🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Qwen3 Finetuning Notebook. I’m tuning Qwen 3 for a fast local coding, and here’s a notebook for the process. 🧵More in the thread. More to come

The implied point of view - "we are helpless passive entities unable to change this, merely to predict it" - kinda annoys me. Much like "p of doom". Prediction is much less important than creation. But I can't quite resist the temptation to check...

We went from our first line of code for prime-rl to releasing INTELLECT-2 in around two months. Now that our infra is in place and proven to work, I’m very optimistic that it’s only a matter of time until we will catch up with frontier labs. Some thoughts on this release ⬇️

Huge update to our preprint today on the superhuman performance of reasoning models in medical diagnosis! TL;DR – they don't just surpass humans in meaningful benchmarks, but in actual medical care from unstructured clinical data: A 🧵⬇️: x.com/AdamRodmanMD/s…

#veo3 is truly incredible. Here's my old explanation on why video/audio progress faster than text. 2025 is the year of agents and grokking physics.

We always want to scale up RL, yet simply training longer doesn't necessarily push the limits - exploration gets impeded by entropy collapse. We show that the performance ceiling is surprisingly predictable, and the collapse is driven by covariance between logp and advantage.

I haven't been so worried about videogen effect on misinformation because there are relatively straightforward technological solutions where public figures can verify clips etc. But still, it is now in 2025 that they have to be rolled out

Introducing the world's first reasoning model in biology! 🧬 BioReason enables AI to reason about genomics like a biology expert. A thread 🧵:

anthropic's "race to the top" is not going too well Anthropic

flops per capita

I feel like OpenAI voice mode changed the tone of voice (not the persona) to have a different laid back cool vibe overnight. Don’t know if that is actually the case but if so it not good to do so without giving some kind of notice

Kyutai Speech-To-Text is now open-source! It’s streaming, supports batched inference, and runs blazingly fast: perfect for interactive applications. Check out the details here: kyutai.org/next/stt

It was a dream come true to teach the course I wish existed at the start of my PhD. We built up the algorithmic foundations of modern-day RL, imitation learning, and RLHF, going deeper than the usual "grab bag of tricks". All 25 lectures + 150 pages of notes are now public! 🧵

So about a month ago, Percy posted a version of this plot of our Marin 32B pretraining run. We got a lot of feedback, both public and private, that the spikes were bad. (This is a thread about how we fixed the spikes. Bear with me. )

While I briefly have no employer, let me tell you what's really happening with AI companies training on public data: [1/n]

The AI world needs to learn this lesson before its too late. We don't want an Oppenheimer amongst us. Physics has never fully gotten clean of that sin. We still talk about it. The AI community should learn from this. 6/6

I'm very happy to see this happen. I think that we're in a vastly better position to solve the alignment problem if we can see what our AIs are thinking, and I think that we sorta mostly can right now, but that by default in the future companies will move away from this paradigm