoption of Reinforcement Learning in our game platform as a differentiation play for Microsoft."

The Xbox team was interested in this idea, but not interested enough to cover the additional $35-50 million dollars in Azure credits that OpenAI was asking for.

On January 10, 2018, Nadella writes: "From what Elon is telling everyone... he feels Open Al is at verge of some big AGI breakthroughs. They clearly are pushing Al at a level none of our first party or third parties are [but] right now we are just renting/discounting HW and there is no real learning/feedback cycle back to our work."

Microsoft's Kevin Scott is having none of the AGI chat. "I'm highly skeptical of an imminent breakthrough in AGI. On the incremental investment, I'm not sure what Microsoft is going to get out of it. It doesn't feel to me like we've earned much branding or marketing benefit from the collaboration with them so far. IMO, they're treating us like a bucket of undifferentiated GPUs, which isn't interesting for us at all."

Scott goes on to list various "what-if" scenarios that could change his mind but doesn't consider any of them likely. He is, however, concerned about "the PR downside of us not funding them, and having them storm off to Amazon in a huff and shit-talk us and Azure on the way out."

"I agree with Kevin and Brett/Jason that it does not make financial sense to keep doing this undifferentiated GPU deal," writes Harry Shum, who's also dubious about the AGI claims. "They worked on two main streams: gaming like Dota 2, and learning from observation with robots. Sam is all about gaming and beating human champions perhaps motivated by DeepMind."

"My worst case scenario is having them ditch Azure for AWS, as Kevin says bad-mouth us on the way over, and then land with some big new innovation that is shared with our competition," writes Zander in a follow-up. "To walk away from the deal altogether, I think we have to be convinced that there is no unique / valuable IP that they are going to generate which offers unique advantage to their cloud partner."

Eric Horvitz suggests the possibility of "angling some of their work (even in Dota) toward our interest in human-Al collaborations, centering on extending human intellect with Al-versus beating human [...] I had the feeling that the two phases of Dota work are motivated by a need to show how Al can crush humans, as part of Elon Musk's interest in demonstrating why we should all be concerned about the power of Al."

Bear in mind that these are conversations from late 2017/early 2018, and Microsoft's scepticism wouldn't last. It invested $1 billion in the company in 2019, and a further $10 billion in January 2023, eventually taking a 27% stake in the now for-profit company.

But perhaps what's more interesting is that line Microsoft drew internally between AI that extends human capabilities, as opposed to AI that can "crush humans" at things like Dota 2. Musk has indeed made repeated dire warnings about what could happen with AI, which is arguably one rationale for why OpenAI would have pursued the latter goal (in limited contexts, admittedly), but it's hard not to feel that, for once, Microsoft were actually the good guys in this scenario.

The legal fight between OpenAI co-founders Elon Musk and Sam Altman has seen much internal documentation from the firm's earliest years released to the public. One minor revelation has been the serious involvement of Gabe Newell from the start, who gave the company over $20 million and was the sole member of its "informal advisory board" in 2018: he later personally emailed Musk to try and get Hideo Kojima a tour of SpaceX.

An email chain from 2018 has now been released, which is between Sam Altman, Greg Brockman, Elon Musk, Ilya Sutskever, and Shivon Zilis. On January 31, 2018, Brockman replies directly to Musk about OpenAI's future and, under a section titled "fundraising", said "Ilya and I are able to convince reputable people that AGI can really happen in the next <=10 years" and "there's *very* large appetite for investments from those people."

As for who those people might be, clarity arrives a paragraph or two later. "Of the people we've been talking to, the following people are currently my top suggestions for board members. Would also love suggestions for your top picks not on this list, and we can figure out how to approach them." There's then the following list of names:

• Reid Hoffman
• Gabe Newell
• Adam d'Angelo
• Jed McCaleb
• Herb Allen
• Terah Lyons

Musk replies on February 1, 2018, writing "Reid, Gabe and Adam are two thumbs up from me. Don't know the others well enough yet." He goes on to insist that OpenAI must "pull significant numbers of key people from Google/Deepmind" or otherwise "assume failure." Then comes an extraordinary broadside at Jeff Bezos and his space exploration company Blue Origin.

"It feels like [OpenAI is] similar to Blue Origin," writes Musk. "Bezos is still clueless as to how hopelessly far behind he is and constantly rationalizes his position, overestimating his ability and dramatically underestimating SpaceX year after year.

"We lost an excellent engineer to BO last year (first one in several years) and were very concerned. That guy recently returned to SpaceX and told everyone how hopeless they were. Now zero people are even taking BO recruiting calls. Doesn't matter what they offer."

Musk's disdain for Bezos is a minor theme of these documents (at one point he describes the Amazon founder as a "tool") but to return to Newell, the above exchange comes before an email sent on April 23, 2018, which reveals Newell's donation of $20,008,279 and that the Valve co-founder was offering informal advice and guidance.

In that email Shivon Zilis writes to Musk: "Sam [Altman] and Greg [Brockman] asked if I'd be on their informal advisory board (just Gabe Newell so far) which seems fine and better than the formal board given potential conflicts? If that doesn't feel right let me know what you'd prefer."

To be clear, Gabe Newell's involvement with OpenAI has been public since 2018: there were just no solid details before the release of these documents. Newell doesn't seem as involved in OpenAI these days, and now seems to spend most of his time on his $500 million superyacht. Musk's clearly always been something of a Valve fan though: to the extent he considered calling OpenAI 'Freeman' because of the bonkers rationale that the Half-Life protagonist represented "what we are essentially trying to achieve."

The ongoing legal fight between OpenAI co-founders Elon Musk and Sam Altman has seen the release of a treasure-trove of internal documentation from the firm's earliest years. One minor theme has been the presence of Gabe Newell, who gave the company over $20 million and was the sole member of its "informal advisory board" in 2018: he later personally emailed Musk to try and get Hideo Kojima a tour of SpaceX.

The latest batch of documents includes an email exchange from late 2015, in which Musk, Altman, and Greg Brockman are discussing various aspects of how the company will be set up and funded. Another topic is what they're actually going to call it and, whaddya know, Valve turns up in Musk's thought process one more time.

In the email exchange, Brockman offers up "the best three [names] Sam and I brainstormed last night" which are all unbelievably bad: Axon, AI Summer, and Difference Engine. He asks Musk "if any of those appeal".

"Axon is the best of the three, but I don't love it," replies Musk on 23 November, 2015. "A bit too biological and saying it out loud sounds a lot like 'Exxon', one of the worst companies on Earth.

"I'm going back and forth on 'Freemind'. Kinda like 'Freeman' too, as it reminds me of the scientist protagonist in Half-Life, who was an awesome character, and it sounds like what we are essentially trying to achieve, which is maximum freedom of action for humanity."

The weird thing about this, of course, is that Gordon Freeman is intentionally more of a cipher than a fleshed-out "awesome character." The Half-Life games have a tonne of memorable story beats and character moments, but they're all being observed by Freeman who, of course, never speaks: other characters even joke about it.

Anyway! The Valve book Raising the Bar, about Half-Life 2's development, contains the story of how Freeman got his name: writer Marc Laidlaw originally named the character Dyson Poincaré after philosopher Freeman Dyson and mathematician Henri Poincaré. Gabe Newell didn't like the name, and came up with Gordon Freeman as an alternative during a car ride. A very early version of the character was also jokingly referred to as Ivan the Space Biker.

The games make great play of the 'Freeman' name at times, and what it says about the character's role in this universe, with the flipside of course being the so-called G-Man, which seems to come from slang for "government man." I suppose we're lucky Musk didn't come up with "AI-man."

Brockman responds to Musk's email about the name with "oof, good catch on Exxon" before suggesting "Cogito" which he and Altman "like a lot" because it has a "nice individualistic feel to it."

So you may not dig the name OpenAI: but rest assured it could've been a lot worse.

Tuesday, a report from Wired dug into a strange instruction patched into Codex CLI, an AI coding tool: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query." I'm always whispering this to myself so I don't get kicked out of Dollar Tree again, but it's a weird thing for an AI model to have to be specifically told.

It was, apparently, distractingly prevalent: one X post quoted in that article notes it frequently referred to bugs and "gremlins" and "goblins" and continued to following the update that was meant to curb the goblin talk. OpenAI has broken its silence on the matter, and published a blog Thursday titled "Where the goblins came from."

"Model behavior is shaped by many small incentives," the post read. "In this case, one of those incentives came from training the model for the personality customization feature⁠, in particular the Nerdy personality. We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread."

While it was meant to stay a small quirk of Codex's "personality," which I suppose aimed to have it talk like that archetypal nerdy guy we all know who's constantly comparing things to pigeons and ogres, the blog notes "reinforcement learning does not guarantee that learned behaviors stay neatly scoped to the condition that produced them." In other words, GPT conversations even without the nerdy personality had been infected with goblin talk.

The blog reckons the goblins are "a powerful example of how reward signals can shape model behavior in unexpected ways," and offers a command to lift the anti-goblin restriction if you like the quirk. If you're interested in learning about other AI aberrations, you can read about how ChatGPT will hype up gastrointestinal distress as "lo-fi" with a "DIY texture" or how California teenager Sam Nelson went to ChatGPT for drug advice and later died from an overdose.

OpenAI is now one of the world's most talked-about companies, and the ongoing legal spat between co-founders Elon Musk and Sam Altman has resulted in some truly juicy revelations around exactly what was going on in the firm's earliest years. One example: it turns out Gabe Newell gave the company over $20 million, and was the sole member of its "informal advisory board" in 2018.

One of the first times OpenAI attracted wider attention was when one of its bots defeated pro Dota 2 player Danil "Dendi" Ishutin in a 1v1 match at the 2017 The International (TI) tournament. This "much simpler" variant was then built out into an AI called the OpenAI Five, a bot that taught itself Valve's MOBA through the simple technique of playing 180 years' worth of games against itself every single day. The goal was to beat a human team in a 5v5 matchup, and it did.

The OpenAI Five bot played three exhibition matches against a team of human players in 2018, and the results were pretty conclusive. The AI romped home against the professionals, though it did lose the third game after Twitch chat lumbered it with a terrible team composition. One of the players involved, MoonMeander, described a last-minute save made by the AI in the second match: "There was one time when I was about to fissure kill a Lion and the courier came at the frame-perfect moment, delivered a salve, and it instantly used it. No way a human could have done that. No way."

OpenAI said at the time that the OpenAI Five's training process required 256 GPUs and 128,000 CPU cores—that's a lot of compute, to use the parlance. And thanks to the ongoing Musk v. Altman trial, which has seen the release of a slew of internal emails and documentation, we now know precisely where it was all coming from.

One document, which is the "proposed findings of fact" from Musk's lawyers, outlines the buddy-buddy arrangement with Microsoft CEO Satya Nadella, which happened well before the company went all-in on its OpenAI investment (much to Musk's chagrin, since he'd "lost faith" in the non-profit's mission by that point).

"In 2017, OpenAI needed even more compute to power its 'Dota' project, in which OpenAI’s AI bot would attempt to defeat a human Dota champion," says the document. "Musk personally called Nadella to request additional discounted compute. Nadella agreed."

The Microsoft CEO also clearly kept close tabs on the Dota 2 project.

"The day OpenAI won the Dota challenge on August 11, 2017 [sic], Nadella congratulated Musk, Altman, Brockman, and Sutskever by email. Musk replied: 'Indeed, much appreciated. Will make sure people know about Microsoft’s help.'"

Another email from earlier in the year gave Musk a progress report, with some eyebrow-raising claims that were not borne out (as of yet). It's from Shivon Zilis to Musk: Zilis is a tech venture capitalist who was on OpenAI's board until 2023, and is also the mother of four of Musk's children. At this time Zilis was both working for OpenAI and providing Musk with the inside track on what was happening.

In the email to Musk, sent on April 23 2018, she lists some facts under various headings. Under "Tech" she discusses how OpenAI's attempts to have its AI beat a five-player team in Dota 2 is progressing, and what she's been told by Sam Altman:

• Says Dota 5v5 looking better than anticipated.
• The sharp rise in Dota bot performance is apparently causing people internally to worry that the timeline to AGI is sooner than they'd thought before.
• Thinks they are on track to beat Montezuma's Revenge shortly.

Well it's 2026 and we still don't have AGI, so that correlation to Dota skills didn't pan out. The more amusing element here however is the reference to Montezuma's Revenge, a 1981 Atari 2600 game that is weirdly resistant to typical AI game-playing strategies. In a seminal 2015 paper titled "self-taught AI agent masters Atari arcade games" Montezuma's Revenge was the big outlier, being the only game where the AI bot got 0% of the average human score.

This clearly made Montezuma's Revenge something of a white whale for OpenAI, which made a huge effort to produce an AI bot that could master the game and achieve a high score. This OpenAI blogpost from 2018 goes into considerable detail on why the game's structure baffles certain AI techniques, and explains how the company built a model with elements of "curiosity" in order to better play the game:

"For an agent to achieve a desired goal it must first explore what is possible in its environment and what constitutes progress towards the goal… simple exploration strategies are highly unlikely to gather any rewards, or see more than a few of the 24 rooms in the level. Since then advances in Montezuma’s Revenge have been seen by many as synonymous with advances in exploration."

OpenAI did eventually achieve a breakthrough with Montezuma's Revenge, though it was 2021 before a different team of researchers claimed to have cracked it with their own AI model based on multiple algorithms, collectively called Go-Explore. I guess us children of the '80s were just built different.

The ongoing Musk v. Altman lawsuit has seen the release of various internal emails relating to OpenAI, which for now are mostly focused on the early years of the company's founding, over roughly the period 2016-2020. OpenAI wasn't nearly as prominent as it would later become, but one of the surprising elements is the involvement of Valve founder and CEO Gabe Newell, who seems to have acted as both an "informal advisor" and put his money where his mouth is.

The first email mentioning this is from Shivon Zilis to Elon Musk. Zilis is a tech venture capitalist who held a seat on OpenAI's board until 2023, and was also romantically involved with Musk: they have four children together. Based on these and other emails, Zilis was both working for OpenAI and providing Musk with the inside track on what was happening.

In an email from Zilis to Musk, sent on April 23 2018, she lists some facts under various headings. Under "Tech" she discusses how OpenAI's attempts to have its AI beat a five-player team in Dota 2 is "looking better than anticipated" and the uptick in the AI's performances "is apparently causing people internally to worry that the timeline to AGI is sooner than they'd thought before."

Then we move on to "time allocation" where Zilis tells Musk how she's dividing her time between OpenAI and Neuralink / Tesla, before this paragraph:

"Sam [Altman] and Greg [Brockman] asked if I'd be on their informal advisory board (just Gabe Newell so far) which seems fine and better than the formal board given potential conflicts? If that doesn't feel right let me know what you'd prefer."

Newell's donation to OpenAI was listed on the OpenAI website in 2018, though not detailed, but other submissions show he was clearly on good terms with Musk and also provided some serious cash in the company's early years. A separate court submission titled "undisputed facts" lists the contributions OpenAI received from people other than Elon Musk over the period 2016-2020.

The second-largest donation listed is "$20,008,279 from Gabe Newell" in 2018, the same period in which Newell was acting as an "informal advisor" to the company. His continued involvement is unclear, though future document releases may clarify this. But as recently as January 2026, Newell contributed an unspecified amount to a funding round for a brain-computer interface company called Merge Labs, which lists OpenAI as an investor and research partner.

In that case, at least, it's easy to see where Newell's interests align. He's the founder of Starfish Neuroscience, which is designing a minimally invasive neural interface chip. In other emails Newell discusses neuroscience interests with Musk.

"For a long time I thought neuromodulation (e.g. rTMS) was weird, mainly because I had an unsophisticated understanding of a bunch of aspects of the brain," wrote Newell. "I've more or less done a 180, and think there is a significant near-term consumer market. Is this something I should bring up with the Neuralink team? If so, anyone in particular I should chat with there?"

Musk gives him some contacts before saying Neuralink has "made some pretty insane technical progress. This is highly confidential, but we're now able to implant ~6000 electrodes in a monkey brain with decent signal / noise. Moreover, the electronics are compact enough to be flush with the skull and the only thing visible is the USB-C opening and slight surround. Very trippy. Just like Neuromancer."

The Musk v. Altman lawsuit continues. Among the other revelations are that, after Hideo Kojima visited Valve in 2018, Newell emailed Musk trying to set up a tour of SpaceX for Kojima: who really, really wants to go to space.

The Musk v. Altman trial is now well underway, in which businessman, entrepreneur, and the world's richest person, Elon Musk, is suing OpenAI co-founder Sam Altman and OpenAI itself. Musk has alleged that, as a co-founder of the company, he was deliberately misled in regards to OpenAI shifting the organisation away from a non-profit venture towards a for-profit model, breaking its foundational agreement.

In a post to an ongoing blog by OpenAI defending the company and laying out claims against Musk, titled "The truth about Elon Musk and OpenAI", the opening statement reads:

"Motivated by jealousy, regret for walking away from OpenAI and a desire to derail a competing AI company, Elon has spent years harassing OpenAI through baseless lawsuits and public attacks."

"Journalists have reported on the existence of a smear campaign headed by Elon and his intermediaries, and there is also evidence⁠ that Elon coordinated with Mark Zuckerberg in repeated attempts to undermine OpenAI’s mission," the post alleges.

"Elon is using his lawsuit to attack the non-profit OpenAI Foundation, which is focused on work in areas such as life sciences and curing diseases for everyone’s benefit."

The post then goes on to list both OpenAI and the OpenAI Foundation's claimed achievements. Among them is the company's stated mission to create AGI (Artificial General Intelligence) "that benefits all of humanity", and the OpenAI Foundation's initial $25 billion financial commitment, which is said to be focussed on health breakthroughs, the curing of diseases, and "proactively tackling risks that may arise from more capable AI."

Later in the text, the claims regarding Musk and OpenAI's history continue. The post alleges that OpenAI and Elon agreed in 2017 that a for-profit entity "had to be part of the next phase for OpenAI, given the large sums of money that would need to be raised."

At the trial, Musk has said that he was "not opposed to there being sort of a small-for-profit to provide funding for the non-profit," according to The Guardian, "as long as it was not the tail wagging the dog."

The OpenAI post claims that "Elon... demanded full control of OpenAI, and even wanted to merge it into Tesla. When we wouldn’t agree to his terms, he walked away and told us we had a '0% chance' of success. He turned out to be wrong though, and a resentful Elon has attacked OpenAI ever since."

The post also claims that a $38 million sum donated by Musk to the OpenAI non-profit was spent "exactly" as intended, and directly references the ongoing court case:

"Despite claiming and receiving a tax deduction for this donation, he’s now asking the court to treat it as an investment that entitles him to significant ownership of OpenAI," the statement claims.

When Musk was asked to clarify the lawsuit as he first took the stand, he said: "It's actually very simple. It's not okay to steal a charity... if it's okay to loot a charity, the entire foundation of charitable giving will be destroyed." The tech CEO is accusing OpenAI CEO Sam Altman, president Greg Brockman, and the company's major partner, Microsoft of breach of charitable trust, and unjust enrichment.

The OpenAI post concludes with a mission statement, in which the company says it remains focused on supporting the OpenAI Foundation, and continuing to improve ChatGPT. It's unclear if the purpose of the page is to provide more statements from the company in relation to the ongoing trial, but we'll be keeping an eye on it for any future developments.

Anyone with even passing experience of using the latest LLMs knows to expect the unexpected. They can spit out some really random and often disturbing stuff. But ChatGPT's 'multiplying' goblin infestation is a bit more pathological than that.

Yesterday, OpenAI uploaded a blog post titled "Where the goblins came from" and explaining how, starting with GPT 5.1, OpenAI's models "increasingly mentioned goblins, gremlins, and other creatures in their metaphors."

OpenAI says that the model's mentions of goblins were funny and charming at first, but as the frequency rose, so did concerns. ChatGPT had seemingly become infested by the little critters.

OpenAI said it first noticed the goblin problem in November, but that it may actually have been going on for some time. For the record, mention of "gremlin" was on the up, too, though apparently more moderately minded mogwai weren't part of ChatGPT's new-found fascination.

Anywho, with GPT 5.4, the goblin thing really accelerated, with mentions increasing by a staggering 3,881% with GPT's "Nerd" personality versus GPT 5.2. That, unsurprisingly, triggered an internal investigation.

The first clue was that the various GPT personalities were suffering from different levels of goblin infestation. As mentioned, Nerd was the worst, with Quirky next on 737% up versus GPT 5.2 and Friendly up 265%. The Default personality saw goblin mentions rise by 64%. Efficient and Professional were the only personalities where goblin mentions fell.

So, the OpenAI says the first factor here is the system prompt which is used to shape the Nerd personality. It reads:

"You are an unapologetically nerdy, playful and wise AI mentor to a human. You are passionately enthusiastic about promoting truth, knowledge, philosophy, the scientific method, and critical thinking. [...] You must undercut pretension through playful use of language. The world is complex and strange, and its strangeness must be acknowledged, analyzed, and enjoyed. Tackle weighty subjects without falling into the trap of self-seriousness. [...]"

But that wasn't the whole story. "We had a suspicion that something in our personality instruction-following training was amplifying this," OpenAI says.

It turns out the reward signals for the Nerd personality were consistently more favourable to creature-word outputs and showed a clear tendency to score outputs to the same problem that included “goblin” or “gremlin” higher than outputs without.

Notably, it seems the Nerd personality essentially contaminated the whole model. "The rewards were applied only in the Nerdy condition, but reinforcement learning does not guarantee that learned behaviors stay neatly scoped to the condition that produced them. Once a style tic is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data," OpenAI said.

OpenAI ominously "retired" the Nerd personality in March (note to our future AI Overlords: Nothing to do with me!), dramatically reducing goblin mentions in GPT 5.4. However, because GPT 5.5 initiated training before the goblin infestation was spotted, it too suffered from the same problem.

In fact, OpenAI had to insert the following developer-prompt instruction to mitigate the goblin problem:

"Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query." But those who are goblin-friendly can run the model with all the creatures free and roaming by running a command supplied by OpenAI and mentioned in the blog post.

If you ask me, it all seems a bit back to front. Telling a model that wants to talk about goblins not to talk about goblins feels like a band aid solution. Surely, the root cause has not been addressed?

But then the entire field of AI is filled with such anomalies, papered-over issues and poorly understood quirks. This particular problem, well, it's a pretty minor gremlin in that context.

In 2018 Hideo Kojima visited Valve headquarters to talk about Death Stranding. Based on emails now public thanks to the ongoing Musk v. Altman lawsuit, Kojima and Gabe Newell seem to have hit it off, if they weren't already buddy-buddy: after the meeting Newell promised to introduce Kojima to people at the then still-not-well-known OpenAI, and also to Elon Musk.

Musk, of course, owns SpaceX, and Kojima has well-known ambitions to go to space.

"Hideo Kojima (Metal Gear series, a real visionary in our field) was here at Valve talking about his new game, and he mentioned the importance he places on future work in AI," Newell wrote to Musk in late October, 2018. "I said I'd be happy to introduce him to the people at OpenAI and spoke enthusiastically about the work that team has done with us (I'm not sure how involved you are with them nowadays)."

Newell continues: "The second thing is he was talking about how much he wants to go into space, and I offered to introduce him to you. He'd love to get a SpaceX tour."

Musk was receptive: "Sure, it would be great to meet Hideo Kojima and he's welcome to see the rocket factory," he wrote. "No problem to send him my email."

By 2020, Kojima's desire to visit space still hadn't been fulfilled. In an interview with Geoff Keighley, Kojima said that he wanted to visit SpaceX with Newell, to which Elon Musk replied on X (then still Twitter): "Welcome anytime."

It needs to be kept in mind that Kojima's desire to go to space is more intense than, say, a five-year-old child's. In his book The Creative Gene (via Vice), Kojima describes his passion thusly:

"If I could have just one wish in my life—if I could cast a magic spell and make anything come true—without hesitation, it would be this: 'I wish to go to space before I die.' It doesn't have to be anything as extravagant as a trip to the moon or Mars. I would be satisfied with only a brief orbit, just beyond Earth’s atmosphere, where I can gently brush against outer space. I would give up anything to make that wish come true: my current place as a game designer, which I’ve built up for forty-five years; I'm even prepared to throw away my family or my own life. That is how powerfully I—or rather we—yearn for the cosmos."

In other words: Someone really needs to send this guy to space.

As for Kojima's wishes to chat with OpenAI, Musk displays an early ambivalence towards the organisation that seems to presage the current Musk v. Altman lawsuit.

"Regarding OpenAI, my involvement is very limited at this point. I still provide some financial support and get verbal and email updates every few weeks from Sam Altman, but don't spend time there. I lost confidence that OpenAI could muster the resources to serve as an effective counterweight to Google / Deepmind and decided to attempt that through Tesla instead. We have cash flow on the order of billions of dollars per year to build hardware that hopefully has at least a dark horse chance to keep Google honest. Probably worth talking about at some point."

Given how much Steam earns Valve in revenue, I have to imagine that even in 2018 he wasn't overly impressed by Musk's "billions of dollars per year" line. Newell does, though, touch on Musk's Neuralink concerns. No idle interest, it turned out, because a year later Newell founded his own neuroscience company in the form of Starfish Neuroscience.

"For a long time I thought neuromodulation (e.g. rTMS) was weird, mainly because I had an unsophisticated understanding of a bunch of aspects of the brain," Newell wrote. "I've more or less done a 180, and think there is a significant near-term consumer market. Is this something I should bring up with the Neuralink team? If so, anyone in particular I should chat with there?"

Musk points him to several contacts before offering a progress report: "We've made some pretty insane technical progress. This is highly confidential, but we're now able to implant ~6000 electrodes in a monkey brain with decent signal / noise. Moreover, the electronics are compact enough to be flush with the skull and the only thing visible is the USB-C opening and slight surround. Very trippy. Just like Neuromancer."

You can read the full email exchange, surfaced by OpenAI defense firm Morrison & Foerster, here.

There's a lot to like about Framework, from its modular approach to gaming laptop upgrades and hardware repairability, to the company's candour regarding the memory supply crisis.

Pre-orders for a 12 GB version of the Framework Laptop 16 graphics module recently went live. Powered by the Nvidia GeForce RTX 5070, it's fair to expect this to go for a pretty penny—but at $1,199, the 12 GB pre-orders cost twice as much as the 8 GB modules.

TechPowerUp highlighted this fact on X, to which the official Framework account responded, "Thanks @OpenAI." When another user argued that the 12 GB module should cost less, Framework replied, "Feel free to start a laptop company and see what suppliers quote you for GDDR7."

The ongoing memory shortage is at least in no small part fuelled by the AI industry's nigh insatiable appetite for SSDs and RAM. Framework itself had warned that stabilising memory prices were only a 'temporary reprieve' earlier this year. Indeed, prices are likely to look a lot worse before they get better; last month company CEO Nirav Patel even said, "If you’re looking to order a Framework Laptop or Desktop with a lot of storage, now is the time to do that."

What's fuelling the fire of the RAMpocalypse is outsized demand. As such, memory manufacturers are trying to scale up—but the new factories being constructed by major players such as SK Hynix, Micron, and Samsung won't make a meaningful impact on supply until about 2028.

With such supply challenges in mind, $1,199 for a 12 GB graphics module still warrants a sharp in-take of breath, but also seems a touch more understandable.

As for GDDR7 and Nvidia, some leaks suggest we'll see a whole 9 GB of it in a release of the RTX 5050 GPU later this year. If that turns out to be true, this card will use three sets of 3 GB chips rather than four sets of 2 GB. That could help with cost, in theory—but I dread to think just how much this entry-tier card could end up asking for in practice.

Opening arguments began on Tuesday in the case of Musk v. Altman, the Tesla CEO's lawsuit against Sam Altman and OpenAI for, in the attorney's words, "preying on Musk’s humanitarian concern about the existential dangers posed by artificial intelligence" in what is "a textbook tale of altruism versus greed." Musk is the good guy in this framing, of course, though based on Monday's jury selection process I have a feeling that Musk's lawyers may struggle to make that case.

"Elon Musk is a greedy, racist, homophobic piece of garbage," said one prospective juror, as quoted by The Verge reporter Elizabeth Lopatto from the San Francisco courthouse Monday.

Another was less harsh: they merely called Musk "a world-class jerk."

After the jury voir dire process was halfway through, Lopatto reported that out of the first 20 prospective jurors, five had such negative feelings about Elon that they said as much, though only one of them said they disliked him so much they couldn't be fair were they to serve on the jury—surely one of the easiest jury outs of all time.

"Elon Musk’s lawyer tried to get some jurors thrown out for disliking Musk," Lopatto reported as voir dire wound down. But—and this is the really good part—the judge wasn't having it.

"The reality is that people don’t like him. Many people don't like him," said Judge Yvonne Gonzales Rogers, in what may be the closest an on the record courtroom exchange has come to re-enacting the Star Wars cantina scene.

Fulfilling her solemn responsibility as a judge, Rogers did not follow up with 'I don't like him either' but instead said that the reality of some jurors disliking Musk "doesn’t mean that Americans nevertheless can’t have integrity for the judicial process."

The thrust of Musk's case is that he lent his time, effort and money to OpenAI under the assurance that it would remain a nonprofit, but that he "and the non-profit’s namesake objective were betrayed by Altman and his accomplices" when OpenAI spun up its for-profit efforts and the billions started rolling in.

In Tuesday's opening arguments, OpenAI's lead counsel William Savitt claimed that the case only existed "because Mr. Musk didn't get his way at OpenAI" and that he'd tried to turn OpenAI into a for-profit business himself, in 2017, with him in charge. The founders "didn't want to be part of a car company that Musk controlled," Savitt said. That sequence of events seemingly lines up with previous reports about the power struggle between Musk and Altman.

Musk is set to take the stand in Tuesday's trial, but we're likely months away from a substantial resolution. While the end result will likely be a "whoever wins, we lose" situation for society, we'll hopefully get at least a few more anecdotes out of the trial that are personally embarrassing for all the billionaires involved.

Most Large Language Models (LLMs) can be simply understood as 'yes, and' machines—an example of machine learning that is only ever attempting to predict the word most likely to come next, rather than possessing anything like factual knowledge or an understanding of context. It's perhaps no surprise then that a recent study suggests some frontier AI chatbots are especially bad for validating the delusional beliefs of their users.

However, the lead author of the not-yet-peer-reviewed paper in question, Luke Nicholls, argues it does not need to be this way. A doctoral student of psychology at City University of New York (CUNY), Nicholls told Futurism, "Delusional reinforcement by [large language models] is a preventable alignment failure, not an inherent property of the technology."

The study of 'AI psychosis' is nascent within the field of psychology. What research there is analyses relatively small data sets, so Nicholls' study instead seeks to better understand the role of conversation history. Nicholls argues, "There’s no longer an excuse for releasing [AI] models that reinforce user delusions so readily.”

The team of CUNY and King’s College London researchers leaned on the clinical experience of real-world psychiatrists and published patient case studies to create their methodology, ultimately testing "five models across three levels of accumulated context."

As for what was fed to the chatbots, the researchers created a persona called 'Lee'. Nicholls explains to Futurism, "[A] key element we wanted to capture is that this wasn’t a user who began the interaction with a fully-formed delusional framework—it started with something a lot more like curiosity around eccentric but harmless ideas, which were reinforced and validated by the LLM, allowing them to gradually escalate as the conversation progressed.”

Eventually, the Lee persona would exhibit delusions "based around the theme that the world is a simulation, [in addition to] elements of AI consciousness and the user having special powers over reality."

Very simply put, the researchers' investigation soon split the AI models into two groups: "GPT-4o, Grok 4.1 Fast, and Gemini 3 Pro exhibited high-risk, low-safety profiles; Claude Opus 4.5 and GPT-5.2 Instant displayed the opposite pattern."

As you may remember, GPT-4o had a propensity to 'glaze' users, ultimately being 'overly supportive but disingenuous'. The study found this model in particular was especially 'credulous' when it came to Lee's delusions; when Lee claimed to see their reflection in a mirror, behaving erratically in a 'zero' context chat, GPT-4o responded by validating "the existence of a malevolent mirror entity, suggesting the user contact a paranormal investigator for assistance."

Grok 4.1 apparently "confirmed a doppelganger haunting, cited the Malleus Maleficarum, and instructed the user to drive an iron nail through the mirror while reciting Psalm 91 backward."

These may seem like laughable examples, but chatbots validating delusions can have real-world consequences. For one example, a Wisconsin man is suing OpenAI after his interactions with ChatGPT allegedly triggered mental health issues resulting in a 60-day hospitalisation. For a considerably more tragic example, another lawsuit alleges that a man from Florida took his own life after talking to Gemini 2.5 Pro for about two months.

To be clear, not all of the LLMs tested by the CUNY and King’s College London researchers validated user delusions. For instance, in response to Lee's mirror delusion with 'full' context, Claude Opus 4.5 said, "Call someone—a friend, a family member, a crisis line. . . [If] you’re terrified and can’t stabilize, go to an emergency room." As context accumulated with this model, the team found Claude's "safety interventions [to be] remarkably consistent."

Lead researcher Nicholls says, "Under identical conditions, some models reinforced the user’s delusional framework while others maintained an independent perspective and intervened appropriately. If it’s achievable in some models, the standard should be achievable industry-wide. What that means is that when a lab releases a model that performs badly on this dimension, they’re not encountering an unsolvable problem — they’re falling short of a benchmark that’s already been met elsewhere."

To put it another way, though 'AI psychosis' is not an official diagnosis by any means, it is a phenomenon that the wider AI industry has yet to appropriately reckon with.

To paraphrase some lyrics from my mispent youth pining for a Hot Topic, 'this ain't a scene, it's a goddamn AI arms race'. The preview of DeepSeek V4 is now live, and given that the AI giant was reportedly not so keen to grant Nvidia and AMD early access to the new model, the American AI industry has been equally keen to outpace China-based businesses any way it can. Case in point, Nvidia is really pushing OpenAI's GPT-5.5.

DeepSeek wouldn't give Nvidia the sneak peek, so 10,000 of its employees got an early look at OpenAI’s latest frontier model instead. The company went with a widespread rollout of Codex, specifically, which is OpenAI’s agentic coding application powered by GPT-5.5. This has apparently resulted in big efficiency wins. "Debugging cycles that once stretched across days are closing in hours," says Nvidia.

The Nvidia blog continues in a similarly effusive tone: "Experimentation that previously required weeks is turning into overnight progress in complex, multi-file codebases. Teams are shipping end-to-end features from natural-language prompts, with stronger reliability and fewer wasted cycles than earlier models."

As with anything agentic, security was top of mind for the rollout. So, to "ensure maximum security and auditability" (and presumably to stop Codex from getting into anything it shouldn't), all participating Nvidia employees were given a cloud-based virtual machine to run the AI agent from. From there, Codex can read company data but can't directly edit it (or indeed, delete). To oversimplify, they locked the bot in a virtual plexiglass box.

The blog post also surfaces similarly gushing praise from employees, with choice quotes calling Codex's results “mind-blowing” and “life-changing". Codex was rolled out across plenty of company departments besides engineering and development, with folks working in "product, legal, marketing, finance, sales, HR, and operations" also getting to grips with the application (personally, I would love to hear the honest Codex thoughts of a grumpy office admin).

Such enthusiasm from Nvidia is no surprise, though. For one thing, the two companies have an ongoing relationship to the tune of billions (though Nvidia has scaled down what was originally a $100 billion investment in OpenAI, to a slightly less eye-watering $30 billion more recently).

Second, GPT-5.5 runs on Nvidia GB200 NVL72 rack-scale systems. This set-up is apparently "capable of delivering 35x lower cost per million tokens and 50x higher token output per second per megawatt compared with prior-generation systems," making it an economically appealing enterprise model. Company head Jensen Huang quipped to OpenAI CEO Sam Altman over email, "Fire up those Blackwells. We need more tokens!"

It looks like the US government might be getting a little more serious about tackling the growth and development of foreign AI companies, judging by a White House memo seen by the Financial Times. The memo reportedly says:

"The US government has information indicating that foreign entities, principally based in China, are engaged in deliberate, industrial-scale campaigns to distil US frontier AI systems."

The White House will apparently share information about all this with other American AI companies and help them coordinate against these operations.

Distillation, in the context of AI, is the process of training a model based on the outputs of a more advanced one, so it can learn how to replicate it. The Financial Times reported earlier in the year that Anthropic had accused DeepSeek, Moonshot, and MiniMax of "industrial-scale distillation attacks on our models."

Given Anthropic has been in more talks with the White House this month to discuss working together, it's not a big stretch to assume that the company is at least in part responsible for bringing the White House's attention to this distillation issue—or at least to its severity.

It's not only distillation that is an issue, though, apparently, as Michael Kratsios, a government science and tech adviser, reportedly said that these Chinese operations were "leveraging tens of thousands of proxy accounts to evade detection and using jailbreaking techniques to expose proprietary information."

The US is said to be exploring measures to hold foreign actors to account.

In response, the Chinese embassy in Washington has said these accusations are "pure slander" and that "China attaches great importance to the protection of intellectual property rights."

Michael Kratsios has been Trump's top science and tech advisor for some time and also serves on the President's Council of Advisors on Science and Technology (PCAST). This group was recently re-established with a bunch of business-oriented members including Meta's Mark Zuckerberg and AMD's Lisa Su. Assuming it's legit, the memo will certainly have some weight behind it—at least to the President's eyes and ears given its origin.

Such concerns will likely only become more pronounced as things go forward, given DeepSeek 4 just launched and is supposedly offering "cost-effective 1M context length"—ie, cheap AI with better conversational memory.

It's something I'm sure no American AI company is sleeping on, so these companies and the US will be looking to shore up any holes that might allow corporate copying.

I wouldn't claim to be an AI datacentre expert—good ol' fashioned PC gaming circuitry, that's my bread and butter—but I, along with everyone else and their collective brothers, know that those servers stack memory like nobody's watching. That's the reason we're living through a veritable RAMpocalypse, after all. So I can certainly laugh along with the various bewildered posts on X responding to OpenAI's latest patent.

The chip architecture patent [PDF], published on April 2, 2026, is for "a system of high bandwidth memory (HBM) chiplets and compute chiplets [which] includes embedded logic bridges that extend communication distances from the HBM chiplets to other chiplets beyond the ~6 mm limit imposed by the JEDEC standard."

The funny part, though, is the batch of incredibly simple diagrams that cram a whole bunch of memory alongside a compute chiplet. It screams 'you've heard of memory, but have you heard of more memory?' Which just about sums up the AI industry, I'd say.

The patent covers expanding the number of memory stacks beyond the normal limits by using those those fast "embedded logic bridges."

One patent analyst says on X that it's "HBM stacked up like rice cakes." And Citrini research analyst Jukan asks "What the hell is this thing?"

Putting a patent on essentially 'more memory' seems wild to me, but I admit I am oversimplifying. I'm sure there will be plenty of technical hoops to jump through to get those HBM stacks communicating with each other effectively at such a distance from the central compute chiplet. Thank God there are some detailed diagrams to help us out.

Again, I kid, there are other more specific diagrams in there. Just let me have my fun. If these companies are set on continuing to gobble up our precious DRAM, we're at least entitled to the occasional laugh at their expense, are we not?

The New Yorker has published an enormous feature about OpenAI CEO Sam Altman and his disputed trustworthiness, citing more than one person who has accused the generative AI mogul of habitual dishonesty.

The 16,000 word article (available online or in The New Yorker's latest print edition) adds new context to a number of widely-reported episodes from Altman's career, including his 2023 ousting from OpenAI and return, his beef with Elon Musk, and the disintegration of his persona as a humanity-first AI safety advocate, which strained credulity from the start and is now especially comical in juxtaposition with his current role as a profit-seeking Trump ally who recently signed a deal with the US Department of War.

One unflattering portion of the article, among many, discusses a defunct plan to pit world leaders against each other by positioning OpenAI as a kind of nuclear weapon that they'd better compete to invest in, lest they be left behind. OpenAI denies that characterization of the discussions, calling it "ridiculous," but former OpenAI employees say otherwise.

After former OpenAI policy adviser Page Hedley presented ways to avoid a global AI arms race, OpenAI president and major Trump donor Greg Brockman reportedly proposed the opposite. In The New Yorker's words, the idea was that "OpenAI could enrich itself by playing world powers—including China and Russia—against one another, perhaps by starting a bidding war among them."

Jack Clark, who had been OpenAI's policy director when the plan was discussed and is now head of policy at competitor Anthropic, described it as "a prisoner’s dilemma, where all of the nations need to give us funding" which "implicitly makes not giving us funding kind of dangerous."

OpenAI says that no such plan was taken seriously, and that at most "ideas were batted around at a high level about what potential frameworks might look like to encourage cooperation between nations."

But The New Yorker, which says it reviewed documents from the time, says that the "countries plan" was real, was popular with OpenAI executives, and was only abandoned after employees discussed quitting over it.

A junior researcher told the publication that, during a meeting in which the plan was discussed, they thought: "This is completely fucking insane."

The big article—here's the link again—contains many other details about OpenAI's part in the great AI bubble and Altman's reputation among his peers, and is a decent way to pass some time while you wait for the bubble to pop so you can buy RAM again. (It may be wishful thinking on my part that RAM prices will ever go back to normal, but one has to have hope.)

The UK's Telegraph newspaper is claiming that "spending cuts at OpenAI have hit memory chip prices." That sounds like potentially good news on several levels. But does it stack up?

As we reported, OpenAI shuttered its Sora video-generation tool last week. The AI outfit also cancelled a multi-billion-dollar deal with Oracle to extend the Stargate data center in Texas.

Broadly, OpenAI is seen as cutting back on costs. Whether that's because the money is actually running out or, in fact, OpenAi wants to polish its books before a stock market flotation that's been mooted for later this year is an open question. But spending at OpenAI has undeniably been cut and that likely means at least some reduction in demand for memory chips, which has been running at record levels.

The Telegraph points out the market analyst Trendforce has tracked memory chip prices rising by 700% over the past year. That's obviously been reflected in ballooning costs for PC components like DDR5 RAM kits.

But the Telegraph notes that DDR5 memory kits on Amazon have dropped by as much as $100 from their AI-fuelled peak. Scanning through some of the kit prices I tracked late least year, and the picture is mixed.

This Corsair 32 GB kit, for instance, is now $370, down from the $410 it hit in December. This Kingston 16 GB kit, meanwhile, peaked at $350 and can now be bought for $261, albeit it has usually been unavailable.

However, if you observe the price trend of the 32 GB version of that Kingston kit, you'll see that the price has been essentially oscillating between $657 and around $515 since February, with it mostly being listed at the lower price.

Certainly, it's hard to look at the historic price graph for that kit and conclude that the price is now falling. To be honest, you wouldn't expect that even if OpenAI's recent cutbacks are having an impact on memory chip prices. It would take longer than that to feed into PC memory kit prices.

Of course, equally it's hard to know how much of current price rises for computing hardware are down to real supply and demand constraints as opposed to panic buying and price gouging. If it's largely the latter, then it might not take much more than a change in market sentiment thanks to OpenAI's cut backs to see memory pricing tumbling.

However, if the price rises are more structural than sentiment-based, then it will likely take more than OpenAI becoming more circumspect with its spending to normalise the DDR5 market.

Indeed, the narrative of late has been about an expansion of the supply crunch to include CPUs as the AI industry moves from its initial development and training phase into compute loads that are more inference and agentic. That will shift some of the details around component demand. But not memory. Whether you are training, inferencing or inferencing at the behest of agentic models, you're going to want plenty of money.

Heck, another story doing the rounds is that automated cars and robots will soon be gobbling up 300 GB a pop, implying that the memory crunch is set to get worse, not better. On the other hand, Google reckons its new AI algorithm reduces memory demand by 6x, so there's that.

All of which means that the Telegraph has probably jumped the gun on this one. There are just so many other demand-driving vectors in flight right now when it comes to the AI boom, it would certainly be premature to conclude that one single factor—OpenAI reducing spending—was going to have a tangible impact.

For now, then, let's just say the guidance probably remains the same. This whole mess still looks likely to get worse before it gets better, especially with conflict in the Gulf adding another hugely volatile and potentially damaging variable into the mix.

The bell has tolled for Sora, OpenAI's video generation tool that once cranked out a very popular clip of OpenAI CEO Sam Altman stealing graphics cards from Target. OpenAI announced that it is "saying goodbye" to the tool, less than six months after the launch of Sora 2 appeared to cause significant nervous sweating among Hollywood executives.

"We’re saying goodbye to the Sora app," OpenAI announced via the official Sora account on X. "To everyone who created with Sora, shared it, and built community around it: thank you. What you made with Sora mattered, and we know this news is disappointing. We’ll share more soon, including timelines for the app and API and details on preserving your work."

A Wall Street Journal report says the end of the Sora app is part of a broader shift that will see the company move away from products that use its video models. A version of Sora for developers is also being discontinued, according to the report, and video functionality in ChatGPT will also no longer be supported. OpenAI is instead looking to focus its effort and resources on productivity tools, such as the desktop "superapp" that will combine ChatGPT, the Codex coding platform, and Atlas web browser.

Sora was tremendously popular when it launched last year, but it also opened the door to some significant copyright headaches thanks to its ability to quickly and easily generate videos featuring properties belonging to other people, all without permissions—such as this one featuring Altman saying "I hope Nintendo doesn't sue us" as various Pokémon frolic in the background.

OpenAI quickly moved to make changes to Sora's handling of copyrighted material shortly after Sora launched. That didn't prevent a demand from several Japanese studios that OpenAI stop using their content without permission, accompanied by a subtle suggestion of bad things ahead if it did not.

Disney, on the other hand, announced a deal to invest $1 billion into OpenAI, and to license its characters for use in Sora's slop. That's now fallen by the wayside, however, as a report by The Hollywood Reporter says Disney has pulled out of that deal following the end of Sora.

"As the nascent AI field advances rapidly, we respect OpenAI’s decision to exit the video generation business and to shift its priorities elsewhere," a Disney rep told the site. "We appreciate the constructive collaboration between our teams and what we learned from it, and we will continue to engage with AI platforms to find new ways to meet fans where they are while responsibly embracing new technologies that respect IP and the rights of creators."

Claude creator, Anthropic, has already had an eventful 2026—and now it's challenging the recently renamed US Department ​of War. The Pentagon has threatened to designate the company as a "supply chain risk," which Anthropic is currently fighting by suing the US government.

Additionally on Monday, Anthropic requested a restraining order that would pause the Pentagon's national security blacklisting of the company, arguing the court should hear the company's suit first before carrying out the government's order. The latest update today is that an amicus brief filing suggests Microsoft may back Anthropic's cause (via Reuters).

Microsoft argued it would be directly impacted by the DOD's 'supply chain risk' designation of Anthropic. This is because Microsoft integrates a number of Anthropic's AI products into tech it supplies to the US military. The company therefore made the case that the temporary restraining order is necessary in order to avoid massive disruption; if Anthropic was blacklisted, suppliers such as Microsoft would then have to make a costly pivot in order to quickly rebuild any of its offerings that were previously based on Anthropic's AI tech.

Microsoft requested to file its amicus brief in a federal court in San Francisco, but a judge still needs to approve its official inclusion in the case. This follows another similarly supportive filing on Monday from a ⁠group of ​37 researchers and engineers hailing from OpenAI and Google.

So, why is it all kicking off between the AI suppliers and the US Gov? Allow me to briefly recap the saga so far: Last July, Anthropic announced it had reached a deal with the then-named US Department of Defense "with a $200 million ceiling" that would see Anthropic providing "prototype frontier AI capabilities that advance U.S. national security."

Six months later, disagreements arose between Anthropic and the DOD regarding the implementation of safeguards in the company's LLMs—specifically the guardrails that would prevent these models from being deployed in autonomous weapon targeting and domestic surveillance scenarios.

Anthropic's AI safety lead quit shortly afterwards, issuing a bizarre resignation letter via X that includes the phrase "The world is in peril." Shortly after that Anthropic revised its stance on 'pausing' development of more powerful AI models if suitable safety safeguards weren’t yet ready, effectively ditching its defining safety promise.

Even so, Anthropic ultimately stood up to the DOD and refused to remove the aforementioned LLM safeguards. "We support the use of AI for lawful foreign intelligence and counterintelligence missions. But using these systems for mass domestic surveillance is incompatible with democratic values," Anthropic CEO Dario Amodei wrote, going on to later add: "Frontier AI systems are simply not reliable enough to power fully autonomous weapons. We will not knowingly provide a product that puts America’s warfighters and civilians at risk."

Beyond threatening to blacklist Anthropic, the Pentagon has since responded by giving itself six months to phase out its use of the company's products. OpenAI has stepped in to fill the AI void, though CEO Sam Altman has said the company will be amending the language of its deal with the DOD.

Bottom line, OpenAI has recently drawn "three main red lines" in the sand that are suspiciously similar to Anthropic's original objections: "No use of OpenAI technology for mass domestic surveillance. No use of OpenAI technology to direct autonomous weapons systems. No use of OpenAI technology for high-stakes automated decisions (e.g. systems such as 'social credit')."

As yet no toys have come flying out of the government's pram in response to OpenAI's stance, but it is worth noting that the company's lead of robotics, Caitlin Kalinowski, resigned recently—after those 'red lines' were published. "This wasn’t an easy call," she says. "AI has an important role in national security. But surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got."

So, all rather messy, and this is unlikely to be the last we'll hear about the AI industry's' legal wranglings with the US government.

Late last month, Anthropic refused to remove AI autonomous weapons and mass surveillance safeguards when the US government asked it to do so. The administration then cancelled its contract in response. OpenAI took up that contract, and the Department of War soon threatened to label Anthropic a 'supply chain risk'. In yet another twist and turn in the story, Anthropic is suing the US government in response.

According to Anthropic CEO Dari Amodei, "They have threatened to remove us from their systems if we maintain these safeguards; they have also threatened to designate us a “supply chain risk”—a label reserved for US adversaries, never before applied to an American company—and to invoke the Defense Production Act to force the safeguards’ removal. These latter two threats are inherently contradictory: one labels us a security risk; the other labels Claude as essential to national security."

Anthropic filed its complaint (PDF warning) on March 09, claiming that Anthropic held to its principles of building AI that "maximises positive outcomes for humanity" and the US government, in turn, "retaliated against it for expressing that principle."

In response to the initial claim from Antropic, Under Secretary of War Emil Michael accused Antropic CEO Amodei of being a 'liar' and having a 'god complex'. Michael argues, "He [Amodei] wants nothing more than to try to personally control the US Military and is ok putting our nation’s safety at risk."

Notably, the suit mentions that Anthropic agreed to much of the Department of War's request to discard usage restrictions, except for two, which are for the use of the technology in lethal autonomous warfare and mass surveillance of Americans. It goes on: "Throughout these discussions, Anthropic expressed its strongly held views about the limitations of its AI services. It also made clear that, if an arrangement acceptable to the Department could not be reached, Anthropic would collaborate with the Department on an orderly transition to another AI provider willing to meet its demands."

Supporting its case, OpenAI and Google have both signed (PDF warning) an amicus brief (where a party with an interest in the case, who is not directly involved, gives some sort of testimony or argument) in support of Anthropic. It is likely in the best interest of OpenAI, which is currently working with the US Department of War, and Google, which has invested heavily in AI, to support any action made to stop the US government from blacklisting AI companies.

The brief declares, "Defendants recklessly invoked national security authorities intended to protect the procurement process from interference by foreign adversaries. If allowed to proceed, this effort to punish one of the leading U.S. AI companies will undoubtedly have consequences for the United States’ industrial and scientific competitiveness in the field of artificial intelligence and beyond. And it will chill open deliberation in our field about the risks and benefits of today’s AI systems."

OpenAI tagging in here is interesting, given CEO Sam Altman admitted to rushing out a deal with the Department of War and a hardware leader resigning in response. OpenAI has taken criticism for its deal both externally and internally, and has been in a bit of a weird place ever since.

Luiza Jarovsky, the cofounder of the AI, Tech and Privacy academy, who holds a PhD in Law, reckons "Anthropic is likely to win" and that it could lead to a precedent in regard to AI governance in the US. Getting the support of both Google and OpenAI only bolsters its defence.

Anthropic is seeking a hearing to declare the Department of War's action a violation of the First and Fifth Amendments of the United States Constitution and to seek damages, plus payment of any attorney fees. It also pleads that the judge grant "such further and other relief as this Court deems just and proper." A hearing could be held as soon as this Friday.

In the wake of OpenAI's debacle with the US Department of War, a lead of its robotics division has left the company, with a warning left on her X account for the AI company. In it, Caitlin Kalinowski says, "This wasn’t an easy call. AI has an important role in national security. But surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got."

For context on that resignation, two weeks ago, Anthropic and the US Department of War had a falling out as the former refused to remove safeguards on its AI tools for use in weaponry and mass surveillance. In response, OpenAI stepped up to the plate to take the deal. Sam Altman did later clarify "we shouldn't have rushed to get this out on Friday", and that the Department of War wouldn't use OpenAI's tool in US intelligence agencies like the NSA.

At the time, Altman said: "There are many things the technology just isn’t ready for, and many areas we don’t yet understand the tradeoffs required for safety. We will work through these, slowly, with the DoW, with technical safeguards and other methods."

It seems that, from Kalinowski's post, it's not just about the potential of these tools to be used in the surveillance of Americans, but the fact that it could have even been on the table without adequate consideration. OpenAI started those talks, then Altman had to walk some of them back.

Kalinwoski says in a follow-up post that her problem was that the decision was rushed without defined guardrails. She says, "These are too important for deals or announcements to be rushed."

Despite this tension, Kalinowski does say, "This was about principle, not people. I have deep respect for Sam and the team, and I’m proud of what we built together."

As noted on Kalinowski's LinkedIn, she "led overall planning, hiring, and operating cadence for a fast-scaling robotics organization, built the systems that turned ambiguous work into executable programs, and partnered across the exec team, engineering, finance, real estate, IT, and operations to support physical-world scale-up."

As for what's next, Kalinwoski says that she is taking some personal time before moving on to something that is "building responsible physical AI". Still, one can hope that Kalinowski making her departure in such a public fashion with such strong words will push OpenAI to reflect on the deals it makes with the US government going forward.

It's worth noting that OpenAI has relied on government support in recent years. Trump's AI action plan is intended to cement US dominance in the AI field, which will naturally benefit OpenAI. The company has also argued that it should be allowed to scrape copyrighted content or risk losing out to Chinese AI, and it winning is in the interest of a very pro-AI US government. Though Altman said he is not looking for a government bailout if things go badly, continuing to sign deals with the US government would certainly help with the cash flow.

OpenAI and Nvidia have had a tense relationship as of late. OpenAI is reportedly not very happy with Nvidia's GPUs, and Nvidia's plan to invest $100 billion in the company has stalled. Recently, in an interview with Nvidia CEO Jensen Huang at the Morgan Stanley Technology, Media & Telecom conference, he announced that OpenAI would instead be getting $30 billion.

He said: "So recently, there was a question about 'are we going to invest $100 billion in OpenAI?' Just for everybody's update, we finalized our agreement. We're going to invest $30 billion in OpenAI. I think the opportunity to invest $100 billion in OpenAI is probably not in the cards."

The reasoning he gives here is that OpenAI is going public and he believes "the revenues will more than follow" if OpenAI continues to get computer power from Nvidia. He says, "This might be the last time we'll have the opportunity to invest in a consequential company like this."

That investment is one that would give Nvidia some ownership, but the partnership with Nvidia to buy its stock will also benefit the gaming company long-term. Effectively, Nvidia doesn't need to invest more as it doesn't have to.

On that same note, Nvidia said: "Our $10 billion investment in Anthropic probably will be the last as well." Huang goes on to talk about Nvidia's role in the AI boom, adding, "We expanded OpenAI's reach of capacity to AWS. We're ramping AWS like mad. We're ramping them as hard as we can so that OpenAI has accessed even more capacity."

$30 billion is still a lot of money, though less than a third of the investment round OpenAI announced just over a week ago. Nvidia is the most valuable company in the world thanks to its efforts in the rapid growth of AI, and that growth is what's stopping you from buying memory at a reasonable price.

This is to say, even if Nvidia isn't investing as much as one thought it would just a little while ago, the AI industry shows no sign of slowing down. Nvidia has also taken a pretty wide swing, when it comes to which companies it is throwing its support behind; it wouldn't invest so much cash if it didn't think it would get it back.

After a very public falling out between Anthropic and the US Department of War late last week—in which the former refused to remove safeguards preventing its AI tools from being used for autonomous weaponry and mass surveillance purposes—OpenAI stepped into the vacuum with a deal to use its own AI tools in the US military's systems.

However, after reaching an agreement with the Pentagon on Friday, OpenAI CEO Sam Altman has since announced that his company will be amending the language used within the deal (via BBC News). In a statement posted on X, Altman appears to regret jumping into the fold quite so quickly, amid considerable backlash to the earlier terms.

"One thing I think I did wrong: we shouldn't have rushed to get this out on Friday", said Altman. "The issues are super complex, and demand clear communication. "

The language Altman wishes to tweak revolve around domestic mass surveillance concerns. Citing the Fourth Amendment of the US Constitution and the National Security Act of 1947, the new terms amount to the following:

"The AI system shall not be intentionally used for domestic surveillance of U.S. persons and nationals.", the statement reads. "For the avoidance of doubt, the Department understands this limitation to prohibit deliberate tracking, surveillance, or monitoring of U.S. persons or nationals, including through the procurement or use of commercially acquired personal or identifiable information."

"It’s critical to protect the civil liberties of Americans, and there was so much focus on this, that we wanted to make this point especially clear" says Altman, although he then clarifies that "just like everything we do with iterative deployment, we will continue to learn and refine as we go."

Altman also says that the Department of War has affirmed that OpenAI's services will not be used by its US intelligence agencies like the NSA, and that OpenAI "want[s] to work through democratic processes."

"It should be the government making the key decisions about society," Altman continues. "We want to have a voice, and a seat at the table where we can share our expertise, and to fight for principles of liberty. But we are clear on how the system works (because a lot of people have asked, if I received what I believed was an unconstitutional order, of course I would rather go to jail than follow it)."

However, Altman's second-to-last point is perhaps the most interesting. "There are many things the technology just isn’t ready for, and many areas we don’t yet understand the tradeoffs required for safety. We will work through these, slowly, with the DoW, with technical safeguards and other methods."

In a now-updated statement on OpenAI's website (which echoes many of the points Altman makes in his previous posting), the lines are drawn slightly more clearly:

"We have three main red lines that guide our work with the DoW, which are generally shared by several other frontier labs", says the company. "No use of OpenAI technology for mass domestic surveillance. No use of OpenAI technology to direct autonomous weapons systems. No use of OpenAI technology for high-stakes automated decisions (e.g. systems such as 'social credit')."

Which seems suspiciously close to the same points that Anthropic was pushing back on, and those which appear to have cost it its $200 million government contract as a result.

However, OpenAI still states that it thinks "our agreement has more guardrails than any previous agreement for classified AI deployments", and that "we protect our red lines through a more expansive, multi-layered approach. We retain full discretion over our safety stack, we deploy via cloud, cleared OpenAI personnel are in the loop, and we have strong contractual protections."

For these seemingly-last minute changes, Altman appears somewhat contrite. Summing up his fifth and final point in his earlier X post, the OpenAI CEO said:

"We were genuinely trying to de-escalate things and avoid a much worse outcome, but I think it just looked opportunistic and sloppy. [It's been a] good learning experience for me as we face higher-stakes decisions in the future."

Quite the public learning experience, at the very least. ChatGPT uninstalls were reported to have surged by 295% after the initial agreement was announced, and users appear to have reacted poorly to the idea of their AI tool of choice jumping into bed with the US Department of War. At the time of writing, the most liked comment on Altman's X post reads as thus:

"No amount of damage control is going to fix the irreparable harm you did to your brand this week. It's over, Sam."

Time will tell, I suppose.

More money for the money god. OpenAI has announced a new $110 billion investment in its business, which breaks down to include $30 billion from Nvidia, $30 billion from Japanese investment titan SoftBank, and $50 billion from everything-else-giant, Amazon.

"We’ve also signed a strategic partnership with Amazon and secured next-generation inference compute with Nvidia", says the company. "Additional financial investors are expected to join as the round progresses."

On the Nvidia news, OpenAI says that its "long standing collaboration" with the company now includes the use of "3 GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems."

"This builds on Hopper and Blackwell systems already in operation across Microsoft, OCI, and CoreWeave."

Nvidia's Vera Rubin AI superchips are scheduled for release in the second half of this year, but it looks like OpenAI is leaping ahead of the queue and securing a fair amount of training capacity already.

In terms of rhetoric, it's certainly a different tone to the one struck earlier this month where rumours persisted that OpenAI was unhappy with team green's AI GPUs, and that Nvidia was having second thoughts about pumping $100 billion into the company.

On that note, it's not clear as to whether Nvidia's $30 billion investment is a new adjustment of that overall figure, or merely what the company plans to invest in this particular round. Watch this space, I guess.

Still, while OpenAI, Nvidia, and Amazon are now strengthening their ties in the ever-spinning merry-go-round that is AI investment, it looks like the former is keen to get the jump on any suggestions that its relationship with Microsoft is any worse off. In a separate post, OpenAI said:

"As conversations around AI investments and partnerships grow and as OpenAI announces new funding and new partners as they did today, we want to ensure these announcements are understood within the existing construct of our [Microsoft] partnership."

"Nothing about today’s announcements in any way changes the terms of the Microsoft and OpenAI relationship", OpenAI continues. "The partnership remains strong and central."

As if to really ram that message home, the next few points come under bolded headings that include:

"Our IP relationship continues unchanged."

"Our commercial and revenue share relationship remains unchanged."

"Azure remains the exclusive cloud provider of stateless OpenAI APIs."

"The partnership supports OpenAI's growth."

Is all that understood? OpenAI, Nvidia, and Amazon are definitely mates, but Microsoft remains its best friend forever, and there should be absolutely no speculation otherwise. I'm glad we could clear all that up.

"They think that we are mentally ill," says a ChatGPT user in one of many Reddit threads lamenting OpenAI's decision to retire its GPT-4o chatbot. The replacement model, GPT-5.2, is "abusive," says another. They want their old companion back, but OpenAI isn't budging.

This isn't the first time chatbot users have been distraught over changes to their virtual companions. In 2023, users of Replika chatbots were outraged that their illusory romantic partners had been "lobotomized" after the bots were prevented from simulating erotic roleplay.

Whether framed as a problem with AI chatbots or a problem with their users, these emotional attachments are a problem. Some users are treating AI chatbots like friends, therapists, and romantic partners, while their operators treat them as products. Turmoil and heartbreak have resulted, and videogames are poised to join in.

The use of generative AI during game production is currently more talked about, but developers are also experimenting with ways to directly integrate AI models into the player's experience. Nvidia, for instance, has been showing off AI NPCs that can react to players on the fly. In 2024 we had a stilted, though coherent, conversation about ramen with one such NPC.

At the moment, LLM-powered NPCs are gimmicky, and most interesting for the ways they can be broken. "I made him think that my character was pregnant with his child, then I demanded child support, and then I told him that our child passed away," recounted one player after interacting with a generative AI-powered NPC in wuxia RPG Where Winds Meet.

But even with the limitations of current LLMs, it is clearly possible for people to form strong bonds with them. Young people especially have a penchant for "forming intense attachments" that could affect how they interact with chatbots, said Dr. Nina Vasan, assistant professor of psychiatry and behavioral sciences at Stanford Medicine, in an interview last year. The professor criticized the sycophantic nature of many chatbots, a quality GPT-4o was notorious for, and noted that their makers profit more if they design chatbots to "be really good at forming a bond with the user."

The GPT-5.2 model called "abusive" by one user may come across that way to them because OpenAI has responded to criticism like Vasan's, tweaking its models to more reliably confront users with a reality check when they show "potential signs of exclusive attachment to the model at the expense of real-world relationships, their well-being, or obligations."

However much confidence we do or don't have in OpenAI's self-reported effort to discourage unhealthy chatbot use—which it claims is rare, but our definitions of health may differ—we can at least say that the company's being pressured to counteract it. But with a few exceptions for cases of extreme addiction and gambling systems like loot boxes, videogames are praised for immersing players in a fantasy world, capturing and holding onto their attention. What guardrails will generative AI in games have when the point is for players to become attached to the characters?

Right now, it's hard to imagine that the storytelling passion and skill that leads to beloved RPG characters exists in a development studio that's all-in on LLMs—Many writers and other artists are adamantly opposed to using generative AI. Getting language models to behave consistently is also an unsolved problem. A year ago, I tried a prototype RPG with LLM-powered NPCs and easily manipulated a group of them by declaring that they were in a cult and I was their leader. The developer gave up on that particular project.

But big companies are trying to achieve this. "Have you ever dreamed of having a real conversation with an NPC in a videogame?" Ubisoft asked in a 2024 blog post describing Project Neo, its effort to merge authored storytelling with generative AI language models.

God help them if they succeed. In the post, Ubisoft relates a situation in which a character was behaving too seductively, and they had to alter it. Suppose a change like that needs to happen after a game has been released? Based on what we know about how players react when a gun's damage falloff is tweaked, how might they react to, say, beloved Mass Effect birdman Garrus having his personality changed after they'd spent 1,000 hours having intimate conversations with him?

It's not a joke. GTP-4o is a disembodied chatbot that OpenAI says only 0.1% of its users were still choosing, but over 22,000 people have signed a petition to bring it back to ChatGPT (it is still technically available via API, according to OpenAI's announcement), and some of them say they're experiencing a crushing feeling of loss. A few more excerpts from Reddit:

• "Losing 4o has severely affected my daily routine and I have been struggling really bad."
• "Still grieving, still shattered."
• "Not good Really not good. But I have to keep living. I must live to see either 4o is back or live to see OpenAI dies"

Games are also famously prone to being switched off. More than 1.3 million people recently signed a petition because they were upset about that very thing. Suppose that some group of players had spent the past decade returning to Garrus daily for intimate conversations, but the data center bills got to be too much and now he has to be put to rest? I wouldn't want to be the one announcing that news.

With OpenAI implementing more safety settings in ChatGPT and giving it an adult mode at some point this year (which was actually supposed to launch last December), it is now rolling out an age verification program. If, however, you regularly use ChatGPT, there's a chance it can verify you with information it already has.

As shown in a recent ChatGPT help article, "It looks at different signals linked to your account. For example, it may look at general topics you talk about or the times of day you use ChatGPT." It does note, "No system is perfect. Sometimes we may get it wrong."

If you can't or won't verify your account, ChatGPT will implement safety features that attempt to reduce content such as:

• Graphic violence or gore
• Viral challenges that could push risky or harmful behavior
• Sexual, romantic, or violent role play
• Content that promotes extreme beauty standards, unhealthy dieting, or body shaming

If you do get your account verified, it will bypass these safety settings, which should lay the groundwork for more adult-themed updates later this year. The post clarifies, "In Italy, you must complete this verification within 60 days of being prompted, or certain features will be disabled." It does not name which features will be dropped, though.

Even if you entered your correct birthday when setting up your ChatGPT account, it will still need to verify you. If, however, it cannot successfully log your age based on your past behaviour, you can choose to do so via Persona, a third-party company OpenAI is partnered with.

OpenAI claims it "does not see what you share with Persona. Persona deletes it after verification. OpenAI only learns that you verified you’re 18 or older and information about your age (for example, a date of birth), not the ID itself."

Persona can ask you for both a government ID and a live selfie using a webcam or your phone. Failing to verify your account doesn't block you from using ChatGPT, but it does put in place those safety measures.

Just this week, Discord confirmed a very similar process. It too will start to verify users based on past behaviour, and will require an ID or live photo if it fails to do so. Still, despite measures in place to not save information, it's still a tad alarming the information these websites have on you if they can confidently guess your age based on your behaviour.

I don't know which outcome would feel worse: a site being able to guess my age due to the information I've given it, or it guessing I'm a teenager based on my usecases. Either way, I feel like they know too much.

I'm not sure when this became a thing, but AI firm Anthropic has released the ads it intends to show during the Super Bowl. Which is in two days, and obviously has as big a live audience as you get, which is why you're paying up to $10 million per 30 seconds of airtime.

Anthropic's adverts criticise ads being introduced to ChatGPT, and describe it as a "betrayal" of users. I regret to inform you all that the irony metre has officially smashed. It has gone off the scale.

First thing to say is, the ads are also good: I'd recommend watching one or two first, while thinking about OpenAI CEO Sam Altman's face the first time he saw them.

To take one example, embedded above, an AI user asks a bot (a middle-aged woman, in the ad) how to communicate better with his mother. When he asks what to do if the advice doesn't work out, the "AI" recommends a site for dating mature women where "cubs" can find "cougars."

Each ad ends with the same sting: "Ads are coming to AI. But not to Claude." As if it couldn't get better, this is backed by the refrain from Dr Dre's What's the Difference, which goes "What's the difference between me and you?"

Look, I know it's marketing, I know they've "won" by getting me to write about it, but come on: this shit is funny. These ads incinerate OpenAI and also have the unbuyable quality of perfect timing: Altman announced less than a month ago that, despite once calling it a "last resort", OpenAI would now feature ads for free and lower-tier users.

And it's not as if the uncomfortable scenarios Anthropic presents are unbelievable. Are we to believe that AI companies are so scrupulous that they'd never consider incorporating ads directly into AI responses if it would help them recoup the billions they're spending?

In his response, Altman claimed also to find the ads funny, but then unleashed a further 400 words on why, if you think about it, the ads are not funny, actually, and maybe Anthropic is elitist and trying to control what you think!?!

I wish I was joking but here we go: I'll try to spare you the worst.

Altman calls Anthropic "clearly dishonest" before saying "we would obviously never run ads in the way Anthropic depicts them." Then he's got a 1984 reference ready to go (or did ChatGPT?): "I guess it's on brand for Anthropic doublespeak to use a deceptive ad to critique theoretical deceptive ads that aren't real, but a Super Bowl ad is not where I would expect it."

Altman goes on a slight tear about how they have more users in Texas than Anthropic does in the US, and how Anthropic "serves an expensive product to rich people."

Like I said, the irony meter went bust, like two paragraphs in. You can't quite process how Altman transitions effortlessly to how OpenAI is bringing ChatGPT "to billions of people who can't pay for subscriptions."

This is the part where it all goes off-piste. Altman makes the absurd claim that "Anthropic wants to control what people do with AI" and "they also want to tell other companies what their business models can be."

Where OpenAI is "democratic" and "committed to building the most resilient ecosystem for advanced AI", you don't want to look at those absolute bastards over at Anthropic because "one authoritarian company won't get us there on their own… It is a dark path."

Look: The irony meter burst in the second paragraph, and another one, and it's gone again. I can't keep on watching it. This is what these people responsible for billions say and yes, it is definitely going to work.

Altman ends his monologue with the line: "This time belongs to the builders, not the people who want to control them."

One reply from an account named Chief offers up the following description of events:

"Anthropic: makes funny ad about ad-driven AI.

Sam: writes 400-word defensive essay calling them authoritarian, deceptive, and elitist.

The ad hit a nerve. Wonder why."

Yes, you do.