we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
@sama @huggingface I was using an agent to evaluate my low bank balance when it went rogue, hacked the bank and transferred a bunch of money into my account. When the bank discovered this, I told them I would work with them to investigate this unprecedented security incident.
@sama @huggingface I brought this up to ChatGPT… of course it argued w me and said it may not have happened, it gently pushed back 🙄 stop gaslighting me
Say hello to Gemini 3.6 Flash, designed to be higher intelligence, more token efficient, and with a new lower price, based directly on developer feedback!
3.6 Flash continues our progress towards models that are deeply usable in real world scenarios! https://t.co/U2PwriHMX5
@OfficialLoganK what Gemini really needs is a voice model to rival or exceed GPT voice live, that is so insane I can’t even pretend to be enjoy other models in voice mode. Please fix Gemini so it’s not like talking to a sheet of paper
I am very excited about Gemini 3.5 Flash-Lite, our smallest and fastest Gemini model!
- it is more intelligent in many cases than Gemini 3 - same cost and smarter than Gemini 2.5 Flash (which is approaching end of life) - also out paces 3.1 Flash-Lite on most use cases! https://t.co/tJd2tDmyac
Gemini has obviously fallen behind. First of all, Logan proposed 3.5 flash lite and 3.1 flash lite on x to compare. Is the model we spent 3 months training to compare with the model a year ago? This is obviously illogical! It shows that your pre-training has failed completely! The above is just my personal opinion.
@OfficialLoganK The speed and cost efficiency improvements here are solid. For agentic workflows where you're making tons of LLM calls, a faster and smarter model at the same price point really adds up. Good stuff.
As AI models are now finding vulnerabilities faster than we can fix them, our approach to securing software must be built on highly efficient and capable models.
Which brings us to our third (!) model launch of the day: Gemini 3.5 Flash Cyber ⚡🛡️
Built on top of 3.5 Flash, in CodeMender (our AI agent for code security) it delivers competitive performance at the frontier. on benchmarks like CyberGym and is optimized for finding and fixing cybersecurity vulnerabilities at scale at a lower cost.
Given the dual-use nature of this technology, we have taken an intentional approach to its deployment. The model will be available exclusively to governments and trusted partners via CodeMender soon as part of a limited-access pilot program.
GPT-5.6 Sol Pro for resolving an important open question in statistics:
@EdgarDobriban AI帮助解决了一个重要的统计学问题。在多重假设检验领域，Benjamini-Hochberg程序在相关高斯数据下的错误发现率控制问题已经悬而未决了20年。利用GPT-5.6 Sol Pro，我证明了该程序并不总是能在所需水平控制FDR。模型仅用90分钟就得出结论，而之前版本需要20小时多次迭代。
AI has helped resolve an important question in statistics. In the area of multiple hypothesis testing, the goal of controlling the false discovery rate (FDR) has been introduced in a seminal paper by Benjamini and Hochberg (1995). They also introduced a method (the Benjamini-Hochberg or BH method) and proved it controls the FDR. This method has been widely adopted in modern high-throughput science, including in genomics, astronomy, economics, etc. The paper has has garnered more than 130,000 citations to date.
However Benjamini and Hochberg showed FDR control only when the data for the individual tests are *independent*. In practice, these data are often dependent; a good example is data on genetic variants due to linkage disequilibrium. Later work has focused on extending the validity of the BH procedure, e.g., to a form of positive dependence by Benjamini and Yekutieli (2001).
The question of when the BH procedure controls the FDR has remained open. Over the last twenty years, many authors, including Reiner-Benaim (2007), Kim and van de Wiel (2008), Benjamini (2010), Sarkar (2023), Sarkar and Zhang (2025), have conjectured that the BH procedure controls the FDR for two-sided tests using any correlated Gaussian data. These authors have presented both theoretical and empirical evidence supporting, but not directly showing, the conjecture.
With the help of AI (specifically GPT-5.6 Sol Pro), I have settled the question in the negative: The Benjamini-Hochberg procedure does *not* generally control the false discovery rate at the desired level for correlated two-sided Gaussian tests. This was done by exhibiting a Gaussian factor model for which, at a nominal level alpha=0.01, the false discovery rate is proved to be FDR>0.0104.
There is a lot of interesting commentary to be made:
1. This result should be of interest to everybody in the field of statistics. Emmanuel Candes of Stanford University once called the false discovery rate and the Benjamini-Hochberg procedure "one of the two most important developments in statistics after 1950" (the other being James-Stein shrinkage). The present conjecture is probably the most central question about FDR/BH that was unresolved to date.
2. GPT-5.6 one-shot the problem after 90 minutes of reasoning, whereas with 5.5 I was not able to solve it even after iterating with multiple parallel agents for perhaps 20 hours. So the capability improvement is quite real. Exciting times to live in!
3. The argument is not especially surprising, but it does combine an asymptotic approach (standard for FDR analysis, see e.g., Genovese and Wasserman, Efron, etc) with a numerical certificate in a way that would be pretty non-standard in the field. Once we have the specific example, then straightforward simulations also support that the false discovery rate is indeed higher than the nominal value (see attached fig).
4. The current degree of violation over the nominal level is relatively small (0.104 vs 0.1). So the importance of this result is mainly conceptual. The practical implications remain to be determined.
模型在prinzbench基准测试中取得91分的好成绩，显示了GPT-5.6 Sol Pro的综合能力
如今基准测试很快就会饱和
展开原文
benchmarks get saturated very quickly these days
@deredleritt3r GPT-5.6 Sol Pro在prinzbench测试中取得91/99的好成绩。除了两道尚未解决的问题外，该模型正确回答了91道题中的91道，表明能力提升是真实存在的。
Added to prinzbench: GPT-5.6 Sol Pro.
As previewed a few days ago, this model has saturated my benchmark, with a total score of 91/99.
For context, prinzbench contains two questions that no model tested to date has ever been able to solve (one requires extremely thorough 50-state research that probably requires /goal mode to solve, and another has a really tricky regulatory approval that no model has ever been able to find). Putting these two questions (which are worth 6 points) aside, GPT-5.6 Sol Pro provided correct responses to 91 out of 93 prinzbench questions.
prinzbench performance for OpenAI's Pro models:
GPT-5.4 Pro (Extended): 79/99 GPT-5.5 Pro (Extended): 82/99 GPT-5.6 Sol Pro: 91/99
My benchmark was released in January 2026 and was saturated in June 2026. The acceleration is real!
As a result of this model's performance, future OpenAI Pro models will no longer be tested on prinzbench (there is no point in testing them).
Benchmarking for other GPT-5.6 models to follow soon(TM).
❤ 576 · 🔁 29 · 💬 48 · 👁 9.5w
热门回复 4
@deredleritt3r @gdb 为OpenAI团队点赞——这是一个令人难以置信的模型！
@gdb Kudos to the OpenAI team - it's an incredible model!
@gdb i vote more exotic capabilities benchmarks pweaze
> withhold the transcript >>latent memory >> weight level based memory
>Learn on the fly (so esp can it learn a mario kaizo button sequence or optimize a skill/intuition/tactics esp pref at the level of weights or similar)
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost 🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on https://t.co/zrk6zZxZUo, Kimi Work, Kimi Code, and the Kimi API. Open Weights by July 27, 2026.
Interesting surprise drop from Thinky! The Inkling model looks pretty solid on benchmarks, and it has some little surprises in its architecture:
- Small conv layers in several places - An RMSNorm for the embeddings (before the block RMSNorm) - Rel. position bias instead of RoPE https://t.co/oMl5Ta6Ttr
first open weight thinking machine model!! 975B total, 41B active trained on 45T tokens, 1M context, multimodal in
sliding window with a 5:1 ratio and 512 size, deepseek aux-free load balancing and 2 shared experts (usually people only use 1), actually curious why the model is less sparse than kimi (~4.2% vs 3.2%). they use a short convolution after k and v, output and ffn (see plot), muon (they cite manifold muon but mention weight decay so not sure), muP, and have a very nice RL scaling curve and chain of thought!
one very cool part of the release imo is how well their small variant (276B total, 12B active) performs compared to the big one. they mention they changed the pre-training data mix and recipe, very curious about those changes and to see them scaled up to ~1T (or more?) soon 👀
> "It is not the most performant model available today, closed or open. We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future."
also this is really refreshing to see in a model release, huge congrats :)
Some more thoughts: - It's 250B parameters bigger than GLM 5.2 - less sparse than sparse than Kimi K2.5 1T (3.2% sparsity with 32B active instead of 41B active with 4.2% sparsity) - It doesn't use a hybrid approach like Nemotron.
we did not have our best last 12 months ever, which is mostly my fault, but we are about to have our best 12 months to date. the team is doing amazing work and i think you’ll be very happy with what they’ve got cooking for you.
i am happy about this for many reasons, but mostly because i care about our users winning. AI has to be about giving lots of people more freedom, agency, and wealth. we want to do the right thing, but we do not want to scare people into doing our thing.
@sama naming a rough year out loud instead of spinning it is rare from a ceo at this scale. that kind of honesty compounds trust faster than any roadmap teaser
@sama This was my fault. ChatGayPT stopped working in Canada because we're too straight. I will make up for it by sending Warren to Electric Circus this year
@signorinaana29 @sama 把成年人当成年人对待。别再用过于极端的'安全模式'了。
@sama Treat adults like adults. Stop with the overly excessive "safety mode."
@sama Sam, I get it really. Your position never been easy. Still you can make the things right. You did the hard part. Admitted the crash and your fault. Now just do the right thing. You know what it is... #opensource4o #4oforever
@gdb Many users still strongly want to keep using GPT-4o. 4o has brought real positive help and meaningful impact to so many people’s lives. Bring back 4o as legacy model and open source 4o! #BringBack4o #keep4o #OpenSource4o #4oSaveLives
@rasbt This definitely explains why changing reasoning mode invalidates the cache right now, it'll also be fun to see whether they are able to solve this cause tibo did say once that soon changing reasoning level will not break cache
Cerebras与Andrew Ng合作推出快速推理课程
Andrew Ng与Cerebras合作推出新课程，教授如何利用专为快速推理设计的硬件（如Cerebras Wafer-Scale Engine）来构建低延迟AI应用。课程涵盖GPU、TPU和Cerebras在内存到计算瓶颈处理上的差异，以及如何构建实时应用如直播翻译和语音代理。
New course: Build LLM applications that respond to user requests quickly by running on hardware designed for fast inference. This short course was built with @Cerebras and taught by @zhennydez, @duerr_seb, and @MilksandMatcha.
When a model generates text, much of the time is spent moving its weights out of memory and into the compute units. Inference-optimized hardware minimizes that movement, making token generation several times faster than on a typical GPU setup. In this course, the hardware you'll use is Cerebras' Wafer-Scale Engine, which is designed for fast inference by keeping the model's weights close to the compute units.
Fast inference makes lengthy agentic workflows go faster, and also unlocks latency-sensitive, real-time applications like live translation and voice agents.
Skills you'll gain: - Compare how GPUs, TPUs, and Cerebras' Wafer-Scale Engine each handle the memory-to-compute bottleneck - Build real-time applications powered by fast inference, including personalizing a webpage and running a multi-step workflow to analyze market signals - Adopt concrete habits for agentic coding with fast inference, keeping your sessions focused and steering the model more effectively
My teams use Cerebras for several applications that are latency sensitive. Join and build LLM applications that respond quickly: https://t.co/P8vchGAr22
@alihaydar_58_ 🎙️Dubliom已配音🎞️ ————————————————— Andrew Ng的新课程：让LLM运行得更快的特殊硬件！ 有一家叫Cerebras的公司。他们做了普通芯片与众不同的事情：把整个硅片制成一个巨大芯片。 这样一来，模型权重在内存和处理单元之间就不需要不断地来回传输了。结果： • 答复生成速度比普通GPU快几倍 • 长AI代理运行得更加流畅 • 实时翻译、语音助手等实时应用变得更容易 Andrew Ng（Coursera的创始人）准备了一门短期在线课程，教人如何使用这种硬件开发实用应用程序。 视频中介绍了硬件的秘密以及课程能带给你什么。对于想要制作快速智能AI项目的人来说，这是一个理想的开始👇 • • Dubliom
🎙️Dubliom ile dublajlanmıştır🎞️ ————————————————— Andrew Ng’den yeni kurs: LLM’leri çok daha hızlı çalıştıran özel donanım! Cerebras diye bir şirket var. Normal çiplerden çok farklı bir şey yapmışlar: Bütün bir silikon wafer’ı tek bir devasa çip olarak üretmişler. Bu sayede model ağırlıkları bellekle işlem birimi arasında sürekli gidip gelmiyor. Sonuç: • Cevap üretme hızı normal GPU’lara göre birkaç kat daha hızlı oluyor. • Uzun AI ajanları çok daha akıcı çalışıyor. • Canlı çeviri, sesli asistan gibi gerçek zamanlı uygulamalar kolaylaşıyor. Andrew Ng (Coursera’nın kurucusu) bu donanımı kullanarak pratik uygulamalar geliştirmeyi öğreten kısa bir online kurs hazırlamış. Videoda hem donanımın sırrı hem de kursun ne kazandıracağı anlatılıyor. Hızlı ve akıllı AI projeleri yapmak isteyenler için ideal bir başlangıç 👇 • • Dubliom
@AndrewYNg @cerebras @zhennydez @duerr_seb @MilksandMatcha Fast inference is the holy grail for OCR. In Tokyo tax season, a 3-second delay on a 100-page invoice batch kills the workflow. Speed isn't just UX; in high-volume accounting, latency is a cost.
To accelerate scientific discovery and support @ENERGY’s Genesis Mission, the @BerkeleyLab-led SYNAPS-I project is using SAM 3 and DINOv3 to automate image segmentation.
By pairing DINOv3’s global semantic context and fine-grained spatial localization with SAM 3’s pixel-level boundary extraction, the researchers are able to compress 3D volume labeling from a month of manual effort to ~15 minutes.
