Wojciech Zaremba

When models start reasoning step-by-step, we suddenly get a huge safety gift: a window into their thought process. We could easily lose this if we're not careful. We're publishing a paper urging frontier labs: please don't train away this monitorability. Authored and endorsed…

We're entering an era where AI outputs are becoming so vast, humans alone can't analyze them. Today's LLMs produce tens of thousands of tokens per task—but complex challenges like comprehensive cancer research, inventing novel molecules, or building entire codebases will soon…

Reasoning models offer a groundbreaking approach to interpretability by analyzing the content of their Chain-of-Thoughts. These models reveal undesired behaviors within their own Chain-of-Thoughts. This is galaxy brain — the models themselves lay bare their misalignment in…

“How We Think About Safety and Alignment” — this is our cornerstone document. Enjoy! openai.com/safety/how-we-…

Today, we’re releasing a computer-using agent as a research preview. Ensuring safety for agentic models is far more complex than for chatbots. Errors can lead to serious consequences—for instance, the agent might make costly real-world decisions, like accidentally spending…

Reasoning models are transforming AI safety. Our research shows that increasing compute at test time boosts adversarial robustness—making some attacks fail completely. Scaling model size alone couldn’t achieve this. More thinking = better performance & robustness.…

$500B of Stargate investment turns into compute, which turns into tokens. More tokens mean more intelligence. That’s a massive, massive amount of intelligence. And with models getting cheaper thanks to algorithmic breakthroughs… How much intelligence does $500B buy?

I am very proud of “deliberative alignment” work as it may apply to AGI and beyond. The reasoning models like o1 can be aligned in a fundamentally new way. We teach alignment by providing specifications to o1, which then percolates into its chain of thought, deeply baking in the…

Where does the gap between perception and reality on AGI company safety come from? x.ai — Elon is very vocal about safety, but so far, no one at works on safety. Anthropic — just released a computer-using agent without any safety…

I like how the blog bounded-regret.ghost.io/what-will-gpt-… gives a quantitative feel for how the agi will be. Here is a picture if you don't have time to read the blog.

What I love about the ChatGPT search is that it is instantaneous The search represents millisecond scale recall of the information on a planetary scale, while o1 provides deliberate reasoning. Enjoy openai.com/index/introduc…

Language models trained on the entire internet are learning not just a human intelligence but humanity’s collective intelligence.

Very inspiring take on what the future may hold if AI plays out well. darioamodei.com/machines-of-lo…

Canvas represents a new AI interface. Canvas lets you and ChatGPT work side by side on documents or code. Highlight specific sections to get focused assistance, receive inline feedback like a copy editor or code reviewer, and directly edit your work—all within a seamless…

It’s sad to see Mira, Bob, and Barret go—not only because they are excellent leaders but also because I will miss seeing them day to day. They are my friends. Their departures made me think about the hardships parents faced in the Middle Ages when 6 out of 8 children would die…

It's super cool to see the advanced mode being finally rolled out. Speaking with AI when the delay is almost gone feels much more natural. Also, having the model giggle and use voice intonation makes a gigantic difference. It took tremendous work across the company to pull it…

o1 paradigm of solving problems with a chain of thought offers new avenues to safety/alignment research. It’s easier to ensure such AI behaves as expected because we can see its thoughts. I am feeling pumped.

I deeply respect @elonmusk, and I love @sama. It’s sad to see an unnecessary fight. It would be so much better to put your creative energy into building the future you dream of over a quarrel. May you (both) be happy and find peace ❤️.