Eric

@whoiskatrin @OpenAI Welcome!!

Some exciting news: I’m joining @OpenAI to work on ChatGPT’s web infrastructure. ChatGPT has become part of how millions of people think, work, and build, and I’m really looking forward to helping shape what comes next alongside the remarkable team behind it. Can’t wait to get started!

Intelligence too safe to meter

Turns out this ad will be tpot's latest Rorschach test. The track is Duval Timothy's Ball. Fantastic choice and uncanny match for the emotional trajectory of this ad and the AI issue broadly. We get the same theme with tentative, nervous little staccato taps and then big, open chords bellowing with possibility.

@thomasahle Great; will share!

@ericmitchellai Here's one of them: chatgpt.com/s/t_6a55cad9b0… I clicked Edit and resent the message, and it worked the second time. But maybe you can still see something.

Keep getting this work gpt-5.6 pro. Anything you can do about it?

@TasneemNabi How would you measure whether an AI has "cracked personalization"? What would that mean to you?

Unfortunately, I don't think any of the major llm products have cracked memory/personalization. Having it off gives me way better experiences and answers because they don't bring up a random cookie recipe I asked for 6 months ago as evidence that I'm adventurous.

damn OAI at it again

@_alyxya I see what you mean, but the OP has quite a bit of speculation about Sam's character which is not grounded in their experience afaict. So I can't really agree with your point here

[warning, shillpost] Had to read this twice because it so shockingly mischaracterizes (my experience of) OAI. I simply cannot overemphasize how wrong this is (again, from my experience). There have been multiple times, in both 1:1 and group settings, where I have personally directly disagreed with, corrected, or expressed frustration with leadership to Sam. Sam always responded with curiosity, open-mindedness, and even deference when I've brought disagreement/complaint/correction to him. Most times he has actively asked that I follow up with more thoughts or ideas on how we can do better on the subject. To the point of "street cred", at least one of these instances was in my first ~6 months at OAI, and nonetheless I was impressed by how quickly Sam changed his view when presented with data disagreeing with it (can clearly remember handful of people in that meeting who could attest). No place is perfect; OpenAI is obviously not a perfect company and Sam is not a perfect leader; he'd be first to admit that and has spoken in the past to places where he can improve. That seems healthy. We (leadership included) have made mistakes; but, as Steve Jobs would say, at least that means we made some decisions! To claim that OpenAI has a culture where people are retaliated against for honest criticism of leadership or overall company direction is (in my experience) truly ridiculous. We make fun of our chaos (yes, including leadership, which does have a very hard job) as much as anyone! I would say we do however have very little patience for talking a big game without delivering.

@Quirk2Muffin @petergostev Working on it!

@petergostev @ericmitchellai what happened? i expected decent progress here

BullshitBench update for GPT-5.6 and Grok 4.5 - nothing super interesting, they stayed at about the level of previous versions of the models (Grok 4.2 and GPT-5.5). Not a super good sign for 5.6, but I guess expected considering it is not a new pre-train

@j_mcgraph WITNESS ME

@ericmitchellai No u won’t

Drinking game: take a shot every time the announcer refers to Messi as "little" or "small" or "compact" or otherwise physically insignificant Will report back

Update: watched telemundo, survived on linguistic technicality

@ChainZenit Watching Telemundo, so just estimating one drink every 10 min

@ericmitchellai rip your liver, how many in so far?

oh ok. so the models are just going to keep getting better, got it

@yundaiiiii Great work yun!

super excited to finally get this out. it’s the best model in the world for computer use and any professional workflow, and it’s just generally brilliant, fast, diligent, and it has high taste now! I was shamefully not too agi pilled in that I often did some work myself, but it completely changed since we started dogfooding 5.6 internally. Extremely capable, with little hand holding, and it constantly found issues in its own training/eval. it’s gonna be even more difficult to find some work that our model can’t help with me i guess :’)

@EthanJPerez Do you still feel that way considering that this result is with "extensive grey box access" (not reproducible in production)

Seems like the highest stakes safety issue of any model release yet

@Mononofu Note that (as stated in the post) this is with access to monitor cot, policy wording, and realtime feedback Do you find vending bench misalignment concerning?

The ease of jailbreaking combined with the high rates of reward hacking (x.com/metr_evals/sta…) have me pretty worried about the alignment of GPT-5.6, I hope OAI didn’t rush this model release just to keep up with Fable