rs intern @ meta, phd @Mila_Quebec, prev. @UWaterloo i like generative models and toronto sports teams (not necessarily in that order)
Montréal, Canada
Joined November 2016
- Tweets217
- Following774
- Followers280
- Likes772
Avery Ryoo retweeted
1/4) We're hiring in my lab, **please share**! 🚀
There are two postdoc opportunities in cog neuro and AI.
This is for a new project supported by a Schmidt Polymath Award. This project aims to build an engaging video game to collect data for neurofoundation models. 🧠🤖
Couldn't be happier to announce my new paper on defining compositional structure — and generating it! I see this project as the culmination of my PhD, as it unifies all my interests: compositionality, compression, structure, and open-endedness. More below! arxiv.org/abs/2610.20922
Before Thomas moved onto better things, he was first my roomie in 2024/25 – a period where a) I snorted when he first showed me Claude Code and b) a beer or two (or more) were shared. A million things have changed since, but it makes me smile that these two interests have not :)
New blog post - Chasing the Karpathy Peak (and Falling Off Slop Cliff): Confessions from Slopaholics Anonymous
My story of becoming a slopaholic, who’s most at risk, and how to avoid it (or recover)
thomasjiralerspong.substack.…
Avery Ryoo retweeted
Most papers treat the privileged information in OPSD as a fixed parameter and ablate other things around it. When OPSD fails they call the algorithm doomed or a waste of time, without actually touching the privileged information once.
It was pretty clear form our initial work on this that the quality of privileged information is the most important factor to make OPSD work (arxiv.org/pdf/2602.04942).
@zkjzou did the most comprehensive analysis on this figuring out what acc makes OPSD tick and why it may not be doomed.
This is the first work that convinced me OPSD when done right can actually consistently outperform GRPO.
Not All Privileged Information Is Equal: What Makes OPSD Actually Work?
We find that information-rich PI, such as environment observations and mistake analysis, outperforms PI that only provides ground-truth solutions.
Better PI leads to better post-training performance. But can we predict which PI will work best before spending compute on post-training?
We investigated three offline metrics:
🔹 Utility: A model's performance with PI can be misleading due to information leakage.
🔹 Reverse KL: Larger PI-induced distribution shifts often correlate with better post-training performance, but incorrect PI can cause equally large shifts.
🔹 Error-step KL contrast: Measuring whether PI changes the model's behavior specifically at error steps better captures its effectiveness and helps select stronger PI.
What types of privileged information have you found most effective in your OPSD experiments?
📄 Paper: arxiv.org/abs/2610.22512
Avery Ryoo retweeted
Your neo-lab wants to pre-train a frontier model, but no provider will sell you enough co-located chips?
We got you.
Our new work shows that SOAP & SPlus achieve strong worker scaling to 96+ DiLoCo workers (e.g., 96 providers) by whitening the update’s covariance. 🧵1/9
Avery Ryoo retweeted
Mila's annual supervision request process is now open to receive MSc and PhD applications for Fall 2027 admission! For more information, visit mila.quebec/en/prospective-s…
Avery Ryoo retweeted
Today, we introduce Priming Models, a new paradigm for training one-step generative models from scratch.
No diffusion teacher. No flow map. No CFG. No JVP.
We believe this work will challenge how we think about building one-step models.
Joint work with @sitanch, @du_yilun, @parkwbin03, Bohyung Han.
🔥 1.19 FID one-step on ImageNet 256×256 while using substantially less compute, e.g., 12X less than Drifting-B, 6X less than Drifting-L
🎨 Scales to text-to-image generation (GenEval 0.82)
🤖 World modeling on RT-1 / Bridge-V2
Priming Model is built on one core message: [Reconstruction pretraining --> Distributional post-training].
Avery Ryoo retweeted
Really excited to see this progress! We’ve been exploring some of these ideas in our work on neuro-foundation models. Glad to see these models enabling live BCI use.
This new post from @neuralink combines POYO-style spike tokenisation with a Mamba backbone, akin to our POSSM architecture, and explores self-supervised learning for these models, as we did recently with MOJO. Exciting to see these ideas supporting real-time BCI use! 🧵
Avery Ryoo retweeted
This new post from @neuralink combines POYO-style spike tokenisation with a Mamba backbone, akin to our POSSM architecture, and explores self-supervised learning for these models, as we did recently with MOJO. Exciting to see these ideas supporting real-time BCI use! 🧵
Our clinical trial participants have now used their Neuralink implants for over 50,000 hours, building one of the largest datasets of intracortical neural activity ever.
We used it to pretrain a foundation model for BCI: self-supervised neural encoders that turn noisy, drifting spikes into stable representations.
Some initial results:
• Decoders that keep working for weeks instead of days. For some users, calibration dropped from 10 min/day to 10 min/week
• Teleporting cursors that jump straight to the intended target
• A new cursor-control record of 11.32 bits per second (exceeding the performance of most able-bodied people)
We’re just getting started. Future directions include pooling data across participants, one-shot (or zero-shot) calibration, and an API for the motor cortex.
Read more:
neuralink.com/updates/pretra…
what an honour to work with such smart individuals, please come check us out at neurips!!
MOJO has been accepted to #NeurIPS2026! 🎉
We use SSL + unlabelled data for better decoding, stronger few-shot transfer, and more interpretable embeddings with POYO and POSSM (SoTA decoders for neural spikes).
Stay tuned for more results + torch_brain-based code. 👀
... and breathe 💨
favourite paper title in recent memory
Can self-interested, self-improving, self-replicating agents learn to cooperate?
Our new paper, Tapes Together Strong, shows they can: when social behavior, computation, and reproduction share one energy budget, cooperation evolves from scratch.
arxiv.org/abs/2609.10817 🧵
It makes me quite sad to see how the OSC has been treated in recent years. As a kid from North York, I've spent dozens of happy days with my family+friends running around the various exhibits and attending the odd show. Disheartening the next gen. might be deprived of the same.
#BREAKING: Ontario Science Centre laying off entire science department cp24.com/politics/queens-par…
Avery Ryoo retweeted
i hate to give the NBA grifters the attention they’re so obviously craving, but the fact that there are currently zero trans women in the WNBA makes it clear that this has never been about “protecting” anyone. it’s about exploiting the league’s growth to advance their anti-trans agenda. and it’s pathetic
Avery Ryoo retweeted
Does a voice agent's quality come from the orchestration framework or the models underneath?
We built 𝐕𝐀𝐦𝐨𝐒 𝐁𝐞𝐧𝐜𝐡 to measure it.📏
Holding the models fixed and swapping the harness moved completion 1.7 pts. Swapping the model stack moved it up to 28. 🧵🧵
Excited to announce the 2nd Foundation Models for the Brain and Body Workshop at #NeurIPS2026 in Sydney, Australia! 🧠🤖
We invite short papers & demos on AI for neural, physiological, or behavioral data.
📅 Papers: Sep 4
💡 Demos: Sep 19
Learn more 👉 brainbodyfm-workshop.github.…
Avery Ryoo retweeted
When talking to LLMs, It’s not What You Say, but How You Say It!
Check out our new ACL paper! 🌀
Does the *way* we say something to an LLM affect whether the LLM agrees with it?
Our ACL 2026 paper - “It’s Not What You Say, It’s How You Say It”, answers with a resounding yes! ✅
aclanthology.org/2026.acl-lo…
Co-led with @ClaraKuempel, as well as @MichelleWastl and @a_stadt 🧵