Jack

@sama GPT-6 wen

so fun to see the reception to 5.5! there is almost nothing that feels more gratifying to me than builders saying they find our tools useful.

@elliotarledge What is the percentage reporting?

KernelBench-Hard coming soon.

@scaling01 @VictorTaelin I think Anthropic is the least trustworthy of the major players actually

@VictorTaelin it's not about gatekeeping it's about trust and safety

Nothing changed for me. To be very clear: I still want Anthropic to win. They are taking the safer and more principled approach and I trust them a lot more. But I'm not going to sit here and pretend GPT-5.5 or any other OpenAI model sucks because of that. Since GPT-5.2-xhigh I have been saying that OpenAI has smarter models, and this trend has continued. My only complain was the reasoning-efficiency, but they fixed that. I have also made predictions about OpenAI pulling ahead way before all of this happened. Saying their models are better or enjoying Sam's recent drunk comments or post doesn't mean I agree or forgive them for all their past reckless actions. But I do appreciate Sam coming forward. It shows good will. Of course I'm not blind to why Sam might be doing all of this. The recent personal attacks on Sam's home, the upcoming Elon vs OpenAI trial and growing anti-AI sentiment are very good reasons to up your PR game and reflect on your actions. I would do the same. So I'm not discounting the possibility that he's doing this to save his own skin, but I also don't want to discount the possibility of this being a genuine attempt to fix things.

@sengpt “In fact, one can prove something slightly stronger.” Aura

23 yaşında bi genç 60 yıldır çözülemeyen Erdös problemlerinden birini chatgpt 5.4 pro ile çözmüş. hem de tek atışta. chatgpt'nin soruyu çözmek için harcadığı süre 1 saat 20 dakika. işin ilginci ai, herkesin bildiği ama kimsenin bu probleme uygulamadığı bi formülü kullanarak problemi çözmüş. burada chatgpt yazışması; chatgpt.com/share/69dd1c83… bu da problem; erdosproblems.com/1176

@daniel_mac8 I would say it’s smarter than the majority of humans, yes, but not the smartest human.

GPT-5.5 Pro is smarter than the smartest human.

@scaling01 @merlindru SVG’s have definitely been benchmaxxed (I remember when 3.1 pro dropped its pelican on a bike was phenomenal but other simple animals sucked), but I doubt voxelbench has. More likely proof of gains from visual task centric RL IMO

@merlindru some labs benchmaxx SVGs but I guess voxelbench is a bit harder to benchmaxx. it's harder to grade

2101 rating vs second place 1722 gpt-5.5 might be him him although i suspect some benchmaxxing for all these super visual vibe tasks

@RatthewVT @kubaswift @Colgate The team in charge of posting advertisements to Facebook has absolutely no influence over the product quality, manufacturing, distribution, or research.

@kubaswift @Colgate Maybe this is dumb but, are we supposed to keep trusting these companies with our oral health when they are too lazy to even hire real artists or just slapping something together in canva? This is such a red flag to me

I am genuinely in shock. @Colgate is a $66 billion company.

@Angaisb_ I wouldn’t be so sure. 5.5 is a brand new pre-train; coding-centric RL could boost coding performance a lot

OpenAI will not be releasing GPT-5.5 Codex They unified models a month ago, it would've been weird if they separated them again

@froggoidiot @redtachyon “My bitch child I hate”

@redtachyon Me red, child blue is the funniest answer. I've got to assume everyone who answered that is joking

Red button, blue button, blah blah you know the drill. However, for whatever reason you're also asked to vote on behalf of your child, or any other person that you love very much. What do?

@zeta_globin There must be no charitability extended to these people:

sometimes I just have to just suddenly leave meetings for the sake of my criminal record if people are eating while chewing with their mouths open because the instinctual rage is that bad

please don't be racist but genuinely I have to ask and I'm pretty sure there is due to the ubiquity of it: is their like a sinonasal architecture reason certain demographics can't chew with their mouths closed like I'm trying to be charitable even with my level 11 misophonia

@zeta_globin People who chew with their mouth open infuriate me. Like I actually feel enraged because it bothers me so strongly. There was a time I was IN CLASS (graduate school!) and the girl ahead of me took out a packed lunch and started loudly eating with her mouth open. I was apoplectic.

@LinkofSunshine That poll had a profound effect on me

The most worrying poll I’ve ever seen is one when they asked doctors “1% of the population had disease X. The test has a 99% accuracy rate. If someone tests positive for disease X, what are the odds they’re have disease X” And like 95% of doctors said ~99%. This is their job.

@synthwavedd I’m sure they are confident on benchmarks. Benchmarks seem to be what Gemini is best at.

🚨"We have a new version of Gemini coming very, very soon", confident on benchmarks - CEO of Google Cloud

@QuantumPionier @testingcatalog Probably because 5.5 is the first post-train of a brand new pre-train, so additional coding-centric RL stands to provide a lot more utility.

But why? gpt 5.5 is amazing good at coding already. And gpt 5.5 codex would be the same model, just little bit more RL on coding. I think they already max their models at coding primarily, so no need for a codex version. And it feels a lot faster too. Probably just more efficient, but feels good 👍

OPENAI 🚨: GPT-CODEX-5.5 HAS BEEN SPOTTED IN THE WILD. Friday feature drop? 👀

@TheAaryanKapoor @chatgpt21 A model can also reach 0% hallucination rate by simply refusing to answer any question, so take that as you will.

@TheAaryanKapoor @chatgpt21 That’s not what this benchmark is saying. It’s saying that a higher percentage of GPT’s incorrect responses are hallucinations, it does NOT account for baseline accuracy. If GPT-5.5 was 99.99% accurate and the remaining 0.01% was hallucinations, that’s 100% hallucination rate.

How did nobody catch this? OpenAI just took first place on AA-Omniscience Accuracy - from Gemini 3.1 pro! This bench measures how often a model correctly answers hard cross-domain factual questions across all questions, not just the ones it chooses to answer. GPT-5.4 xhigh: 50% GPT-5.5 xhigh: 57%

@theo Did your smartest hacker friends also get the second hint?

The hint in my latest video helped a small handful of people out. 5 have found the answer so far. Genuinely so impressed. This challenge was SUPER hard. Threw it at my smartest hacker friends and none could figure it out.

@AcerFur Understandable. I’ll still take this conversation as 100% confirmation of my pre existing suspicions and let my hype train run wild with no foreseeable GPT-6 launch date.

@AcerFur Probably fully post-trained Spud if I had to guess, unless you’re aware of some internal architecture breakthrough that you can’t disclose

@AcerFur Is GPT-6 the “what’s next” you were referring to? Or is what you’re excited for likely coming sooner than that?