Nathan Lambert
@natolambert.bsky.social
A LLN - large language Nathan - (RL, RLHF, society, robotics), athlete, yogi, chef
Writes http://interconnects.ai
At Ai2 via HuggingFace, Berkeley, and normal places
Writes http://interconnects.ai
At Ai2 via HuggingFace, Berkeley, and normal places
Pinned
Nathan Lambert
@natolambert.bsky.social
· Apr 16
First draft online version of The RLHF Book is DONE. Recently I've been creating the advanced discussion chapters on everything from Constitutional AI to evaluation and character training, but I also sneak in consistent improvements to the RL specific chapter.
rlhfbook.com
rlhfbook.com
Added a slides gallery to my website with links to all the recordings. Enjoy the backlog from 2025 if you’re interested in post-training or open models :)
natolambert.com/slides
natolambert.com/slides
December 28, 2025 at 4:46 PM
Added a slides gallery to my website with links to all the recordings. Enjoy the backlog from 2025 if you’re interested in post-training or open models :)
natolambert.com/slides
natolambert.com/slides
a nice post snow run at home
December 28, 2025 at 1:13 AM
a nice post snow run at home
Everyone next week is going to be like "So what did Claude 4.5 Opus build for you in your time off."
My answer is helping perform a major polish for the rlhfbook (really nice editor), and some fancy website automations.
My answer is helping perform a major polish for the rlhfbook (really nice editor), and some fancy website automations.
December 27, 2025 at 1:31 PM
Everyone next week is going to be like "So what did Claude 4.5 Opus build for you in your time off."
My answer is helping perform a major polish for the rlhfbook (really nice editor), and some fancy website automations.
My answer is helping perform a major polish for the rlhfbook (really nice editor), and some fancy website automations.
ngl ChatGPT wrapped was pretty on point
Happy holidays all!
Happy holidays all!
December 23, 2025 at 2:19 PM
ngl ChatGPT wrapped was pretty on point
Happy holidays all!
Happy holidays all!
80 posts this year on Interconnects, my most ever, and likely a local maximum.
A short reflection on what happens when your two long-term areas of focus — open models and RL — become the centerpiece of the AI boom.
Thx for your support!
www.interconnects.ai/p/2025-inter...
A short reflection on what happens when your two long-term areas of focus — open models and RL — become the centerpiece of the AI boom.
Thx for your support!
www.interconnects.ai/p/2025-inter...
2025 Interconnects year in review
Three years writing every week about AI.
www.interconnects.ai
December 18, 2025 at 3:11 PM
80 posts this year on Interconnects, my most ever, and likely a local maximum.
A short reflection on what happens when your two long-term areas of focus — open models and RL — become the centerpiece of the AI boom.
Thx for your support!
www.interconnects.ai/p/2025-inter...
A short reflection on what happens when your two long-term areas of focus — open models and RL — become the centerpiece of the AI boom.
Thx for your support!
www.interconnects.ai/p/2025-inter...
I'm making big updates to the RLHF book in the coming weeks, so if you've been reading it & have feedback, please open a GitHub issue. Or, if you took a look at only the initial versions, look again as its come a long way with a trickle of improvements. 🍒
rlhfbook dot com
rlhfbook dot com
December 15, 2025 at 9:12 PM
I'm making big updates to the RLHF book in the coming weeks, so if you've been reading it & have feedback, please open a GitHub issue. Or, if you took a look at only the initial versions, look again as its come a long way with a trickle of improvements. 🍒
rlhfbook dot com
rlhfbook dot com
Can't believe a product called Nano Banana is what is going to take plucky underdog Google Gemini to finally beat the incumbent ChatGPT.
Fast, precise image gen and editing is such a joy to use and surely mass market adoption is cranking.
More warranted to call a Code Red for that than Gemini 3.
Fast, precise image gen and editing is such a joy to use and surely mass market adoption is cranking.
More warranted to call a Code Red for that than Gemini 3.
December 15, 2025 at 4:49 PM
Can't believe a product called Nano Banana is what is going to take plucky underdog Google Gemini to finally beat the incumbent ChatGPT.
Fast, precise image gen and editing is such a joy to use and surely mass market adoption is cranking.
More warranted to call a Code Red for that than Gemini 3.
Fast, precise image gen and editing is such a joy to use and surely mass market adoption is cranking.
More warranted to call a Code Red for that than Gemini 3.
A very nice long read on Olmo 3 from my friend Cameron. He puts in so much care to do a great job on these.
cameronrwolfe.substack.com/p/olmo-3
cameronrwolfe.substack.com/p/olmo-3
December 15, 2025 at 4:08 PM
A very nice long read on Olmo 3 from my friend Cameron. He puts in so much care to do a great job on these.
cameronrwolfe.substack.com/p/olmo-3
cameronrwolfe.substack.com/p/olmo-3
Open models year in review.
What a year! We're back with an updated open model builder tier list, our top models of the year, and our predictions for 2026.
www.interconnects.ai/p/2025-open-...
What a year! We're back with an updated open model builder tier list, our top models of the year, and our predictions for 2026.
www.interconnects.ai/p/2025-open-...
December 14, 2025 at 8:28 PM
Open models year in review.
What a year! We're back with an updated open model builder tier list, our top models of the year, and our predictions for 2026.
www.interconnects.ai/p/2025-open-...
What a year! We're back with an updated open model builder tier list, our top models of the year, and our predictions for 2026.
www.interconnects.ai/p/2025-open-...
Too many cases of starting something in Codex that you think is going great then Opus needs to save the day.
December 13, 2025 at 3:45 PM
Too many cases of starting something in Codex that you think is going great then Opus needs to save the day.
Reposted by Nathan Lambert
Olmo 3.1: even more RL = even more RL-Zero!
@saurabhshah2.bsky.social and I tweaked some hyperparams and prompts, @hamishivi.bsky.social and @finbarr.bsky.social improved the code and boom!
New Olmo 3.1 RL-Zero 👾 An updated, solid baseline for your RL and reasoning research
@saurabhshah2.bsky.social and I tweaked some hyperparams and prompts, @hamishivi.bsky.social and @finbarr.bsky.social improved the code and boom!
New Olmo 3.1 RL-Zero 👾 An updated, solid baseline for your RL and reasoning research
Olmo 3.1 32B Think shows that not just frontier labs can scale RL.
My favorite RL run yet over 7+ years of doing RL.
The biggest fully open RL run ever?
Gold stars on downstream evals is our original release, this latest one is the final checkpoint on the plot.
My favorite RL run yet over 7+ years of doing RL.
The biggest fully open RL run ever?
Gold stars on downstream evals is our original release, this latest one is the final checkpoint on the plot.
December 12, 2025 at 8:42 PM
Olmo 3.1: even more RL = even more RL-Zero!
@saurabhshah2.bsky.social and I tweaked some hyperparams and prompts, @hamishivi.bsky.social and @finbarr.bsky.social improved the code and boom!
New Olmo 3.1 RL-Zero 👾 An updated, solid baseline for your RL and reasoning research
@saurabhshah2.bsky.social and I tweaked some hyperparams and prompts, @hamishivi.bsky.social and @finbarr.bsky.social improved the code and boom!
New Olmo 3.1 RL-Zero 👾 An updated, solid baseline for your RL and reasoning research
My favorite of the Olmo 3 paper: Transparent auditing of the cost for the v3 models (not 3.1), based of wall clock time for pre/post train, evals, cluster issues, etc, as a counter to the famous $5.576M for DeepSeek V3
at $2/H100 hour, Olmo 3 start to end would cost $2.75M
allenai.org/papers/olmo3
at $2/H100 hour, Olmo 3 start to end would cost $2.75M
allenai.org/papers/olmo3
December 12, 2025 at 6:43 PM
My favorite of the Olmo 3 paper: Transparent auditing of the cost for the v3 models (not 3.1), based of wall clock time for pre/post train, evals, cluster issues, etc, as a counter to the famous $5.576M for DeepSeek V3
at $2/H100 hour, Olmo 3 start to end would cost $2.75M
allenai.org/papers/olmo3
at $2/H100 hour, Olmo 3 start to end would cost $2.75M
allenai.org/papers/olmo3
Olmo 3.1 32B Think shows that not just frontier labs can scale RL.
My favorite RL run yet over 7+ years of doing RL.
The biggest fully open RL run ever?
Gold stars on downstream evals is our original release, this latest one is the final checkpoint on the plot.
My favorite RL run yet over 7+ years of doing RL.
The biggest fully open RL run ever?
Gold stars on downstream evals is our original release, this latest one is the final checkpoint on the plot.
December 12, 2025 at 5:15 PM
Olmo 3.1 32B Think shows that not just frontier labs can scale RL.
My favorite RL run yet over 7+ years of doing RL.
The biggest fully open RL run ever?
Gold stars on downstream evals is our original release, this latest one is the final checkpoint on the plot.
My favorite RL run yet over 7+ years of doing RL.
The biggest fully open RL run ever?
Gold stars on downstream evals is our original release, this latest one is the final checkpoint on the plot.
Remember when OpenAI said GPT-5 was when they were going to fix their naming scheme.
December 11, 2025 at 6:07 PM
Remember when OpenAI said GPT-5 was when they were going to fix their naming scheme.
It's finally here: The public (and most complete) version of my talk covering every stage of the process to build Olmo 3 Think. This involves changes and new considerations of every angle of the stack, from pretraining, evaluation, and of course post-training.
December 10, 2025 at 7:36 PM
It's finally here: The public (and most complete) version of my talk covering every stage of the process to build Olmo 3 Think. This involves changes and new considerations of every angle of the stack, from pretraining, evaluation, and of course post-training.
Building Olmo 3 Think
Foundations of Reasoning in Language Models @ NeurIPS 2025
Today 13:45 - 14:30
Foundations of Reasoning in Language Models @ NeurIPS 2025
Today 13:45 - 14:30
December 7, 2025 at 3:26 PM
Building Olmo 3 Think
Foundations of Reasoning in Language Models @ NeurIPS 2025
Today 13:45 - 14:30
Foundations of Reasoning in Language Models @ NeurIPS 2025
Today 13:45 - 14:30
Good researchers obsess over evals
The story of Olmo 3 (post-training), told through evals
NeurIPS Talk tomorrow.
Upper Level Room 2, 10:35AM.
Slides: docs.google.com/presentation...
The story of Olmo 3 (post-training), told through evals
NeurIPS Talk tomorrow.
Upper Level Room 2, 10:35AM.
Slides: docs.google.com/presentation...
December 6, 2025 at 8:35 PM
Good researchers obsess over evals
The story of Olmo 3 (post-training), told through evals
NeurIPS Talk tomorrow.
Upper Level Room 2, 10:35AM.
Slides: docs.google.com/presentation...
The story of Olmo 3 (post-training), told through evals
NeurIPS Talk tomorrow.
Upper Level Room 2, 10:35AM.
Slides: docs.google.com/presentation...
Reposted by Nathan Lambert
Loved seeing our OSU collaborator Sam Stevens featured at the #NeurIPS SAIL Media Booth w/ @natolambert.bsky.social! Great convo on how AI accelerates scientific discovery (not just bigger models 🙃).
Clips ⬇️
youtube.com/shorts/v0k5D...
youtube.com/shorts/4R-BF...
#AI4Science #NeurIPS2025
Clips ⬇️
youtube.com/shorts/v0k5D...
youtube.com/shorts/4R-BF...
#AI4Science #NeurIPS2025
SAIL Booth with Sam Stevens
YouTube video by SAIL Team
youtube.com
December 5, 2025 at 3:26 PM
Loved seeing our OSU collaborator Sam Stevens featured at the #NeurIPS SAIL Media Booth w/ @natolambert.bsky.social! Great convo on how AI accelerates scientific discovery (not just bigger models 🙃).
Clips ⬇️
youtube.com/shorts/v0k5D...
youtube.com/shorts/4R-BF...
#AI4Science #NeurIPS2025
Clips ⬇️
youtube.com/shorts/v0k5D...
youtube.com/shorts/4R-BF...
#AI4Science #NeurIPS2025
On a prompt count basis this mean reasoning models are not close to a majority on OpenRouter, as reasoning models can use 10-1000x the tokens of non-thinking models per prompt.
Lots of need for fast, efficient open models. Reasoning model usage is likely closed labs more.
Lots of need for fast, efficient open models. Reasoning model usage is likely closed labs more.
December 5, 2025 at 4:23 PM
On a prompt count basis this mean reasoning models are not close to a majority on OpenRouter, as reasoning models can use 10-1000x the tokens of non-thinking models per prompt.
Lots of need for fast, efficient open models. Reasoning model usage is likely closed labs more.
Lots of need for fast, efficient open models. Reasoning model usage is likely closed labs more.
Olmo 3 32B Think is now free on Open Router until 12/22.
Let us know what you think and what to improve!
(Hosted by Parasail)
This may give it the hug of death... would be my dream.
openrouter.ai/allenai/olmo...
Let us know what you think and what to improve!
(Hosted by Parasail)
This may give it the hug of death... would be my dream.
openrouter.ai/allenai/olmo...
December 5, 2025 at 12:49 AM
Olmo 3 32B Think is now free on Open Router until 12/22.
Let us know what you think and what to improve!
(Hosted by Parasail)
This may give it the hug of death... would be my dream.
openrouter.ai/allenai/olmo...
Let us know what you think and what to improve!
(Hosted by Parasail)
This may give it the hug of death... would be my dream.
openrouter.ai/allenai/olmo...
Where to find me at NeurIPS! Talks, panels, socials
Talks: I'm giving two talks on the last day at Workshops Dec. 7th
1. Good researchers obsess over evals.
10:35am-11:05am: Evaluating the Evolving LLM Lifecycle
2. Building Olmo 3 Think.
1:45pm-2:30pm: Foundations of Reasoning in Language Models
Talks: I'm giving two talks on the last day at Workshops Dec. 7th
1. Good researchers obsess over evals.
10:35am-11:05am: Evaluating the Evolving LLM Lifecycle
2. Building Olmo 3 Think.
1:45pm-2:30pm: Foundations of Reasoning in Language Models
December 2, 2025 at 4:25 PM
Where to find me at NeurIPS! Talks, panels, socials
Talks: I'm giving two talks on the last day at Workshops Dec. 7th
1. Good researchers obsess over evals.
10:35am-11:05am: Evaluating the Evolving LLM Lifecycle
2. Building Olmo 3 Think.
1:45pm-2:30pm: Foundations of Reasoning in Language Models
Talks: I'm giving two talks on the last day at Workshops Dec. 7th
1. Good researchers obsess over evals.
10:35am-11:05am: Evaluating the Evolving LLM Lifecycle
2. Building Olmo 3 Think.
1:45pm-2:30pm: Foundations of Reasoning in Language Models
Excited to share another NeurIPS event I'm helping with.
We're hosting a dedicated booth to record researchers talking about their work, share that audio&video content on our socials, and start great conversations.
calendly.com/contact-read...
We're hosting a dedicated booth to record researchers talking about their work, share that audio&video content on our socials, and start great conversations.
calendly.com/contact-read...
December 1, 2025 at 9:27 PM
Excited to share another NeurIPS event I'm helping with.
We're hosting a dedicated booth to record researchers talking about their work, share that audio&video content on our socials, and start great conversations.
calendly.com/contact-read...
We're hosting a dedicated booth to record researchers talking about their work, share that audio&video content on our socials, and start great conversations.
calendly.com/contact-read...
New Artificial Analysis Openness Index! As it should be, Olmo is king.
This plot feels like a solid representation of ability v openness. My favorite part is Olmo 3 mogging on Llama 4 Maverick in every plot 🤭. It's pretty sad that the two values are pretty much perfectly anti-correlated.
This plot feels like a solid representation of ability v openness. My favorite part is Olmo 3 mogging on Llama 4 Maverick in every plot 🤭. It's pretty sad that the two values are pretty much perfectly anti-correlated.
December 1, 2025 at 4:39 PM
New Artificial Analysis Openness Index! As it should be, Olmo is king.
This plot feels like a solid representation of ability v openness. My favorite part is Olmo 3 mogging on Llama 4 Maverick in every plot 🤭. It's pretty sad that the two values are pretty much perfectly anti-correlated.
This plot feels like a solid representation of ability v openness. My favorite part is Olmo 3 mogging on Llama 4 Maverick in every plot 🤭. It's pretty sad that the two values are pretty much perfectly anti-correlated.
Does anyone use the Deep Research features anymore. What's SOTA? My guy would guess Gemini 3 is decent?
I've just been so GPT Pro pilled. Why use DR?
I've just been so GPT Pro pilled. Why use DR?
November 30, 2025 at 3:28 PM
Does anyone use the Deep Research features anymore. What's SOTA? My guy would guess Gemini 3 is decent?
I've just been so GPT Pro pilled. Why use DR?
I've just been so GPT Pro pilled. Why use DR?
Olmo 3 is a banger paper. Finally getting to read the rest of it 😅.
We love releasing things that serve as a comprehensive snapshot of public knowledge on training leading language models.
There's an award for whoever finds all the secrets first in the new arxiv version
allenai.org/papers/olmo3
We love releasing things that serve as a comprehensive snapshot of public knowledge on training leading language models.
There's an award for whoever finds all the secrets first in the new arxiv version
allenai.org/papers/olmo3
November 29, 2025 at 7:57 PM
Olmo 3 is a banger paper. Finally getting to read the rest of it 😅.
We love releasing things that serve as a comprehensive snapshot of public knowledge on training leading language models.
There's an award for whoever finds all the secrets first in the new arxiv version
allenai.org/papers/olmo3
We love releasing things that serve as a comprehensive snapshot of public knowledge on training leading language models.
There's an award for whoever finds all the secrets first in the new arxiv version
allenai.org/papers/olmo3