Nathan Lambert

Perfect timing for COLM2025 here in Montreal.

Thoughts on The Curve The conference and the trajectory.

Augmentation of humans and a restructuring of the research org is far more likely.

Among other thoughts on a great conference!

buff.ly/mzrRumA

The Curve is a new style of mini AI conference to debate AI progress.

Here I reflect on it and explain why the argument that AI will fully replace human research engineers, and then scientists, is far fetched in the years of compute scarcity.

I can't believe we're actually as a general public going to need to be using quadrillion all the time in every day discourse, starting with tokens processed.

We'll blink and measuring in trillions won't cut it anymore.

Oh yeah I do this but wear many hats and things change fast these days…

+1

You should be spending 10+minutes on slides per minute of your talk.
Doing too many talks makes it so you don't have time for top quality.

One of the mindviruses we may never fully squash is that people think you should and CAN actually ban open weights AI models in the U.S.

Good luck with that buckaroo. We have plenty of time to prepare for any potentially dangerous open models of the future.

I gave a talk today at The Curve on the state of open models.
Here are the slides, recording soon.

Topics include: Chinese ecosystem, reflections on DeepSeek, the demise of Llama, who will fill the U.S. market, what local models do, ATOM project & ai2, and more topics
buff.ly/8BiC67C

A ton of attention over the years goes to plots comparing open to closed models.
The real trend that matters for AI impacts on society is the gap between closed frontier models and local consumer models.
Local models passing major milestones will have major repercussions.
buff.ly/ccMJydQ

What changed? Despite many wonderful models, Anthropic never really remotely translated to LMArena.

The core question -- has LMArena's users or Anthropic's models shifted? Or both?

Mostly releasing base models, but it’s not a substantive analysis just a tweet

Seeing Qwen as trying to build Android for AI models: Cheap, everywhere, powerful, modifiable, should give you a good sense of their strategy.

"Be It Enacted" A Proposal for Federal AI Preemption

Recent dean: buff.ly/n2ygaU5
My original post: buff.ly/rUgJUxQ
More on character training with a model spec update: buff.ly/YwVvjiM
On sychophancy in GPT4o, and how it related to the model spec: buff.ly/p1b4M1H

As Olmo gets better, this has been on my list to craft one and share the process, difficulties of following it, and so on. I welcome pressure on needing to deliver this in order to set a better example.

The behavior of these models is actually remarkably steerable (sharing more research I'm involved with on this soon!) and the lack of model specs is pretty awful as a community standard.

Links to Dean's piece, and my older pieces on model specs are all below.

Largely these seem to be blocked on politics -- both internal where teams actually cant agree what the model should do -- and external, where labs fear pushback.

The model spec sets the intentions of how the model should behave, irregardless as to if it succeeds in it.

Again, I'm happy to discuss this with labs as a free consult as I think its great for the world.

I haven't posted about Model Spec's in a while, but Dean gave me a shoutout on my earlier writing on them, so its time to say definitively again that every frontier lab should have a model spec. It builds long term trust with users, developers and regulators.

The fact that llama 4 and qwen 3 were only a few weeks is crazy looking back at history. What an unbelievable shift in power.

Nvidia should acquire huggingface as a cheap way to build deeper integrations between CUDA and the open-source ecosystem.
Almost as cheap as training leading open models themselves, and in line with increased openness from Nvidia in 2025.

Nice to see another fully open, multimodal LM released! Good license, training code, pretraining data, all here.
LLaVA-OneVision-1.5: Fully Open Framework for Democratized Multimodal Training

Slowly, the community is growing.
arxiv.org/abs/2509.236...

I expect interests of people
To shift over time but this is a valid critique for sure

100%.

I mean people love to buy stuff on instagram this is letting you find stuff rather than the advertiser finding your interests.

I know the average Bluesky user absolutely detests this but also it’s impactful for the majority of people using ChatGPT causally.

Still hoping to avoid dystopia, not that I have thatttt much trust in Altman.

We aren’t the audience yeah. I would use ChatGPT a bit for shopping, but the Bluesky folks are definitely NOT the target

ChatGPT: The Agentic App ChatGPT's long awaited move into user monetization.

Read in full here! buff.ly/4AX415Z

ChatGPT: The Agentic App
Every step ChatGPT takes to realize its vision as the most valuable consumer app since Facebook can have massive repercussions on the AI industry.

ChatGPT's long awaited move into user monetization and what it shows about the future of ChatGPT (and AI products writ large).

ChatGPT launching integrated checkouts is a way bigger deal than the other model/research news coming today.