DINOv2 weight space programming. With the help of Opus 5.5, its features both look better and actually improve downstream task performance.. Not as good as DINOv3 (trained on 10x more data) still, but was shocked to see tbh
We're sharing new data on how AI is changing research at OpenAI.
We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area.
This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems.
Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry.
https://t.co/xyGPcCkPCK
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
absolute fanfic. AI has real, immediate, tangible dangers that we need to focus on, for instance the impact on the job market and making sure random script kiddie can't use it for cyber offense.
But this "oh once it's superintelligent it will take over and ensave us, maybe kill us all!" is pure sci-fi fanfic and distracts from the real (and difficult) issues, which linear probing won't solve.
It's worth repeating the LLMs don't just "decide" to do something and then just "do it" in a void. They generate tokens based on our prompts, only as long as we call `model.forward()`, and on giant clusters with terabytes of weights, that's impossible to "quickly self-replicate".
We also need to regularly remind that the OAI/HF incident was specifically a _guardrail-free_ cyber eval where the model is _instructed_ to hack, to find out how good it is at it.
Some recent personal event made me realize there needs to be more push-back on this doomsday nonsense, and we need to re-focus the discussion on real dangers/issues. I won't argue much in the replies though, not interested in peer-reviewing Terminator fanfic.
I don't think this matters? You want them to keep their "world knowledge" that the vlm imbues. Whether they still remember how exactly an AI2D quiz looks like, i would say it's not clear that it should be any useful. I think the "keeping world knowledge" part should be tested differently eg generalization to variations of the robotics tasks.
Really amazing work! 👏
Moving the ViT’s parameter budget into vision-only modules in the early decoder layers preserves performance.
But could MoE routing learn that specialization on its own? So an encoder-free MoE might already work well without this.
Also the ~3× compute gap is for the linear-patch baseline. At equal FLOPs, a hierarchical MLP on 640px images matches the 224px ViT on VQA. I’d count that as an encoder-free design too.
Regarding pretrained vs scratch encoders:
The gap narrows in Fig. 2. With longer VLM pretraining, would the pretrained encoder eventually pull ahead at higher lifetime FLOP budgets?
The MetaCLIP comparison follows a single pretraining trajectory, and the main CapPa runs use web alt-text at 224px. I’d be very curious to see whether “better encoder ⇏ better VQA” still holds for an encoder pretrained at higher resolution with NTP loss on OCR and grounding data.
This is great research
Run simple experiments, share the findings right away
No novelty-seeking, no 50-pages-long paper. No need to wait for reviews
JEM looking good, congrats JEM authors!
A VLM can figure out semantics: which object to grasp, where to put it. But how do we turn this knowledge into fluent and dexterous actions?
At CoRL on Tuesday, Andy and Will will talk about Flow Reversal Steering: a method for passing VLM "hints" to a diffusion VLA.
Congrats @thinkymachines team on Inkling 2 Small!
A 594B model pretrained in just 68 hours with NVIDIA NVFP4 on GB300 NVL72 racks.
Read how the team made low-precision training work at this scale. 👇
Safety evaluations are critical for understanding whether a model is ready to deploy, which safeguards to build for it, and where to focus training and research.
On our Engineering Blog, we offer 10 recommendations for how to design, run and report them.
https://t.co/MtSQmqqi2D
Reframe: Teaching visual agents to use what they see
A new post-training framework that grounds tool rewards and repairs observations during distillation. Beats prior art on 14 benchmarks without SFT cold start.
Google just released Multimodal Knowledge Agent on Hugging Face
A benchmark for AI agents tackling real-world visual tasks across 10 professional domains—with CAD blueprints, RF schematics, and more.
https://t.co/k6U0SRQybr
@MicrosoftAI will be at NeurIPS in December both Sydney and Paris, I will personally be in Paris walking the floor! Fill my form pinned if interested in meeting us to discuss our models 👇
We are #hiring#ai - follow me for live updates!
https://t.co/8aS7TRc8Ku
In FP8 and FP4, attention is no longer bottlenecked by matrix multiplication. It is bottlenecked by the transcendental and vector math of softmax and the on-chip memory traffic required to shuttle intermediate scores between execution domains:
(Article 🧵) https://t.co/A4ZEmbpCvo
That playbook just hit a physical wall of its own making. By scaling matrix throughput 2× to 4× per generation while leaving Special Function Units (SFUs), vector ALUs, and shared-memory bandwidth behind, chipmakers inverted the attention roofline.
In FP8 and FP4, attention is no longer bottlenecked by matrix multiplication. It is bottlenecked by the transcendental and vector math of softmax and the on-chip memory traffic required to shuttle intermediate scores between execution domains:
When the non-matmul work outruns the matmuls, low-precision attention stops being a GEMM problem.
The industry already voted with production code: drop tokens across the sequence (DSA/QSA via a cheap ReLU indexer with 0 SFU calls), then run the surviving tokens as 100% dense tiles at peak low-bit Tensor Core speed.
Highly recommended read.
Learn how one of the most advanced agent memory systems works.
Your memory system is one of the biggest levers you have in the harness.
I also don't think a one-size-fits-all memory system exists today.
Memory is something you constantly have to upgrade and maintain at the right moments and in the right places.
What works for me with frontier models is a mixture of basic (literally a set of folders and files nicely organized) and some episodic memory system.
Like agent skills, keep your memory system extremely basic. Otherwise, you will only get in the system's way and confuse it.
As models continue to improve, these memory systems will keep evolving. One thing is clear to me: no matter how powerful these models get, your agent will need some form of memory to keep things personalized and useful for you.
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work.
GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
https://t.co/Nzab3Z1Td6
I am very proud of this work. FlashProve is a research prototype, but it shows that proving which model produced a response can be practical.
In the blog, the first usecase we discuss is AI pacing: such proofs get us part of the way to checking that hardware serves inference, and is not being used for training.
This is just a starting point, and I'm very excited for others to take on the mantle and outdo us.
https://t.co/0pn81QXTKK