Nathan Lambert

this'll be a big week for American momentum with open models, multiple things are going to start to fall into place

No just chatting with people who have

OpenAI is going to release an open model (1st since GPT 2) and GPT-5 within weeks of eachother. The open model is indicative of where impact is possible in the AI space, where GPT-5 isnt a huge step like GPT-4.

Open models & agents/products are big stories of 2025 so far.

xAI's Grok 4: The tension of frontier performance with a side of Elon favoritism An o3 class model, the possibility of progress, chatbot beige, and the illusiveness of taste.

I discussed some of this in my Grok 4 post on Interconnects.

Will be a fun space to follow. It’s exciting that Gemini had much bigger gains. www.interconnects.ai/p/grok-4-an-...

There’s so much more to play with in the parallel compute space - raw parallelism (like BoN sampling), independent agents with an orchestrator (deep research), different base models w more finetuning, how much compute they’re willing to throw at one prompt, reward model types.

Was reading Swyx’s AI News summary on Gemini DeepThink and it’s understated how DeepThink, Grok Heavy, and o3 pro are more likely to be more different in how they use parallel compute relative to the underlying models (which end up being similar).

Case 912901 of academia being disconnected from reality. Getting weird ethics flags on normal ass papers while people in the real world are building MechaHitler.

You get AGI, but the only form factor you can use to interact with it is short term videos.
Would you use it?

I've finally figured out how to know when OpenAI will release.
Case 1: normal announcement "coming soon" is an Elon style deadline. Totally mysterious, not soon.
Case 2: google has a major announcement, 100% OpenAI lands the exact deadline

Right now we're in case 1

I downgraded from OpenAI pro a few days ago. Already hit my rate limits on o3. Turns out I wasn't wasting money after all.

The funny thing about Meta sounding like it'll go closed w AI is how Zuck has craved the next "platform" ever since their problems with Apple, eg investing $20B+/yr in VR, but Llama is the closest they have come to making one, and they're just going to let it what, fade away?

2025: buff.ly/oxjt89B
2024: buff.ly/kgmaitj

New Zuck post, what a difference a few years makes:

Today: "We'll need to be rigorous about mitigating these risks and careful about what we choose to open source."

2024: "Meta is committed to open source AI... and therefore a platform that will be around for the long term."

Recent AI news (Chinese models and OpenAI’s coming releases)
“Do and don’t” of LLM training organizations
Reasoning research and academic blind spots
Research people aren’t paying enough attention to
Non language modeling news & other topics

Enjoy!

YouTube: buff.ly/OWMU0qZ
Post: buff.ly/vBBYAHP

Back in the interview chair with a general recap of the AI research ecosystem with one of my favorite people to talk to. Ross Taylor (former Llama reasoning lead) is back for round 2!

In this episode we cover some of everything.

Any day now

17 of the top 20 open models on the Artificial Analysis leaderboard are made by Chinese companies. 2 from the US and 1 from Europe. This is what current domination of open models looks like, but it can be changed quickly with focused funding and effort.

Leaderboard: buff.ly/hVyaNl8

I bet pretty soon a Chinese research org drops a LLM scaling laws for RL paper.

Closed frontier labs have definitely done this and wont share it, academics havent mastered the data + infra tweaks yet (or lack the $).

GitHub copilot is hit or miss but if it’s the only one available it’s better than nothing. Haven’t tried cline

Sonnet 4 probably fine, but depends on the interface.

Whenever I start mentoring a new junior student or researcher one of the first things I do is to check that they use sufficiently strong AI.

If they don't I'll pay out of pocket to make sure they aren't shooting themselves in the foot. The payoff is worth it.

Chinese open models being at the frontier today isn’t thaaat surprising for folks who were following closely. What is more surprising is how quickly Llama disappeared from the clear top spot in the conversation. Two trends combine for massive impact.

Pretty much lol

Important distinction: open AI models, not OpenAI models

Slack shipped a summary feature, AGI coming 0.1% sooner, finally man.

progress is far more valuable than the money cost

Which kind of non profit are you, OpenAI or Ai2?

some parts of it are totally fine? I won't speak for all of it, but e.g. open models / research investment are good

This is a major shift in the tenor of how the U.S. government feels about and prioritizes open AI models. This narrow piece of the action plan is a huge W overall, with a couple nit picks. Happy to share my thoughts as I scope my own path to building this reality.

buff.ly/Rq5T5Yp

The new AI action plan is out from the White House
buff.ly/YXNpTtF