Gio

@thsottiaux Tibo can we get the "Request blocked." issue fixed? Even gpt doesn't know what is happening.

Hello beautiful people! We have reset usage limits across Codex and ChatGPT Work. And another one will come later in the day. Rejoice. Now that I have your attention, a quick update on ChatGPT Work, Codex and all the updates we shared yesterday. We’ve spent the last 24 hours reading feedback, looking at usage patterns, and talking with many of you. The short version is that there is a *lot* of excitement for GPT 5.6 Sol, ChatGPT Work on mobile & web, but also that we didn't get everything quite right. - We made it too easy to use the highest-compute settings without making the impact on usage limits sufficiently clear. - We reorganized the desktop app in one bold move, making familiar things like chats and projects harder to find. - Our launch framing was focused on ChatGPT Work and to some of our Codex fans it made it feel like Codex was going away over time. Absolutely not our intention, we love Codex and it is here to stay. - And we introduced regressions for some existing multi-agent workflows, alongside a collection of rough edges in plugins and other parts of the experience. We’re landing a first set of improvements today. We’re resetting usage twice so people can keep experimenting, changing defaults and the model picker so they don’t push people toward unnecessarily expensive settings, fixing several plugin submission issues, improving how we represent Codex in the product, and cleaning up some of the most immediate desktop problems. A larger set of improvements will land next week. We’re bringing chats and projects back into the sidebar in a more familiar and customizable way, making usage and reset timing much more visible, clarifying when to use ChatGPT Work and when to use Codex, and addressing the many other smaller pieces of great feedback we've had. The ambition behind this launch hasn’t changed. We think bringing ChatGPT and Codex together into a workspace where people and agents can collaborate is a very important step forward. But an ambitious direction doesn’t excuse avoidable confusion or regressions in the first version. Please keep the feedback coming. We’re moving quickly, and you should see the experience already get better with a few updates today; and substantially better again next week.

@Ayanjkk @thsottiaux I tried to lower it to 5.5 that's why it's like that. Of course it didn't work

@giosullutrone99 @thsottiaux you are using 5.5 in screenshot

Today, the OpenAI product and engineering team is humming, there is so much to coordinate and keep running across the release of GPT 5.6 Sol/Terra/Luna, ChatGPT Work, the new ChatGPT desktop app, hosted sites and a million other things... together with the new ChatGPT Voice just launched yesterday. We will not get it all right, give us feedback and we will iterate quickly. Thank you for being an amazing community.

@CameronEhrlich @thsottiaux A text-to-sql benchmarking pipeline. Probably not the most dangerous repo on GitHub.

@giosullutrone99 @thsottiaux Inow I’m curious… What’s your repo about?

Ehm @thsottiaux I think I broke something

@SharonYixuanLi @Samuel861025 However, in our internal experiments, we observed much more unpredictable behavior across different sets of retrieved documents, which makes me think that your analysis may have been the missing piece.

@SharonYixuanLi @Samuel861025 Then everyone should strive to write more like them, I'll for sure try to do so. The paper is aclanthology.org/2025.findings-… . In Section 6 Q3, we didn't distinguish between different document types but focused on the number of documents retrieved.

Excited to share our #ACL2026 paper: "How Retrieved Context Shapes Internal Representations in RAG" 🧵 w/ @Samuel861025 RAG is everywhere, but I am often puzzled by what happens inside the model when we stuff retrieved documents into the context. Did it integrate the evidence? Ignore it? Get confused by it? To answer these questions, we need to look at what's happening inside. We built a controlled framework to study how the hidden states of LLMs change under different retrieval conditions: relevant, distracting, and random documents. 🤔 The counterintuitive finding: Relevant documents barely move the needle on internal representations. They largely reinforce what the model already knows rather than injecting decisive new information. Meanwhile, random, irrelevant documents cause large representation drift -- far larger than relevant or even distracting ones. Why? Because LLMs internally recognize uninformative context and shift into a refusal mode. That representation drift is tightly linked to abstention behavior. The model is essentially saying: "I see this context, and I know it's useless." Other key insights: → Later transformer layers increasingly prioritize parametric knowledge over retrieved evidence, limiting how much external context can influence generation. → In multi-document settings, a single relevant document can anchor the entire representation and suppress surrounding noise. This is good news for imperfect retrieval in practice → Distracting documents (semantically similar but unhelpful) are the silent threat: they don't trigger the model's refusal mechanism the way random docs do, but they quietly degrade quality The takeaway for practitioners: understanding RAG can't stop at output accuracy. The internal dynamics tell a richer and sometimes alarming story about when your system is actually integrating evidence vs. when it's ignoring it, fighting it, or shutting down because of it. We hope our analysis offers mechanistic insights for more principled and reliable RAG system design. 📄 arxiv.org/abs/2602.20091

Beautiful work. I think there are also some interesting points that can be applied to downstream tasks

Text diffusion models are fast, but are less transparent than today's LLMs because they do many forward passes before outputting text. We audit the transparency of DiffusionGemma and find that the intermediates are interpretable. This recovers many of the benefits of CoT! 🧵

As believers of open research, we are disappointed to see Anthropic silently degrading Fable 5 for AI development "Any topic related to building pretraining pipelines, distributed training infrastructure, or ML accelerator design... may have limited effectiveness through Claude via methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning." Not only do they get to decide what you use LLMs for in research, but this also enables them to silently intervene in your research without you knowing. This sets a dangerous precedent. If a model refuses openly, users can understand the boundary. If a model falls back to another model, users can still evaluate the difference. But if a model silently modifies or weakens its own answers while still pretending to help, researchers lose the ability to know whether a failed result came from their own idea, their implementation, or an invisible intervention by the model provider. That is not safety. Safety policies should be transparent, auditable, and user-visible. On top of that, the people most harmed by this are not the largest labs with massive teams and proprietary infrastructure. It is the independent researchers, academic groups, startups, and open-source builders who rely on public tools to compete, innovate, and pioneer AI for everyone else.

@sama Corruption by cross-referencing data

what problem do you most hope AI will solve in the future? maybe we can help!

@sama Don't make models any faster or I won't have the time to go for a walk while codex is working

i get some anxiety not using the smartest-available model/settings. but sometimes i dont mind if it's really slow. i wonder if we should focus more on a price/speed tradeoff relative to a price/intelligence tradeoff.

A strange notification to wake up to @sama

@elon_lit Thanks for sharing. I'll for sure read it and maybe try it out in some of the trainings we are working on!

We developed a unified theory of generalization in deep learning. It explains grokking, double descent, benign overfitting, and implicit bias. But theory is only half the story. It turns out that optimizing the population risk of any neural network amounts to a small change to your optimizer. 🧵

@sama An interactive code canvas that I can use to explore the code produced by codex with different levels of details

i would like to talk to people who have built amazing things with 5.5 that weren't possible with earlier models. i am especially interested in examples that took ludicrous token budgets. thanks.

@sama @kr0der I have a couple of jobs running for 2+ hours so it would help, thanks

@kr0der i personally think worth it yeah maybe we will reset limits soon

should i try this? i’ll need a 2nd $200 subscription to make it sustainable

@zeeg @garrytan You still haven't reached your full potential. @thsottiaux do I get a reset for this abomination?

im coming for you today @garrytan

@simpsoka That's extremely cool and thanks to that I just noticed that I had servers open that I completely forgot about

We are truly shipping constantly on the Codex team. It doesn't feel like releases as much as a steady stream of things getting better. I landed my first 'larger' PR this week, a small QOL update to the in app browser that replaces the blank empty state. It now discovers local servers from chats and scans common localhost ports. Lemme know what you think!

@sama @Makuh90 What are you gonna drink once we achieve AGI?

@Makuh90 i have not been drinking YET but ollie got me a super nice bottle of wine to celebrate 5.5 so maybe there will be some tweets tonight...

He's not been drinking. But he's on the money.