Michael Zhang (@mzhangio) profile picture

new stuff, tasty agents cs phd @hazyresearch, @StanfordAILab; founding tech staff @periodiclabs

Stanford, CA
Joined February 2016
Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)Photo by Michael Zhang (@mzhangio)
according to Fable, we are *all* too dangerous. @HazyResearch in good hands w this one (also excited to share new stuff, soon!)
According to Fable, I am now a Cogito Hazard.
1
13
3,274
iykyk (i don’t actually know, but text in Instinct’s artifacts smells a lot like that passive voice from Claude) ig this is the Cursor way? also Harvey + other “wrappers”, but i feel that term is too diminutive for what’s pretty great: 1. find experience where some automation via AI makes it more delightful 2. use that user interaction data to figure out features + collect training data 3. eventually train own models to make money on the lower cost and that data capture @cursor_ai @harvey @rachelsupark
2
6
876
Replying to @madisonleaho
@madisonleaho lol come back for more research mentoring rolling your own personal agent works p well e.g., playwright-cli + @cartesia + claude* (+ optional continual learning reflection harness :) also featuring @BaigYasa *this was in june; pick your favorite
i’ve tried every ai assistant out there and yet none of them can connect to my school outlook can someone help a college student out and convince stanford IT to get email connector working pls
3
1
1
15
2,580
real demo / extended cut without provocative driving edit TTS kinda fun as a way to demo reasoning + computer-use (ie when the text is too smol, but sorry for bad sync 😅) if you save passwords you can do it all w/o user interaction
1
1
290
something something personal superintelligence? now the fun part: making sure it works for you! vals.ai/models/meta_muse-spa…
(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.
4
5
1
58
15,616
i’d pitch this a step further! where w/ cross-entropy loss, not only does A(s, a) < 0 push token likelihood p(a | s) down… sample-wise, the neg-adv Xent gradient pushes down high-confidence tokens *less* than low-confidence ones (opposite of how we maybe expect Xent loss to behave in SFT) meanwhile, positive-advantage has the expected effect -> entropy collapse!
It took me a while to understand why naive PPO/GRPO tends to decrease entropy even when rewards are random. The mental model is, an overgrown action’s likelihood ratio can be pushed down with negative-advantage samples, but clipping caps the downward corrections, and these actions are more likely to be sampled. So accidental probability concentration is not perfectly reversed and entropy tends to drift downward. Popular variants DAPO/CISPO/DPPO all indirectly helps maintain a non-decreasing entropy (under random rewards), in different ways.
5
4
15
2,680
one can reason about how: 1. this collapses entropy (strongly reinforcing unlikely but ultimately correct tokens, while pushing down incorrect generations only if they’re unlikely). 2. clipping (asymmetrically) can help 3. and how this “accidentally” aligns well with LLMs (we don’t want to nuke “confidently incorrect” tokens, bc we still need to model language) never got around to pushing empirical results, but thought it was a fun observation in LLM RL (that could also comment on how RL random reward helps, wrote a bit last year in dropbox notes below) and ofc curious if you or @johnschulman2 have exploited this in ways i have not dropbox.com/scl/fi/qzqhl59lg…
3
422
congrats on launch! 🌱🌻 excited to see all the fun takes on memory / continual learning you can rag, you can compact, you can continuously rl / sft on updated data (kinda like the days of “old”) you can collect more reasoning traces, and create new envs but friends here are exploring new ideas w new memory components + new ways to train them also could not have had better day-1 labmates @EyubogluSabri @MayeeChen @NeelGuha @dan_biderman (last one maybe day-1460 but who’s counting)
2
2
24
4,992
at @neurips today and tmrw! pls do lmk if you want to talk about: - periodic / vision / ai scientists rlyfun.substack.com/p/monkey… - technicals, i.e., how to: --- will be at luma.com/p9i679us?tk=GmgDG7 with the rest of the @periodiclabs folks, some socials
if you’re interested in autonomous science at frontier scale, come find the @periodiclabs team at neurips! look for: @vwxyzjn to discuss training big MoEs @xiangfu_ml for atomic GNNs @mzhangio for AI scientists @VincentMoens for RL systems me for midtraining sample efficiency
1
2
2
29
5,735
bonus (1/2): also here to run! lmk if you want to join (will be w a fun crew of Gen Z PhDs; really want to see some nice water bois again)
1
7
575
bonus (2/2): always thought conferences would be a great time to enable in-person collabs / research jams (when else are we all together) maybe a bit tight this year, but have a big backlog of blog ideas / early methods to derisk, and a stash of tinker credits to burn - curious if ppl interested? (ty @dchaplot, @thinkymachines for all the support, maybe it uniquely enables this kind of collab? and @jerrywliu for enabling this post lol) some past examples (last one especially great❤️) 1. dropbox.com/scl/fi/gvdl0tcc2… 2. dropbox.com/scl/fi/hkvehtp2f… 3. hazyresearch.stanford.edu/bl… (tbt the goats @EyubogluSabri @simran_s_arora )
1
2
8
750
do chat w Jerry! he's been doing a lot of great foundational work on numerical precision in ML settings, think this would go a long way w improving science modality foundation models (as you may imagine, while softmax attn and bf16 are great for decoding to tokens, it kinda wrecks what the models can pick up for some modalities)
I’ll be at NeurIPS until Sunday! My work focuses on understanding the core mechanics of ML architectures, recently: (1) pushing ML toward better numerics, and (2) *constructively* understanding mechanisms for fact-storage. If you're excited about how deeper algorithmic insight can unlock new capabilities (esp. in AI+science), let’s chat. Also open to spring/summer 2026 internships!
9
1,778
v cool to see Mayee's work shape up to one of the more interesting + timely Q's how to curate pretraining data to optimize for post-trained performance? - yes you want deduplication, yes you want diverse data - but there's so much more fun stuff to do (rt @hazyresearch and the 2023 coreset notes...) - and open questions! (how should pretraining distributions change if we're targeting specific domains? conversely, or multiple RL tasks? cost-wise, should we pour more tokens into "it just works" next-token prediction if it means needing to do less RL later on?)
Thrilled to have contributed to Olmo 3! The best fully open 32B model (data, training recipes, checkpoints and more!) As an intern at AI2 these last 8 months, I’ve grown to deeply appreciate the careful science, iteration, and collaboration that go into models like this and have learned so much from the team. I am more optimistic than ever about the future of open-source and data-centric research right now. My particular contribution was working on the Dolma 3 data mix 👩‍🍳 I was able to apply ideas from some of my earlier mixing work, explore new problem settings, and see firsthand the data challenges that arise when building datasets intended for real models at scale. More on this coming soon!
1
1
18
3,429
Replying to @simran_s_arora
We were surprised to learn that the very best AMD kernels out there are hand-optimized by a handful of engineers **in raw assembly**. Unsurprisingly, this is difficult to scale to the breadth of AI workloads we need to support. It's difficult to read and modify.
1
1
10
1,589
some scary tricks hiding in those treats this year @simran_s_arora
1
3
2
45
23,474
dw the difference is “subtle”
Accidentally said "hard" instead of "non-trivial" and they kicked me out of SF
4
1,064
yess; excited to see Reflection take on American open-source i think the bossman @HazyResearch put it best below (source in reply). We owe a lot to Qwen, K2, and Deepseek today. But we also owe it to ourselves to at least have a parallel set of Western models. This keeps things competitive, and lets us continue picking models, doing research, and building progress based on quality alone. @achowdhery has been a personal inspiration; excited to see what she, @MishaLaskin, and the team share (even better if it’s a family of models from 4B to 235B MoE that researchers and startups can do post-training scaling curves on 🙏😅)
Today we're sharing the next phase of Reflection. We're building frontier open intelligence accessible to all. We've assembled an extraordinary AI team, built a frontier LLM training stack, and raised $2 billion. Why Open Intelligence Matters Technological and scientific progress is driven by values of openness and collaboration. The internet, Linux, and the protocols and standards that underpin modern computing are all open. This isn't a coincidence. Open software is what gets forked, customized, and embedded into systems worldwide. It's what universities teach, what startups build on, what enterprises deploy. Open science enables others to learn from the results, be inspired by them, interrogate them, and build upon them in order to push the frontier of human knowledge and scientific advancement. AI got to where it is today through scaling ideas (e.g. self-attention, next token prediction, reinforcement learning) that were shared and published openly. Now AI is becoming the technology layer that everything else runs on top of. The systems that accelerate scientific research, enhance education, optimize energy usage, supercharge medical diagnoses, and run supply chains will all be built on AI infrastructure. But the frontier is currently concentrated in closed labs. If this continues, a handful of entities will control the capital, compute, and talent required to build AI, creating a runaway dynamic that locks everyone else out. There's a narrow window to change this trajectory. We need to build open models so capable that they become the obvious choice for users and developers worldwide, ensuring the foundation of intelligence remains open and accessible rather than controlled by a few. What We've Built Over the last year, we've been preparing for this mission. We’ve assembled a team who have pioneered breakthroughs including PaLM, Gemini, AlphaGo, AlphaCode, AlphaProof, and contributed to ChatGPT and Character AI, among many others. We built something once thought possible only inside the world’s top labs: a large-scale LLM and reinforcement learning platform capable of training massive Mixture-of-Experts (MoEs) models at frontier scale. We saw the effectiveness of our approach first-hand when we applied it to the critical domain of autonomous coding. With this milestone unlocked, we're now bringing these methods to general agentic reasoning. We've raised significant capital and identified a scalable commercial model that aligns with our open intelligence strategy, ensuring we can continue building and releasing frontier models sustainably. We are now scaling up to build open models that bring together large-scale pretraining and advanced reinforcement learning from the ground up. Safety and Responsibility Open intelligence also changes how we think about safety. It enables the broader community to participate in safety research and discourse, rather than leaving critical decisions to a few closed labs. Transparency allows independent researchers to identify risks, develop mitigations, and hold systems accountable in ways that closed development cannot. But openness also requires confronting the challenges of capable models being widely accessible. We're investing in evaluations to assess capabilities and risks before release, security research to protect against misuse, and responsible deployment standards. We believe the answer to AI safety is not “security through obscurity” but rigorous science conducted in the open, where the global research community can contribute to solutions rather than a handful of companies making decisions behind closed doors. Join Us There is a window of opportunity today to build frontier open intelligence, but it is closing and this may be the last. If this mission resonates, join us.
5
5
65
10,884
original blog post for screenshot: hazyresearch.stanford.edu/bl… (i don’t endorse all the aesthetics but agree w the overall message above, and am thankful to have an advisor + lab that welcomes discussion and discourse :))
2
418
thinky tinky
GPUs are expensive and setting up the infrastructure to make GPUs work for you properly is complex, making experimentation on cutting-edge models challenging for researchers and ML practitioners. Providing high quality research tooling is one of the most effective ways to improve research productivity of the wider community and Tinker API is one step towards our mission there. Tinker API is built on top of our experimental results on fine-tuning with LoRA: thinkingmachines.ai/blog/lor… Beta starts and you can join the waitlist today: thinkingmachines.ai/tinker/
1
2
1,059
Michael Zhang retweeted
On your first day @periodiclabs you get to pick an element from the periodic table for your desk plate
29
13
9
474
101,223