Nathan Lambert and Sebastian Raschka are machine learning researchers, engineers, and educators. Nathan is the post-training lead at the Allen Institute for AI (Ai2) and the author of The RLHF Book. S
Large language models don't "think" the way humans do. We start with thoughts and ideas, then encode them into language and words to communicate. the entire reason we have language is to encode meaning and ideas and share that with others. so language IS encoded knowledge
LLMs do this in the opposite direction. trained on massive text, they excel at language prediction—completing patterns based on what's probable in their training data. Given a prompt, they generate coherent continuations that mimic human writing. Because language encodes meaning, knowledge, and logic, the output often appears intelligent. it is a result of finishing the pattern and structure of the response. It's a powerful trick cracking open the language and reading the embedded knowledge it contains. humans go from ideas to language, but the LLMs generate language, and the meaning emerges as a byproduct.
LLMs could be said to "think" in some emergent way through their complex computations, but they're not thinking about "theory of mind" or reasoning like humans. Even "deep reasoning" models just run extra cycles of the same pattern-completion trick, layering predictions to simulate step-by-step logic. It feel
Очень многогранный разговор.Я не профессионал индустрии,но мне очень хорошо зашла,как "температура" самого разговора,так и его формат.И я почерпнула много для себя интересного.Спасибо Вам большое.
Fortunately we have people like you, Lex... please please KEEP GOING ON with these podcasts... nowadays there is NOTHING as valuable as this to keep informed about what's going on in the world! Thanks a lot Lex and keep doing this SO WELL AND SO GOOD!.
So good, Lex Nathan and Sebastian! Great honor and learning opportunity, by sharing you are helping the world by creativity of building it yourself! Thankyou.
I listened to an episode you had on Python years ago, while I tried to learn enough coding to use Pi and arduino... I kept listening, the Neil Gershenfeld interview blew my mind. If it wasnt for your show, I wouldnt have been able to understand this shift. Thanks!
Dear Lex Fridman — I first found your long-form interviews through your conversations with chess minds I’ve respected for years, and I stayed because your work has a rare quality: it makes people think without making them feel preached at.
In a world optimized for outrage and speed, you’ve chosen depth, patience, and real curiosity. The questions are technical when they should be technical, human when they should be human — and the tone is steady enough that the truth has room to show itself. That’s not just content; it’s craft, and it’s service.
Thank you for building a place on the internet where nuance still survives. If there were a serious award for information stewardship — for turning noise into understanding — you’d be at the top of the list.
With respect and gratitude from Malaysia — from a fellow AI enthusiast, chess player, and grappler.
I started working with AI last year by getting an NVIDIA DGX spark. My sons and I took Python coding classes over last summer. We are developing a Llama model for my business to submit data to multiple platforms we work on. It is going well and has been a learning experience.
Hellllll yeah, back to the AI podcast days. Please do more technical conversations. Interviewing public figures and politicians might be great for reach, I get that. I’ve been listening since your very first podcasts and my favourite have always been conversations like these. More researchers, engineers, CTOs and the occasional philosopher
I have a deal with myself at the gym - I can’t get off the treadmill until the video ends. Accidentally clicked this video, and now my legs have fallen off
Thank you for listening ❤ Check out our sponsors: lexfridman.com/sponsors/ep490-sa
See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc.
0:00 - Introduction
1:57 - China vs US: Who wins the AI race?
10:38 - ChatGPT vs Claude vs Gemini vs Grok: Who is winning?
21:38 - Best AI for coding
28:29 - Open Source vs Closed Source LLMs
40:08 - Transformers: Evolution of LLMs since 2019
48:05 - AI Scaling Laws: Are they dead or still holding?
1:04:12 - How AI is trained: Pre-training, Mid-training, and Post-training
1:37:18 - Post-training explained: Exciting new research directions in LLMs
1:58:11 - Advice for beginners on how to get into AI development & research
2:21:03 - Work culture in AI (72+ hour weeks)
2:24:49 - Silicon Valley bubble
2:28:46 - Text diffusion models and other new research directions
2:34:28 - Tool use
2:38:44 - Continual learning
2:44:06 - Long context
2:50:21 - Robotics
2:59:31 - Timeline to AGI
3:06:47 - Will AI replace programmers?
3:25:18 - Is the dream of AGI dying?
3:32:07 - How AI will make money?
3:36:29 - Big acquisitions in 2026
3:41:01 - Future of OpenAI, Anthropic, Google DeepMind, xAI, Meta
3:53:35 - Manhattan Pro
We use cookies, device fingerprinting and cross-device tracking to personalise content, serve targeted advertising, and analyse how you use our site across your devices. By clicking Accept all you consent to the use of these tracking technologies. Your data may be used to build a profile of your interests and show you personalised ads on other sites.
Privacy policy
Comments (20)
Daily Junction discussion mixed with clearly attributed comments from the original video provider.
Join in — free. Comments on Daily Junction are for members, so real names stay rare and bots stay out.
One field. We email you a 6-digit code — no password needed. Your comment is kept while you do it.
Under 13? You’ll need a parent’s OK first — it takes them one click.
LLMs do this in the opposite direction. trained on massive text, they excel at language prediction—completing patterns based on what's probable in their training data. Given a prompt, they generate coherent continuations that mimic human writing. Because language encodes meaning, knowledge, and logic, the output often appears intelligent. it is a result of finishing the pattern and structure of the response. It's a powerful trick cracking open the language and reading the embedded knowledge it contains. humans go from ideas to language, but the LLMs generate language, and the meaning emerges as a byproduct.
LLMs could be said to "think" in some emergent way through their complex computations, but they're not thinking about "theory of mind" or reasoning like humans. Even "deep reasoning" models just run extra cycles of the same pattern-completion trick, layering predictions to simulate step-by-step logic. It feel
In a world optimized for outrage and speed, you’ve chosen depth, patience, and real curiosity. The questions are technical when they should be technical, human when they should be human — and the tone is steady enough that the truth has room to show itself. That’s not just content; it’s craft, and it’s service.
Thank you for building a place on the internet where nuance still survives. If there were a serious award for information stewardship — for turning noise into understanding — you’d be at the top of the list.
With respect and gratitude from Malaysia — from a fellow AI enthusiast, chess player, and grappler.
See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc.
0:00 - Introduction
1:57 - China vs US: Who wins the AI race?
10:38 - ChatGPT vs Claude vs Gemini vs Grok: Who is winning?
21:38 - Best AI for coding
28:29 - Open Source vs Closed Source LLMs
40:08 - Transformers: Evolution of LLMs since 2019
48:05 - AI Scaling Laws: Are they dead or still holding?
1:04:12 - How AI is trained: Pre-training, Mid-training, and Post-training
1:37:18 - Post-training explained: Exciting new research directions in LLMs
1:58:11 - Advice for beginners on how to get into AI development & research
2:21:03 - Work culture in AI (72+ hour weeks)
2:24:49 - Silicon Valley bubble
2:28:46 - Text diffusion models and other new research directions
2:34:28 - Tool use
2:38:44 - Continual learning
2:44:06 - Long context
2:50:21 - Robotics
2:59:31 - Timeline to AGI
3:06:47 - Will AI replace programmers?
3:25:18 - Is the dream of AGI dying?
3:32:07 - How AI will make money?
3:36:29 - Big acquisitions in 2026
3:41:01 - Future of OpenAI, Anthropic, Google DeepMind, xAI, Meta
3:53:35 - Manhattan Pro