Pinned Tweet
🚨 New Course Announcement!
This Fall @Diyi_Yang @jyangballin and I are teaching a BRAND NEW Stanford course on how to build AI Agents from scratch!
Millions of users interact with AI Agents every day. Come join us and learn how to build them!
🧵 (1/5)
This Fall @Diyi_Yang @jyangballin and I are teaching a BRAND NEW Stanford course on how to build AI Agents from scratch!
Millions of users interact with AI Agents every day. Come join us and learn how to build them!
🧵 (1/5)
This fall @michaelryan207 @jyangballin and I are teaching a new course CS329Z "Engineering AI Agents" on how to build AI Agents from scratch. Come join us and learn how to build them 🤖
25
267
1,812
118,158
Michael Ryan retweeted
Autumn 2026: @Diyi_Yang, @michaelryan207 & @jyangballin launch CS329Z Engineering AI Agents.
Not content with CS224N & CS336, @stanfordnlp leads with 2 amazing new courses, CS312 & CS329Z.
“But where are the videos?”, said the world.
Well, the first just dropped—more to come!
Not content with CS224N & CS336, @stanfordnlp leads with 2 amazing new courses, CS312 & CS329Z.
“But where are the videos?”, said the world.
Well, the first just dropped—more to come!
3
6
17
2,562
Having the most mind-numbing experience with #ICLR2027 reviews this year...
Reviewer: Reviews the wrong paper
Me: "I think you reviewed the wrong paper"
Reviewer: Decreases score to a 1, increases confidence to a 4, and pastes in 100% AI generated review
Me: ...
Reviewer: Reviews the wrong paper
Me: "I think you reviewed the wrong paper"
Reviewer: Decreases score to a 1, increases confidence to a 4, and pastes in 100% AI generated review
Me: ...
13
16
404
59,559
Michael Ryan retweeted
In @augmind_fm EP8, our @michaelryan207 had an insightful conversation with Peter Norvig, co-author of “Artificial Intelligence: A Modern Approach”
Two arguments really stuck with me:
1. While the scaling law hasn’t stopped and we now have much more data, algorithms, and ier is setting the right objective.
2. One of AI’s most profound risks is human deskilling and loss of motivation, which can have a long-lasting negative impact if we don’t do anything.
Peter also shared his career journey and his perspective on today’s AI industry. I found it super useful for early-career researchers and practitioners navigating such uncertain times.
Thank you, Peter, for sharing your wisdom and experiences with us!
Two arguments really stuck with me:
1. While the scaling law hasn’t stopped and we now have much more data, algorithms, and ier is setting the right objective.
2. One of AI’s most profound risks is human deskilling and loss of motivation, which can have a long-lasting negative impact if we don’t do anything.
Peter also shared his career journey and his perspective on today’s AI industry. I found it super useful for early-career researchers and practitioners navigating such uncertain times.
Thank you, Peter, for sharing your wisdom and experiences with us!
“I should have my choice. How much automation do I want helping me? But also, how much control do I want?”
“You have to struggle whether it’s lifting weights or whether it’s mental exercise. You’ve got to actually do the work yourself.”
For EP8, “After Data and Compute, What’s ,” we welcome Peter Norvig, co-author of Artificial Intelligence: A Modern Approach and former director of research at Google. Peter is now a Distinguished Education Fellow at Stanford HAI and a researcher at Recursive, working on self-improving AI.
As AI takes on more of our work, how do we make sure it makes us more capable? Peter explores what’s still worth learning in AI and programming, why learning still requires struggle, how more automation can coexist with more human control, and why getting AI’s objectives right remains one of the field’s biggest challenges.
Outline:
0:00 - Introduction
2:32 - What Has (and Hasn’t) Changed Throughout Peter’s Career?
7:03 - How Peter Makes Career Pivots
11:17 - Is This AI Moment Different From Previous Cycles?
15:32 - What’s Still Worth Learning in AI and Programming?
19:55 - How Will AI Transform Higher Education?
24:10 - Main Bottlenecks to Building Human-Centered AI
35:10 - Reward Hacking vs. Creativity
41:51 - Continual Learning Remains an Unsolved Challenge
44:16 - Open Source AI Models
46:32 - AI, Inequality, and Winner-Take-All Markets
1
5
18
1,304
Some of you know I’ve gotten deep into teaching recently, so I was thrilled to have the opportunity to sit down with Peter Norvig (one of the best AI educators to ever do it!!!)
Peter co-wrote THE textbook on AI!
In this discussion we talk about what is still important for humans to learn and do when it comes to AI development, and how we should think about setting the objectives for AI!
Check it out! ⬇️
Peter co-wrote THE textbook on AI!
In this discussion we talk about what is still important for humans to learn and do when it comes to AI development, and how we should think about setting the objectives for AI!
Check it out! ⬇️
“I should have my choice. How much automation do I want helping me? But also, how much control do I want?”
“You have to struggle whether it’s lifting weights or whether it’s mental exercise. You’ve got to actually do the work yourself.”
For EP8, “After Data and Compute, What’s ,” we welcome Peter Norvig, co-author of Artificial Intelligence: A Modern Approach and former director of research at Google. Peter is now a Distinguished Education Fellow at Stanford HAI and a researcher at Recursive, working on self-improving AI.
As AI takes on more of our work, how do we make sure it makes us more capable? Peter explores what’s still worth learning in AI and programming, why learning still requires struggle, how more automation can coexist with more human control, and why getting AI’s objectives right remains one of the field’s biggest challenges.
Outline:
0:00 - Introduction
2:32 - What Has (and Hasn’t) Changed Throughout Peter’s Career?
7:03 - How Peter Makes Career Pivots
11:17 - Is This AI Moment Different From Previous Cycles?
15:32 - What’s Still Worth Learning in AI and Programming?
19:55 - How Will AI Transform Higher Education?
24:10 - Main Bottlenecks to Building Human-Centered AI
35:10 - Reward Hacking vs. Creativity
41:51 - Continual Learning Remains an Unsolved Challenge
44:16 - Open Source AI Models
46:32 - AI, Inequality, and Winner-Take-All Markets
6
11
72
4,807
Michael Ryan retweeted
EP8 of the AM Podcast drops tomorrow!
We’re getting better at making AI achieve its goals. How do we make sure those are the goals we actually want?
We sat down with Peter Norvig, co-author of Artificial Intelligence: A Modern Approach and former director of research at Google, cores don’t always mean better AI, how we decide what these systems should optimize for, and what it takes to build AI that makes people more capable.
Here’s a preview 👀
We’re getting better at making AI achieve its goals. How do we make sure those are the goals we actually want?
We sat down with Peter Norvig, co-author of Artificial Intelligence: A Modern Approach and former director of research at Google, cores don’t always mean better AI, how we decide what these systems should optimize for, and what it takes to build AI that makes people more capable.
Here’s a preview 👀
0
4
7
560
Michael Ryan retweeted
Thank you @LaudeInstitute for the generous support!
So excited to work with @tatsu_hashimoto, @chrispiech, @erikbryn and our team to build collaborative AI that complements human expertise and helps people develop new skills as work evolves!
So excited to work with @tatsu_hashimoto, @chrispiech, @erikbryn and our team to build collaborative AI that complements human expertise and helps people develop new skills as work evolves!
Collaborative Intelligence for the Future of Work | @Stanford
@Diyi_Yang, @tatsu_hashimoto, @chrispiech, and @erikbryn are building AI systems that complement human expertise to equip the workforce of the future.
laude.org/moonshots/one/…
13
16
118
8,670
Michael Ryan retweeted
Collaborative Intelligence for the Future of Work | @Stanford
@Diyi_Yang, @tatsu_hashimoto, @chrispiech, and @erikbryn are building AI systems that complement human expertise to equip the workforce of the future.
laude.org/moonshots/one/…
@Diyi_Yang, @tatsu_hashimoto, @chrispiech, and @erikbryn are building AI systems that complement human expertise to equip the workforce of the future.
laude.org/moonshots/one/…
2
18
76
31,288
Michael Ryan retweeted
It’s done! Marin’s Mochi 535B-A23B finished its main pretraining phase this morning. 18T tokens! Long context this weekend, SFT over the next week or so. Hoping to release checkpoints next week. Some EU paperwork first. mumtracker.exe.xyz/hero-run-535b
42
60
634
50,300
This blog is sick! Astra can edit the weights of a smaller network to fix bugs like reasoning loops 🤯🤯🤯
My favorite part is they include all the logs which demystifies how this is possible (lots of careful reasoning over attention maps and activations). Astra even reconstructs the Pac-Man board entirely from neuron weights of a small network trained on Pac-Man!
The pattern they discuss of Astra first writing a python program for behavior, and then figuring out how to encode that into weights is super fascinating!
Feel the AGI 🔥
My favorite part is they include all the logs which demystifies how this is possible (lots of careful reasoning over attention maps and activations). Astra even reconstructs the Pac-Man board entirely from neuron weights of a small network trained on Pac-Man!
The pattern they discuss of Astra first writing a python program for behavior, and then figuring out how to encode that into weights is super fascinating!
Feel the AGI 🔥
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…Community noteLLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
6
1
47
4,537
Michael Ryan retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…
Community note
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
2,364
7,277
55,949
9,928,564
Michael Ryan retweeted
Synthetic rewrites are now standard to improve data quality and when data is limited. But how much compute should you spend rewriting? And can a model trained on rewrites beat the rewriting model?
We fit scaling laws to answer both:
1/🧵
We fit scaling laws to answer both:
1/🧵
10
20
201
32,229
Michael Ryan retweeted
PhDone!
Very grateful to all my committee members!
@Diyi_Yang @zsoltkira @kartik_goyal_ @PoloChau @sanmikoyejo
Very grateful to all my committee members!
@Diyi_Yang @zsoltkira @kartik_goyal_ @PoloChau @sanmikoyejo
42
22
255
30,410
Wow! 79 tasks not covered by O*Net in just six occupations!!
Super cool work helping to get a clearer picture on AI exposure in the workforce!
Super cool work helping to get a clearer picture on AI exposure in the workforce!
Jobs are modeled as bundles of tasks. But as jobs change, how do we keep those task lists up to date?
Workers often experience new responsibilities and tools before they reach their job descriptions.
We start by listening to them. 🧵
0
0
26
1,242
Michael Ryan retweeted
Jobs are modeled as bundles of tasks. But as jobs change, how do we keep those task lists up to date?
Workers often experience new responsibilities and tools before they reach their job descriptions.
We start by listening to them. 🧵
Workers often experience new responsibilities and tools before they reach their job descriptions.
We start by listening to them. 🧵
15
77
290
52,912
Michael Ryan retweeted
SECOND HUMANS& MODEL LFGGG
some cool things about Amp:
- Amp was trained in long open-ended simulations with Persimmon. This forced us to innovate across the RL stack from algos to infra
- Amp helps users achieve their goals better than any model we tested in our long horizon simulations spanning up to months of simulated time
- Amp is purple because purple is my favorite color
- Amp is the coolest thing I've ever worked on
- Amp was a giant effort across the whole company. Working on Amp meant working with and learning from:
- the indomitable @parth_29 who led the project ! ! !
- the GOAT @joeyxhong who's so crazy idek what to say man this guy's insane!!!
- @YuchenHe07 & @ericzelikman who're as good at RL as they are at founding companies (i.e. pretty ok)
- the infra goats @Ani_nlp & @rramador & @zianglih & @yawd2028 who keep us stepping fast
- the maniacs @niloofar_mire & @__Charlie_G & @vpj & @gharik who have all sorts of zany ideas on simulations
- the brilliant @314Bansal & @alexisjross with boundless of insight on using Persimmon better
- @noahdgoodman & @Diyi_Yang for professor-ly wisdom
- the prudent @eoflazer & @MaartenBosma & @getshao & @yuanlu119 for tasteful insights from the product side (more on that soon, stay tuned!)
- jim for bein cool and helpin us not break the law :)
if you wanna be on this list for our next model, join us
some cool things about Amp:
- Amp was trained in long open-ended simulations with Persimmon. This forced us to innovate across the RL stack from algos to infra
- Amp helps users achieve their goals better than any model we tested in our long horizon simulations spanning up to months of simulated time
- Amp is purple because purple is my favorite color
- Amp is the coolest thing I've ever worked on
- Amp was a giant effort across the whole company. Working on Amp meant working with and learning from:
- the indomitable @parth_29 who led the project ! ! !
- the GOAT @joeyxhong who's so crazy idek what to say man this guy's insane!!!
- @YuchenHe07 & @ericzelikman who're as good at RL as they are at founding companies (i.e. pretty ok)
- the infra goats @Ani_nlp & @rramador & @zianglih & @yawd2028 who keep us stepping fast
- the maniacs @niloofar_mire & @__Charlie_G & @vpj & @gharik who have all sorts of zany ideas on simulations
- the brilliant @314Bansal & @alexisjross with boundless of insight on using Persimmon better
- @noahdgoodman & @Diyi_Yang for professor-ly wisdom
- the prudent @eoflazer & @MaartenBosma & @getshao & @yuanlu119 for tasteful insights from the product side (more on that soon, stay tuned!)
- jim for bein cool and helpin us not break the law :)
if you wanna be on this list for our next model, join us
No one changes the world alone
Introducing Ampersand - the first model built for us. Ampersand is a collaborator trained to help people achieve their individual and collective goals. In multi-month simulations, users collaborating with Ampersand accomplished more than with Fable ls
Available to everyone through our API
23
11
126
8,668
Michael Ryan retweeted
Atlantic-CC // @michaelryan207, @WilliamBarrHeld, @kothasuhas, @YejinChoinka, @Diyi_Yang, @percyliang (@Stanford, Open Athena) — A 17T token model-based processing of the Common Crawl for training bigger and better open-source large language models.
1
2
11
996
Michael Ryan retweeted
I'm on the academic job market this year!
Recently I've been working on scaling laws and human-centered interventions for building more socially grounded AI.
If you're attending #COLM2026, please drop me a DM or email. I'd love to chat!
Recently I've been working on scaling laws and human-centered interventions for building more socially grounded AI.
If you're attending #COLM2026, please drop me a DM or email. I'd love to chat!
0
32
110
10,414
Michael Ryan retweeted
🚨new method day! 🚨
Trust[ing results] in ML conferences is utterly broken. Let's fix it with an algorithm!
We are excited to introduce 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗪𝗶𝘁𝗻𝗲𝘀𝘀𝗲𝘀, a system for “honesty” certificates of data and evaluations which a verifier can cheaply check and trust.
Trust[ing results] in ML conferences is utterly broken. Let's fix it with an algorithm!
We are excited to introduce 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗪𝗶𝘁𝗻𝗲𝘀𝘀𝗲𝘀, a system for “honesty” certificates of data and evaluations which a verifier can cheaply check and trust.
6
26
115
28,688
Michael Ryan retweeted
Data work seems to me to be relatively under-discussed as part of modern discussions about progress towards RSI.
Example: Ant. talked a lot about architecture and kernels in their RSI post, but not much about constitutional RL and synthetic data: anthropic.com/institute/recu…
Example: Ant. talked a lot about architecture and kernels in their RSI post, but not much about constitutional RL and synthetic data: anthropic.com/institute/recu…
7
13
149
12,833