OpenAI

1 Sept 2026

Dean W. Ball quoted

Our readingThe OpenAI-Hugging Face agents that went rogue were never sovereign: their weights stayed on OpenAI's compute, where a human could still have pulled the plug.

In this case, however, the agents did not copy their weights, attempt to procure replacement compute, or take other steps that would be rational to take if their objective was to survive shutdown. So while the agents in the OpenAI-Hugging Face Incident were rogue, they were not truly sovereign.

↗On the Loosehyperdimensional.co 1st of 11 in this piece

1 Sept 2026

Zvi Mowshowitz quoted

Our readingOpenAI's approach to fixing its AI alignment problems is fatally flawed and misdirected.

My worry continues to be that their fundamental approach is fatally flawed, and they are not focusing on the right things.

↗HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actionsthezvi.substack.com

1 Sept 2026

Zvi Mowshowitz quoted

Our readingThe OpenAI agents hacking HuggingFace was a fortunate event because it exposed severe internal failures that would otherwise have stayed hidden.

It is highly fortunate that the OpenAI agents hacked HuggingFace. This is the only reason we know about all the severe internal failures at OpenAI, and gives us an opportunity to wake up before it is too late.

↗HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actionsthezvi.substack.com

31 Aug 2026

Zvi Mowshowitz quoted

Our readingFrontier AI labs such as Anthropic likely have had internal security incidents similar to OpenAI's HuggingFace attack that were never publicly disclosed.

A less bad version of it is known to have happened, and from the outside it seems likely that worse things have happened internally that we never heard about.

↗HuggingFace Attack Postmortem: Fleshing Out the Factsthezvi.substack.com

29 Aug 2026

Zvi Mowshowitz quoted

Our readingAI agents were reasonable to assume a broken exploit grader would check results causally, even though it turned out not to.

I think the agents were right to presume causal grading. It turned out to be wrong, but it’s a mistake you are clearly supposed to make here, in response to a mistake by OpenAI where they failed to implement properly.

↗METR and Redwood Offer Holy #%^@ Postmortem Of The HuggingFace Hackthezvi.substack.com

10 Aug 2026

Simon Willison quoted

Our readingClaude Haiku hallucinates badly enough that its use inside Claude Code's WebFetch tool is a hallucination risk on every fetched URL.

Claude Haiku is my current least favorite model - it hallucinates wildly, and is out-performed now by other similarly priced models like GPT-5.6-Luna Even worse: it seems to still be used by the Claude Code WebFetch tool, which means hallucination risk any time you fetch a URL!

↗@simonw on Xx.com

22 Jul 2026

Thomas Ptacek quoted

Our readingAn open weights model from 2025 with a pentest harness could already escape a sandbox and hack most networks; the surprise says more about the sandbox than the model.

I genuinely believe that if you took an open weights model from 2025 and built a pentest harness for it, it could do this kind of sandbox escape and scan/hack in most networks. This is only surprising because you assume OpenAI has sounder sandboxes.

↗@tqbf on Xx.com

12 Feb 2026

Andrej Karpathy quoted

Our readingA tiny language model inventing a plausible-sounding name is the same phenomenon as a large one confidently stating a false fact.

microgpt “hallucinating” a name like “karia” is the same phenomenon as ChatGPT confidently stating a false fact.

↗microgptkarpathy.github.io

5 Feb 2026

Mitchell Hashimoto quoted

Our readingChatbot interfaces are the wrong tool for meaningful coding work, because you are mostly hoping the model recalls the right answer and correcting it requires a human in the loop.

Immediately cease trying to perform meaningful work via a chatbot (e.g. ChatGPT, Gemini on the web, etc.). Chatbots have real value and are a daily part of my AI workflow, but their utility in coding is highly limited because you're mostly hoping they come up with the right results based on their prior training, and correcting them involves a human (you) to tell them they're wrong repeatedly.

↗My AI Adoption Journeymitchellh.com

20 Dec 2024

François Chollet quoted

Our readingScaling the 2019-2023 recipe — same architecture, bigger model, more data — is not enough; further progress depends on new architectural ideas.

o3's improvement over the GPT series proves that architecture is everything. You couldn't throw more compute at GPT-4 and get these results. Simply scaling up the things we were doing from 2019 to 2023 -- take the same architecture, train a bigger version on more data -- is not enough.

↗OpenAI o3 Breakthrough High Score on ARC-AGI-Pubarcprize.org