Dohun Lee

Thinking about AGI, ex-Quant

[Dohun Lee @dohunchris]

Found an unwanted behavior in your LLM after SFT? You can just find the training data that caused it, and filter it out, right?

Surprisingly, this worked much worse than we expected! Across many traits and filtering methods, Olmo just keeps inheriting the traits.

[Noam Brown @polynoamial]

An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science.

We believe it will be a major step for scientific reasoning. https://t.co/iP6cyheZ7i

[YouGov @YouGov]

If you had a choice between instantly receiving £50,000 or a 50% chance to win £1m, which would you pick?

£50,000: 73%
50/50 chance of winning £1m: 21%

[Agam Bhatia @chewing_a_gum]

We know the J-space can surface what a model is thinking.

But can it tell us how a model computes its answer?

In a new blogpost, we apply J-Lens to find “meta-tokens” that can directly tell us the algorithm Qwen-3.6-27B uses to complete a task. 🧵

[Timothy Gowers @wtgowers @wtgowers]

Assuming this is correct, it is for me the first example of an LLM solving a problem not in my area that was nevertheless big enough that I had very definitely heard of it. Again it's a counterexample, so not in "end of mathematics" territory, but still pretty amazing.

[Dohun Lee @dohunchris]

Thoughtful

[Kyle Corbitt @corbtt]

I hope you enjoy the story! It's a short read, and I put a lot of love into it.

https://t.co/E83cSjBHMQ

[Thinking Machines @thinkymachines]

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

[Dohun Lee @dohunchris]

@celestepoasts The most quintessential Korean experience

[Daniel Kokotajlo @DKokotajlo]

In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power.

In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.

[Lisan al Gaib @scaling01]

OpenAI says GPT-5.6-Sol autonomously post-trained GPT-5.6-Luna

[Dohun Lee @dohunchris]

Work done as part of Neel Nanda’s MATS Exploration Phase, @jrosseruk (co first author, equal contribution), @JoshAEngels, @NeelNanda5

[Dohun Lee @dohunchris]

Found an unwanted behavior in your LLM after SFT? You can just find the training data that caused it, and filter it out, right?

Surprisingly, this worked much worse than we expected! Across many traits and filtering methods, Olmo just keeps inheriting the traits.

[Dohun Lee @dohunchris]

Read more details, including some of our potential explanations towards why, here: https://t.co/P7iekT2uXf