rl and evals @ epiq ai labs. studied @DukeU. researched at @MSFTResearch, @CuraiHQ, @dynamo_ai.
Queens, NY
Joined February 2022
- Tweets526
- Following646
- Followers399
- Likes25.1K
Pinned Tweet
Excited to announce that two of my papers on synthetic data gen from our team at Epiq AI Labs were accepted to NeurIPS and EMNLP 2026!
Introducing InvestigationWorlds and CorporateBench... 🧵
albert yu sun retweeted
如今的亚洲首富张一鸣在30岁前的微博：
1. 年轻人花钱买时间要激进：省时间，用好时间。
2. 读传记，最能塑造品格、理想和动机。
3. 能快速抓住重点、说到重点，是两个很重要的素质。
4. 把时间换成小时来算，更能感知时间。
5. 延迟满足：沟通时情绪要沉住，人要稳住，思维才不失控。
6. 注意力也要开源节流：少欲望少杂念是节流，锻炼身体、训练注意力是开源。
7. 常问自己马上能做什么，做行动派。
my bro @CRColasurdo just moved back to bangkok, thailand! go follow him to see beautiful stills of thai culture and urbanism 🔥
every few months, I update my personal website with the latest codex models. every few months I'm pleasantly surprised by its improvements. this time, I'm impressed by the improvements in computer use latency.
check out this easter egg I added!
albert yu sun retweeted
Was reading the score-centering RL paper to add some notes on it the my book, and they have a great table in the appendix on the different importance sampling based train-inference mismatch methods in RL. These have been the biggest focus of RL algorithm research since the GRPO takeoff, imo.
albert yu sun retweeted
this is an abuse of technology 😭
one thing @_albertgu won't tell you is how well hnet tokenizes bad apple vs bpe.
suspicious! what else is he hiding???
Get PhD application advice from the GOATed students like @megan_richards_!
Power Broker-maxxing
Highly recommend folks read Caro’s book called Writing. Great self memoir abt how he immerses himself in his research and writes every thought down before editing. Writing = thinking. Great way to approach your life’s work.
Excited to announce that my benchmark, CorporateBench (EMNLP), has been added to the @FireworksAI_HQ Specialized Intelligence Index!
ailabs.epiqglobal.com/announ…
This quoted post is unavailable.
It's elegant how pedagogy analogizes well with IL vs RL. Imitation learning is like taking notes and studying them, whereas reinforcement learning is applying concepts you've learned and getting graded on real projects. Just like w/ robots, we often first do IL and then RL!
Cool to see how we can learn action values (Q-function) by modeling, from real trajectory data, which actions are likely under different return values!
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/…
albert yu sun retweeted
Tinker's new StableReinforce objective is Reinforce + Score Centering. Here's how I reverse-engineered it 🧵
Off-policy RL is difficult; runs can suddenly collapse because of async RL staleness or numerical mismatch.
StableReinforce is a new method on Tinker that keeps your RL runs stable, even when learning drifts from sampling.
tinker-docs.thinkingmachines…
albert yu sun retweeted
LLMs have a score doubling time of 53.9 days on a robotics benchmark between July and October 2026, while speed increased by 4.4x and cost fell by 6.2x.
More results in our new paper: The Bitter Lesson Comes to Robotics 🧵
albert yu sun retweeted
There is an (unpatchable?) supply-chain zero-day in PyTorch autograd that allows a regularization term to use the model's weights as RAM and SGD as its CPU clock to run any computation during training.
albert yu sun retweeted
China’s operating on a whole other level. We just can’t compete with this kind of juice in the US.
Just came out of an LED panel factory in Guangzhou.
Most notable thing after having toured European manufacturers for over a month last year: Nobody minded us taking pictures. In western factories that's taboo, because nobody wants any information about how they produce circulating.
I asked the production manager why they don't care here. His answer: "You can see how we do it, but each of these machines costs ¥3m (~€400k) and we run dozens of them." The moat is capital. There's also a default assumption that everything gets copied anyway, so nobody builds their business on production secrets.
Another big reason is that showing the line is marketing. About 95% of companies selling LED panels here are trading companies. Middlemen = somebody's taking a cut = you're paying a bad price. Letting you see the factory proves they actually produce, shows the quality of the components, and signals you're getting a fair price.
Same with privacy. In Europe, filming workers beyond their hands is a no-go. Here people glanced up at three people filming them and continued working. Try that in a Bavarian SME.
Some numbers:
> 95% automated: that means machines assemble 95% of components, humans 5%
> Factory runs 24/7 on two 12-hour shifts. In Europe this would be 3x 8h
> 60h weeks, two days off, just not always the weekend
> Guangdong minimum wage is ¥2,680 (€360) per month, so ~€1.40/hour
> Employer cost is ¥4,100 though and employee pays ¥700 in taxes. Those are all pension and social contributions. There's no income tax on salaries below ¥5,000
The owner is called Cha-Ming, so he goes by Charming for foreigners. He definitely enjoys leaning into this name for branding.
albert yu sun retweeted
Just came out of an LED panel factory in Guangzhou.
Most notable thing after having toured European manufacturers for over a month last year: Nobody minded us taking pictures. In western factories that's taboo, because nobody wants any information about how they produce circulating.
I asked the production manager why they don't care here. His answer: "You can see how we do it, but each of these machines costs ¥3m (~€400k) and we run dozens of them." The moat is capital. There's also a default assumption that everything gets copied anyway, so nobody builds their business on production secrets.
Another big reason is that showing the line is marketing. About 95% of companies selling LED panels here are trading companies. Middlemen = somebody's taking a cut = you're paying a bad price. Letting you see the factory proves they actually produce, shows the quality of the components, and signals you're getting a fair price.
Same with privacy. In Europe, filming workers beyond their hands is a no-go. Here people glanced up at three people filming them and continued working. Try that in a Bavarian SME.
Some numbers:
> 95% automated: that means machines assemble 95% of components, humans 5%
> Factory runs 24/7 on two 12-hour shifts. In Europe this would be 3x 8h
> 60h weeks, two days off, just not always the weekend
> Guangdong minimum wage is ¥2,680 (€360) per month, so ~€1.40/hour
> Employer cost is ¥4,100 though and employee pays ¥700 in taxes. Those are all pension and social contributions. There's no income tax on salaries below ¥5,000
The owner is called Cha-Ming, so he goes by Charming for foreigners. He definitely enjoys leaning into this name for branding.
albert yu sun retweeted
Today, Dream Labs is coming out of stealth with Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.
Dream API is live for everyone, today. 🧵
albert yu sun retweeted
i regret to inform you that personal growth rarely comes from acquiring new knowledge and almost always from:
-getting humiliated
-showing up terrified and doing it anyway
-admitting you might be the problem
albert yu sun retweeted
We are open-sourcing V1.0 of our agentic robotics framework, ASPIRE, with full real robot skill discovery and self improvement capabilities! On the day Astra arrived, our lab switched the backend model and saturated all the benchmarks to ~99% over night.
And the frontier is already on the move:
- Opus 5.5 slightly edges out Astra on our robot simulation bench.
- Sonnet 5.5 hits the Pareto frontier of cost vs. robot capability, $12/task.
We first introduced ASPIRE and /robot skill library in June, way before the big models could reliably do robotics. A rising frontier tide lifts all boats, but Astra was a 0->1 tsunami on many tasks.
Now you can also feel the Physical Bitter Lesson at home like we did on that fateful day ;)
links in thread