Pinned Tweet
Wrote a short blog about the "shape" of language models, and the tradeoffs they may present in the future. I genuinely think it's a valuable research direction to start thinking about now, especially w.r.t. harness design.
alexzhang13.github.io/blog/2026/shap…
alexzhang13.github.io/blog/2026/shap…
51
257
2,393
291,314
alex zhang retweeted
Earlier this year, @a1zhang introduced the idea of compositional generalization, showing how harnesses can help LLMs generalize to harder tasks by decomposing them.
In our new blog post, we show that harnesses for Vision Language Models (VLMs) are compositional generalizers image-based questions into simpler observations the model can solve. (1/8)
harishkk.com/writing/visual…
In our new blog post, we show that harnesses for Vision Language Models (VLMs) are compositional generalizers image-based questions into simpler observations the model can solve. (1/8)
harishkk.com/writing/visual…
5
14
122
8,795
alex zhang retweeted
The team that launched this challenge took the @GPU_MODE OSS code, asked an agent to rewrite it in Go, copied it, didn't credit us, shipped a closed source binary to hide it all and are not committing releasing the entries.
Full receipts here github.com/msaroufim/open…
Full receipts here github.com/msaroufim/open…
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev
22
46
716
54,804
alex zhang retweeted
I feel deeply honored to co-teach this course on Self-Evolving Agents with @pliang279 at MIT this semester.
I often wished more courses would bring the latest research and discussions from the AI frontier into the classroom. It feels special to help make that happen now.
We designed the course to welcome students from different backgrounds, combining foundational methods and the latest research with hands-on practice in building modern agent harnesses, post-training models, and more. Our hope is that everyone can take something useful back to their own projects and research.
For each module, we invite researchers to share their work and answer students’ questions. Over the past few weeks, these conversations have brought cutting-edge research on context, harness, and model evolution directly into our classroom.
A huge thank-you to our amazing guest speakers:
• @qizhengz_alex for Agentic Context Engineering
• @a1zhang for Recursive Language Models
• @IdanShenfeld for Self-Distillation Enables Continual Learning
I also want to thank @Kimi_Moonshot and @PrimeIntellect for generously supporting the class with model API credits and GPU resources. Their support is essential to the class’s success.
I often wished more courses would bring the latest research and discussions from the AI frontier into the classroom. It feels special to help make that happen now.
We designed the course to welcome students from different backgrounds, combining foundational methods and the latest research with hands-on practice in building modern agent harnesses, post-training models, and more. Our hope is that everyone can take something useful back to their own projects and research.
For each module, we invite researchers to share their work and answer students’ questions. Over the past few weeks, these conversations have brought cutting-edge research on context, harness, and model evolution directly into our classroom.
A huge thank-you to our amazing guest speakers:
• @qizhengz_alex for Agentic Context Engineering
• @a1zhang for Recursive Language Models
• @IdanShenfeld for Self-Distillation Enables Continual Learning
I also want to thank @Kimi_Moonshot and @PrimeIntellect for generously supporting the class with model API credits and GPU resources. Their support is essential to the class’s success.
The next frontier is AI that improves itself, and we believe everyone should know how to build it.
This semester, Ao @ao_qu18465 and I are teaching 'How to AI an AI for Almost Anything' @MIT @medialab -- a new course on self-evolving AI systems that recursively improve from emory, and exploration over long interaction horizons.
The key idea: intelligence shouldn't stop improving when training ends.
We cover four levels of evolution:
(1) context evolution through prompt optimization and context management,
(2) harness evolution through memory, tools & skills, workflow optimization, and multi-agent systems,
(3) model evolution through RL, distillation, test-time training, self-play, and continual learning,
(4) task evolution through generating increasingly complex synthetic tasks, simulations, evaluations, and environments.
Slides, readings, and videos available: mit-mi.github.io/how2ai-course/…
We thank @Kimi_Moonshot and @PrimeIntellect for generously sponsoring API credits!
10
27
220
12,243
alex zhang retweeted
Introducing Slingshots // FOUR: 14 projects across eight top institutions. Self-improving agent systems and harnesses, maximally-open models and training data, applications for research and decision-making – plus kernels and more. Say hello to the batch 🧵
6
31
90
16,242
thanks @NoahZiems for the reminder, but it’s been 1 year since the first blog of my PhD 😅 time really flies
it’s funny that this came out before I took my first exam too lol
it’s funny that this came out before I took my first exam too lol
What if scaling the context windows of frontier LLMs is much easier than it sounds?
We’re excited to share our work on Recursive Language Models (RLMs). A new inference strategy where LLMs can decompose and recursively interact with input prompts of seemingly unbounded length, On the OOLONG benchmark, RLMs with GPT-5-mini outperforms GPT-5 by over 110% gains (more than double!) on 132k-token sequences and is cheaper to query on average.
On the BrowseComp-Plus benchmark, RLMs with GPT-5 can take in 10M+ tokens as their “prompt” and answer highly compositional queries without degradation and even better than explicit indexing/retrieval.
We link our blogpost, (still very early!) experiments, and discussion below.
8
4
122
5,182
alex zhang retweeted
Surprising new results:
Giving tools to LLMs makes them less able to say NO!
We tested 11 multimodal LLMs (Claude, GPT, Gemini, ...). With tools, every one fails to refuse harmful requests more often, up to 68.7% more!
Details and possible reasons 🧵
To appear at NeurIPS’26.
Giving tools to LLMs makes them less able to say NO!
We tested 11 multimodal LLMs (Claude, GPT, Gemini, ...). With tools, every one fails to refuse harmful requests more often, up to 68.7% more!
Details and possible reasons 🧵
To appear at NeurIPS’26.
9
23
167
17,519
the prime boys are absolutely cooking these days
Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust.
With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages, this was our largest test yet of Prime Agent’s multi-agent capabilities and our infrastructure arms.
Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.
2
7
156
12,259
alex zhang retweeted
Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust.
With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages, this was our largest test yet of Prime Agent’s multi-agent capabilities and our infrastructure arms.
Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.
With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages, this was our largest test yet of Prime Agent’s multi-agent capabilities and our infrastructure arms.
Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.
185
208
2,847
556,648
alex zhang retweeted
Added @a1zhang 's Speculative Programmatic Tool Calling to @pydantic AI, and I've been getting nice results from it.
There is pretty much no drawback to Code Mode that I can see.
github.com/mpfaffenberger… has had it for awhile now, and I added a little statusline to show you from speculated tool calls.
It's basically just free speed.
pydantic.dev/docs/ai/harnes…
There is pretty much no drawback to Code Mode that I can see.
github.com/mpfaffenberger… has had it for awhile now, and I added a little statusline to show you from speculated tool calls.
It's basically just free speed.
pydantic.dev/docs/ai/harnes…
4
6
38
3,936
alex zhang retweeted
I’m excited to share that I’m joining @coreauto as a Member of Technical Staff, focusing on RL performance! The team is genuinely amazing, and I couldn’t be more excited about what we’re building.
The past year has been quite wild. I got to work on pre-training a frontier open model, helped build RL infrastructure for 1T+ parameter models, and moved across the world to the US after getting two O-1 petitions approved. I feel incredibly fortunate to be working in this field at a time like this, and I’ve learned more than I could have imagined.
I’m incredibly grateful to everyone at @PrimeIntellect for the opportunity and especially the amazing people I’ve gotten to work with. It’s been a huge part of my journey, and I’ll be rooting for the team from the sidelines as they continue building open superintelligence.
The past year has been quite wild. I got to work on pre-training a frontier open model, helped build RL infrastructure for 1T+ parameter models, and moved across the world to the US after getting two O-1 petitions approved. I feel incredibly fortunate to be working in this field at a time like this, and I’ve learned more than I could have imagined.
I’m incredibly grateful to everyone at @PrimeIntellect for the opportunity and especially the amazing people I’ve gotten to work with. It’s been a huge part of my journey, and I’ll be rooting for the team from the sidelines as they continue building open superintelligence.
82
9
618
48,018
RLMs + the compositional generalizers blog (alexzhang13.github.io/blog/2026/harn…) in this year's @stateofai report!
What a year it's been
What a year it's been
welcome to the 9th annual @stateofai report!
ai building better ai, world models, physical ai, geopolitics, cyber defense and 12-month predictions.
read it at stateof(dot)ai or watch my editor’s cut here.
a few of this year’s biggest findings follow:
11
14
242
10,749
alex zhang retweeted
welcome to the 9th annual @stateofai report!
ai building better ai, world models, physical ai, geopolitics, cyber defense and 12-month predictions.
read it at stateof(dot)ai or watch my editor’s cut here.
a few of this year’s biggest findings follow:
ai building better ai, world models, physical ai, geopolitics, cyber defense and 12-month predictions.
read it at stateof(dot)ai or watch my editor’s cut here.
a few of this year’s biggest findings follow:
46
127
694
146,619
anyone in SF this week should take some time to chat with souradip :)
souradip and I have been working pretty closely the past couple of months, so I can say with confidence that you’ll have a good conversation! he’s one of the rarer individuals in our field with strong fundamentals
souradip and I have been working pretty closely the past couple of months, so I can say with confidence that you’ll have a good conversation! he’s one of the rarer individuals in our field with strong fundamentals
I’ll be in San Francisco this week and also attending the Simons workshop! If you're around and want to catch up, please reach out. I'd love to chat about Pedagogical RL, diversity, alignment and anything beyond standard RL.
3
4
125
12,111
alex zhang retweeted
it’s both surprising and fairly obvious why harnesses work so well to increase the generalization capabilities of LLM
in this 1h30 chat with big boss @a1zhang we cover:
- compositional generalization
- having stuff locally in-distribution
- why code is it
- what’s still missing
in this 1h30 chat with big boss @a1zhang we cover:
- compositional generalization
- having stuff locally in-distribution
- why code is it
- what’s still missing
3
5
86
3,555
alex zhang retweeted
As models get smarter I am more and more convinced that @a1zhang's mismanaged genius nomenclature is representative of whats going on
There is no reason for these models to struggle the way they do at this point
There is no reason for these models to struggle the way they do at this point
1
2
62
6,150
alex zhang retweeted
@StijnSmits @a1zhangh@willcbt@lateinteractionhost bridge that gives it access to harness primitives so it can do lots of things programmatically. MCP tool calling, skill creation, agent profile creation, LLM calling, delegation, and more :) The python repl is also perfect for doing computational science since bulky datasets stay in memory as agents iterate on their analysis!
9
19
220
21,589
alex zhang retweeted
late to reading this, but FUCK YES to playing with more shapes for intelligence!
please play around with it more! append-only chat is not the final interface!
please play around with it more! append-only chat is not the final interface!
Wrote a short blog about the "shape" of language models, and the tradeoffs they may present in the future. I genuinely think it's a valuable research direction to start thinking about now, especially w.r.t. harness design.
alexzhang13.github.io/blog/2026/shap…
23
53
969
70,445
alex zhang retweeted
SFT is not dead! 🥳
We found a way to make SFT rival current prevailing posttraining methods, often generalizing better and forgetting less than RL and OPSD. 🤯
Following our prior work on reasoning with sampling, we now introduce sampling to the posttraining stack.
1/n
We found a way to make SFT rival current prevailing posttraining methods, often generalizing better and forgetting less than RL and OPSD. 🤯
Following our prior work on reasoning with sampling, we now introduce sampling to the posttraining stack.
1/n
74
244
2,166
367,154
