Sign in

Jonathan Shi

@jnshi.bsky.social
14 followers 7 following 17 posts

Optimization/sampling theory for highly non-convex problems from physics. 3 chip tapeouts incl. Intel 3. Won a COVID forecasting contest by 200,000× likelihood.

PostsRepliesMedia
Jonathan Shi @jnshi.bsky.social · 28/09/2026
We can now sample the SK model with vanishing TV error up to the conjectured β < 1! This has been a burning question for @freegaussians.bsky.social, @oldheneel.bsky.social, and me since El Alaoui, Montanari, and Sellke proposed a denoising diffusion algorithm in 2022. Embarrassingly... 1/5
130
Reposted by Jonathan Shi
Juspreet Singh Sandhu @freegaussians.bsky.social · 24/09/2026
Finally! Submitted PHA-4 to arXiv: drive.google.com/file/d/1m3gc... Sampling SK all the way to beta < 1 building on the free-probability and cavity interpolation theory developed in PHA-3, with the amazing @oldheneel.bsky.social and @jnshi.bsky.social
drive.google.com
141
Reposted by Jonathan Shi
angela zhou @angelamczhou.bsky.social · 16/09/2026
Another blog post: separatedhyperplane.substack.com/p/we-cant-le... We can't let frontier labs choose their own auditors - or their own questions. Triangulating between the near-past lens of algorithmic accountability, the financial crisis, and the AI evaluation landscape as it looks now:
separatedhyperplane.substack.com
We can’t let frontier labs choose their own auditors - or their own questions.
We should start with mandatory audits that adequately cover harms under current laws and regulations while we figure out the rest.
2249
Reposted by Jonathan Shi
Juspreet Singh Sandhu @freegaussians.bsky.social · 13/09/2026
Speaking of real math -- We will have a nice PHA installment (PHA-4) on arXiv in the next week or two to finish the SK story! Stay tuned! PS: If I was focused we could've done this sooner but I got into proving infinte-dim LSIs via Malliavin calculus. cc: @jnshi.bsky.social @oldheneel.bsky.social
021
Reposted by Jonathan Shi
angela zhou @angelamczhou.bsky.social · 09/09/2026
blog: I argue that we need auditable AI now, because current AI developments threaten auditability of AI systems for ordinary regulatory compliance, and we can't figure out “what to do next about AI” without common knowledge about AI (mis)alignment. separatedhyperplane.substack.com/p/we-need-au...
separatedhyperplane.substack.com
We need Auditable AI to figure out “what to do next about AI”
Mandatory, not voluntary, third-party governance, accountability, and transparency
1306
Jonathan Shi @jnshi.bsky.social · 14/07/2026
Why does probability theory still call it "stochastic localization" when "diffusion model" is a much more accessible name for the same process? I ask because: we proved something new about stochastic localization/diffusion models! @oldheneel.bsky.social @freegaussians.bsky.social
231
Reposted by Jonathan Shi
angela zhou @angelamczhou.bsky.social · 19/05/2026
Agencies already know they need more staff, but we show how our richer model enables *bundled interventions* that combine extra staff with service design changes. For example, DSS estimated it needed 150 staff to get waits < 2 min: if they could reduce handling time by 50%, they only need 50 more.
141
Reposted by Jonathan Shi
angela zhou @angelamczhou.bsky.social · 19/05/2026
Excited to share our paper! Due Process on Hold: A Queueing Framework for Improving Access in SNAP arxiv.org/abs/2605.15165 Millions of Americans interface with the social safety net via call centers that are too congested. In Holmes v. Knodell, bad operations = procedural due process violation.
arxiv.org
Due Process on Hold: A Queueing Framework for Improving Access in SNAP
The U.S. social safety net delivers essential services at mass scale, but access burdens persist, as congested contact or call centers serve as a primary mode of application completion and assistance....
2295
Jonathan Shi @jnshi.bsky.social · 07/05/2026
We just put out a 180-page paper on sampling from the SK model! One big surprise we ran into: the Hessian Ascent algorithms investigated for non-convex optimization have been diffusion models in disguise the whole time! arxiv.org/abs/2605.03718 @freegaussians.bsky.social @oldheneel.bsky.social
arxiv.org
Potential Hessian Ascent III: Sampling the Sherrington--Kirkpatrick Model at Beta < 1/2
We give a polynomial-time algorithm to sample from the Gibbs measure of the Sherrington--Kirkpatrick model with negligible total-variation distance (TVD) error up to inverse temperature $β< 1/2$. Prio...
031
Jonathan Shi @jnshi.bsky.social · 07/04/2025
i bet Marc Andreessen feels like an idiot now
010
Jonathan Shi @jnshi.bsky.social · 16/02/2025
one watergate scandal ended nixon, but 3½ watergates is just another day for trump
020