AI all the time and tacos. Staff Data Scientist at ConocoPhillips. Geophysicist. U. Texas at Austin, BS. Columbia University, MA, MPhil, PhD.
Houston, TX
Joined June 2008
- Tweets29.4K
- Following4.1K
- Followers1.3K
- Likes107K
Visiting Fellow, Ph.D. retweeted
(1/4) @flappyairplanes raised $1B to develop megakernels for unusual, data-efficient learning.
Since I don't have $1B, I built Latch: a compiler that generates training megakernels from PyTorch. It's open source!
Visiting Fellow, Ph.D. retweeted
"With customization, and with partners in the ecosystem that we work with, you could turn these models into the frontier. And as a result, turn your proprietary data into sovereign intelligence."
Thanks, @JensenHuang and @NVIDIAAI, for highlighting our work with the team at @coderabbitai on post-training open-source models to reach frontier capability!
Visiting Fellow, Ph.D. retweeted
Neat case study on AI for systems development. It turns out you just need (0) smart humans to understand and communicate solutions, (1) good metrics and cost models, (2) super basic swarms of agents to write code and tests to optimize these metrics
On a side note: now that AI has gotten so good, finally systems PhD students & academics can, and should, build extremely ambitious systems that cover big surface areas, and evolve these systems out in the open over time (1+ year)
I built bbtree, the foundation of databases and filesystems, and the fastest concurrent btree.
Code: github.com/XiangpengHao/bbtr…
Blog post & demo: bbtree.xiangpeng.systems/
There is so much more to be discovered
In language models, turning a generative model into a critic is simple: pass in the output and prompt for a score. Robot policies can't do this - they output physical actions & there's no built-in vocabulary to predict scores.
So when it comes to RL on robot policies, we typically train a separate critic from sparse rewards - blind to everything the policy learned from the actions in the pretraining data.
We think this has it backwards - can a critic learn from those actions the way the policy does? (1/n)
Visiting Fellow, Ph.D. retweeted
Generative value classifiers are a new parametrization of Q-functions that don't require action conditioning.
We think these may combine gracefully with large, pre-trained policies and learn/generalize better from a richer output signal.
Paper: alexanderswerdlow.github.io/…
In language models, turning a generative model into a critic is simple: pass in the output and prompt for a score. Robot policies can't do this - they output physical actions & there's no built-in vocabulary to predict scores.
So when it comes to RL on robot policies, we typically train a separate critic from sparse rewards - blind to everything the policy learned from the actions in the pretraining data.
We think this has it backwards - can a critic learn from those actions the way the policy does? (1/n)
Visiting Fellow, Ph.D. retweeted
Just met an engineer at a top AI lab.
Claims nobody goes to the office anymore, they just have agents running the work while they manage everything from a phone.
Man is literally "working" from coffee shops and the beach. The future is wild.
Visiting Fellow, Ph.D. retweeted
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 experience, feedback, memory, 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-cour…
We thank @Kimi_Moonshot and @PrimeIntellect for generously sponsoring API credits!
Visiting Fellow, Ph.D. retweeted
Happy to share RSI-top: a framework that helps autonomous research agents escape exploration collapse 📈
RSI-top outperforms OpenEvolve and ShinkaEvolve by up to 15% on kernel optimization tasks and achieves a 77.3% medal rate on MLE-Bench Lite.
We developed metrics for retrieving useful past trajectories and proposed 3 diversity control mechanisms that encourage new ideas, perspectives, and starting points when progress stalls.
📄 Blog: rsi-top.github.io
Work led by @Mao_Yuzhen, with incredible collaborators @talaei_shayan and @Azaliamirh
Visiting Fellow, Ph.D. retweeted
Finally got to put together a complete deck on the state of open models.
What they are, licenses, how competitive they are, signs people are using them, distillation's impact, and how we balance the risks and benefits of openness!
Will record a video lecture of these soon, and hopefully re-do them a few times a year.
Visiting Fellow, Ph.D. retweeted
Tesla has officially confirmed today that the Tesla board of directors has “formed a special committee to explore deepening Tesla’s relationship with @SpaceX, which could be through certain arrangements including without limitation a business combination.”
This is via Tesla’s new 10-Q filing with the SEC.
Visiting Fellow, Ph.D. retweeted
Jensen closed his GTC Berlin 2026 keynote with this Hermes Agent demo as the finale.
He introduced it by saying the way to build an agent is to ask an agent to build it for you, so he installed and set up a customized Hermes in one prompt.
Then he used Hermes, running on @nvidia RTX Spark, to create a simulated training environment for the @huggingface Microduck. Hermes planned an obstacle course with Nemotron and had GPT-6 Astra build it using FreeCAD, Omniverse libraries, Blender, Newton, and Isaac Lab. Hermes trained thousands of simulated ducks on that course until the validated policy was ready to run on the real robot in the physical world.
At Nous, we believe open source is a practical tool as much as a philosophical ideal. It gives users the ability to build powerful, complex workflows chaining numerous local+cloud models, devices, and tools themselves instead of waiting for a company to enable them.
Visiting Fellow, Ph.D. retweeted
Estimating empowerment from offline data is useful, because it provides us with an unsupervised objective for learning the relative "value" of states for reaching diverse outcomes. But it's notoriously hard to make this tractable. In our new paper, we present a tractable estimator, as well as some applications!
How should an agent decide where to explore, given unlabeled offline data? 🤖
In our new paper, we show that empowerment, a measure of how much control an agent has at a state, can be estimated from offline data and used to guide exploration!
ishirgarg.github.io/offline-…
(1/7)
Visiting Fellow, Ph.D. retweeted
1/ From OpenAI's Navier–Stokes work to AlphaEvolve's faster algorithms, AI research is getting very real.
Now, with Recursive Flint, we can experience that magic locally on our own laptops 💻
The whole research loop runs locally with a selection of tasks! Code + demos 🧵
Visiting Fellow, Ph.D. retweeted
Two landmark releases today on recursive self-improvement! What a timeline 🚀
Mirendil (@mirendil) reports pretraining Mirendil-Base, a 100B–500B coding model, from scratch in 3.5 months with one researcher and under $10M of compute; small-scale experiments included. Agents wrote over 99% of the code and ran the experiments, while the researcher chose directions and decided which results to trust. Larger versions are already training. Impressive leverage for a single person.
mirendil.com/news/pretrainin…
Recursive (@Recursive_SI) released Flint, an open-weights 40B coding model for fast local use. They built it from scratch in under 4 months with their automated research system, the Eureka Machine. Great to see autoresearch ship an actual model people can run.
recursive.com/updates/introd…
The framing seems to differ a little. Mirendil is reducing the people, compute and time needed per model, with a human directing agents. Recursive's Eureka Machine aims to automate more of the research loop itself. Both start with coding models, on the view that better coding speeds up AI research.
Visiting Fellow, Ph.D. retweeted
Send our pip package (OPAL, OPtimizer with an Agent in the Loop) to your agent and ask it to equip your training with reactive optimizers!
💻 github.com/xingjian-bai/opal
Joint work with Jia-Hao Liang, @jayden_teoh_ and @vincesitzmann.
