Co-founder of @Recursive_SI. ex-Meta FAIR Director. ex-Google. Reasoning, Optimization and Understanding LLM. Novelist in spare time. PhD in @CMU_Robotics.
Joined December 2009
- Tweets 1,138
- Following 939
- Followers 44,291
- Likes 3,185
128 Photos and videos
Pinned Tweet
Jun 11
Early results from Recursive 🚀🚀
SotA results from our open-ended knowledge discovery system:
1️⃣NanoChat 5min pre-training (0.9372 bpb -> 0.9109 bpb, 2.8% lower Bits-Per-Byte than long-standing community SoTA)
2️⃣NanoGPT SpeedRun (79.7s -> 77.5s, 2.8% faster than long-standing community SoTA)
3️⃣GPU kernel optimization (Overall 7.8% better than SoTA performance in SOL- ExecBench, hosted by NVIDIA)
To achieve that, our system automatically finds and combines innovations together to create better solutions than current ones carefully designed by expert humans in various domains.
We have open-sourced resulting artifacts found by our system so you can check the output yourself. See a full breakdown and technical writeup:
recursive.com/articles/first…
6
35
351
52,395
Jun 12
Detailed explanation of our first results by our amazing teammate @cong_ml !
Recursive just came out of stealth, and the team has been cooking 🔥
Our first results: an automated AI research system that can improve AI across 3 very different settings across training and GPU kernel optimization.
recursive.com/articles/first…
1
3
66
17,445
initial results on automated ai research from our team on small scale pre-training and kernel optimization. we also open-source the corresponding artifacts.
it's been great seeing all the amazing progress here in such a short time!
2
5
54
7,157
Yuandong Tian retweeted
Jun 11
Excited to share these preliminary results on our internal autoresearch system @Recursive_SI, where we achieve SOTA on nanochat / nanogpt speedrun / kernel benchmarks using the same underlying system without task-specific adaptations.
blog: recursive.com/articles/first…
3
24
108
15,963
Jun 12
In GPU kernel optimization, our framework achieves overall SoTA in NV's SOL-ExecBench, SoTA in all 4 sub-categories, and outperforms solutions that are (1) designed by GPU human experts, and (2) generated by other AI systems designed by GPU experts.
Our system is general. We don't have in-house GPU experts for now (I am regarded as a kernel "expert" internally😆).
🔍Check the details in NV's official leaderboard research.nvidia.com/benchmar…
Jun 11
AI is now doing our AI research.
At Recursive we set out to build recursive self-improving superintelligence (RSI) to automate knowledge discovery. The best way to expand humanity’s knowledge is through the scientific method.
RSI leads to better ideas, explanations and inventions which lead to better RSI. Automating the scientific method requires closing the loop between ideation, implementation and validation, and being able to run it over extended periods of time.
Today, we are excited to share the first outputs of Recursive’s automated open-ended discovery system. To be clear, this system is merely a milestone towards RSI, a v0.1 of what I sometimes call the “Eureka Machine”. It is one program that you can point at any hard problem and get useful inventions out. Though it’s still very early, we've run it on three AI tasks and achieved state-of-the-art results on all three.
These results demonstrate that even this early version of the system can solve a variety of autoresearch problems in AI and improve over prior state of the art. Concretely, it did this on the community benchmarks NanoGPT speedrun, NanoChat, and NVIDIA's Sol-ExecBench.
AI is code and AI can code.
The code and ideas that lead to these results were not invented by our team but by the AI system itself.
To do RSI safely, we need to investigate its inventions. That's best done transparently with the community.
@Recursive_SI we are open-sourcing the system’s discoveries, demonstrating that it finds creative and benign solutions instead of focusing on obvious optimizations or dangerous ideas.
Link below.
10
15
140
24,716
May 26
Great work @YujieZhao455906! In this work, we build
1️⃣ AMA-Bench that evaluates long-horizon memory in real-world tasks;
2️⃣ AMA-Agent that tracks causality and objective info across long trajectories to achieve better performance than existing agentic designs.
The paper has been accepted in #ICML2026. Welcome to use😀 Project link in ama-bench.github.io.
May 26
🚀 Excited to share that our work, AMA-Bench, has been accepted to #ICML2026!
Most benchmarks test dialogue memory, but real agents learn through continuous environment interactions. We actually found that systems acing dialogue benchmarks completely struggle in true agentic settings! 🤯
To fix this, we introduce AMA-Bench to evaluate long-horizon memory in real applications, plus AMA-Agent—a new system designed to track causality and objective info across long trajectories. 🧠
🌐 Check it out: ama-bench.github.io/
See you at ICML! 🎉
3
13
106
22,175
Yuandong Tian retweeted
May 13
some news: I've joined @Recursive_SI as a member of the founding team in London.
We are building safe, recursively self-improving intelligence. So excited about our discoveries so far and even more so about everything that lies ahead.
May 13
Excited to co-found Recursive (@recursive_si) with an exceptional team in London and SF to create AI that experiments on how to safely improve itself, turning compute into knowledge that accumulates in an open-ended process of endless, automated scientific discoveries.
19
17
339
55,022
Yuandong Tian retweeted
May 13
Some updates: I've joined Recursive as a member of the founding team.
My core curiosity about the world centers on how complex patterns and knowledge emerge from the two open-ended processes we know: natural and cultural evolution. I've been lucky to explore this during my PhD through works like ADAS, Darwin Gödel Machine, and The AI Scientist. Excited to keep chasing this thread with the incredible team!
17
16
155
22,753
Yuandong Tian retweeted
May 13
Personal update: I’ve joined @Recursive_SI as a member of the founding team.
We’re working on safe, recursively self-improving intelligence. I’m excited by what we’ve learned so far, and even more excited for the work ahead!
4
9
102
12,720
Yuandong Tian retweeted
May 13
Life update: I’ve joined @Recursive_SI as part of the founding team, based in San Francisco.
At Recursive, we’re building safe, recursively self-improving superintelligence to automate knowledge discovery.
Grateful to be part of this team ❤️, and excited for what’s ahead 🔥.
6
9
91
118,740
Yuandong Tian retweeted
May 13
Today, @recursive_si is coming out of stealth.
Over the past few years, I’ve become increasingly convinced that Recursive Self-Improvement (RSI) will define one of the most important frontiers in AI.
Now feels like the right moment to help turn these possibilities into reality.
I’m excited to join as a founding member and help build practical and safe self-improving AI systems.
Thanks to all of our co-founders for the vision and opportunity ♥️
10
12
66
15,468
May 13
Great to work together along the journey @_rockt!
May 13
Excited to co-found Recursive (@recursive_si) with an exceptional team in London and SF to create AI that experiments on how to safely improve itself, turning compute into knowledge that accumulates in an open-ended process of endless, automated scientific discoveries.
1
43
6,067
Yuandong Tian retweeted
May 13
Today we’re launching Recursive (@Recursive_SI)
We’re building AI that automates science, starting with the science of how to improve itself.
I’ve spent a lot of time building AI products and tools for AI teams. One thing that has always stood out is how much progress depends on the experimental loop: deciding what to try, implementing it, running it, understanding what happened, and repeating.
Recursive is automating that loop, safely and at scale.
Excited to work on this with an incredible team across SF and London.
16
24
173
22,726