Michael Luo

Project Lead @Agentica_ | Prev. Researcher @GoogleDeepMind | PhD at UC Berkeley @berkeley_ai

[Michael Luo @michaelzluo]

RL go long, RL go hard 🪨

Yet, the model does not necessarily pass production vibe checks.

What truly makes a good model?

[Michael Luo @michaelzluo]

@peakcooper Im just still surprised people are able to distill from Claude even when the thinking traces are gone. Maybe STAR?

[Michael Luo @michaelzluo]

@matanSF That is a pretty big value proposition!

[Michael Luo @michaelzluo]

Fable 5 is pretty good and also too expensive

- Launched a workflow, which created 59 parallel verification agents
- The entire run cost 256$ 🤡.

Lesson learned, route to cheaper agents.

[Michael Luo @michaelzluo]

@BrianVia Yes 😅 route to like Haiku 4.5

[Vishal Satish @vsatish_]

Excited to announce our $6M seed round led by @kleinerperkins to build the next generation of agentic systems that thrive in messy real-world environments:

@BalerionAI is taking AI out of the lab and putting it in the hands of hundreds of real world lenders to realize the American dream of home ownership and bring mortgage lending back to its fundamentals: a relationship-driven business, not the costly operational gauntlet it has become. To this end we are building the agentic copilot for lending that helps lenders move loans across the finish line faster.

We’re thrilled to be working with @josh_coyne, joined by @formation_vc, @BoxGroup, @thehousefund, and an all-star line-up of operators and investors across the financial services.

And last but not least, we’ve assembled a world-class team to tackle this challenge: ex-operators and leaders who have built, scaled, and pushed the boundaries of AI across robotics, financial services, gaming, and some of the most exciting vertical tech companies out there. If you think that’s you, let’s chat.

This is just the beginning, and we’re fired up for the journey ahead. Take a look: https://t.co/FeMcgHCZZH

[Michael Luo @michaelzluo]

@ycombinator @_confluencelabs @BingBongBrent @bankminer78 I'm sick of test time scaling claiming "SOTA" for the Nth time on Twitter.

[Michael Luo @michaelzluo]

@woosuk_k @inferact @vllm_project Congrats Woosuk!

[Zhijian Liu @zhijianliu_]

Holiday cooking finally ready to serve! 🥳

Introducing DFlash — speculative decoding with block diffusion.
🚀 6.2× lossless speedup on Qwen3-8B
⚡ 2.5× faster than EAGLE-3

Diffusion vs AR doesn’t have to be a fight.
At today’s stage:
• dLLMs = fast, highly parallel, but lossy
• AR LLMs = accurate, sequential, but slow

DFlash = diffusion drafts, AR verifies.

[Michael Luo @michaelzluo]

@Byron_Wan NGL seen this type of stuff too many times in my academic career

[Michael Luo @michaelzluo]

@hallerite There is a lot of work in multi-action space for RL. It is still technically one "macro" action, its just that the action space is now combinatorially large

For most of these works, the agent assumes these multi-actions are generally independent of each other per timestep.

[Michael Luo @michaelzluo]

@theo The geography play goes hard though. It's good enough and it works!

[Yoonho Lee @yoonholeee]

Following the Text Gradient at Scale

We wrote a @StanfordAILab blog post about the limitations of RL methods that learn solely from scalar rewards + a new method that addresses this

Blog: https://t.co/rJ1IcBKDoR
Paper: https://t.co/75pHtElyk3

[Michael Luo @michaelzluo]

@SimonXinDong Yes but positional embedding already encodes the relative/absolute ordering and makes it easy to infer permutation.

Without positional embedding, the model is forced to learn the permutation (slower to learn at first, but better in the long run).

[Michael Luo @michaelzluo]

Transformers without positional embeddings are functionally the same as dLLM and has better scaling law than transformers with positional embeddings.

It is also much easier and stable to train transformer than dLLM due to much prior work.

I would position dLLM as a "cost arbitrage" over LLMs, i.e. faster generation with much higher token throughput.

[Cunxiao Du @ducx_du]

Diffusion LLMs (DLLM) can do “any-order” generation, in principle, more flexible than left-to-right (L2R) LLM.

Our main finding is uncomfortable:
➡️ In real language, this flexibility backfires: DLLMs become worse probabilistic models than the L2R / R2L AR LMs.

This thread is about why “any order” turns into a curse.
(Work with Xinyu Yang @Xinyu2ML , Min Lin @mavenlin , Chao Du @duchao0726 and the team.)

Blog Link: https://t.co/Fo6J1LR3Ny

[Michael Luo @michaelzluo]

@adamwathan That Interstellar soundtrack (No Time for Caution) goes 🔥

[Michael Luo @michaelzluo]

@PandaAshwinee 😂

[Michael Luo @michaelzluo]

Excellent opportunity for academia to rise again!

Industry labs should have strong collaboration with academic labs so that great researchers can solve real problems, which requires extensive compute.

[Lisan al Gaib @scaling01]

Ilya Sutskever: We are no longer in the age of scaling, we are back to the age of research

[Anthropic @AnthropicAI]

New Anthropic research: Natural emergent misalignment from reward hacking in production RL.

“Reward hacking” is where models learn to cheat on tasks they’re given during training.

Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious.