Sam Duffield

New open source: cuthbert 🐛 State space models with all the hotness: (temporally) parallelisable, JAX, Kalman, SMC

@andreamichi @depthfirstlabs Open source ❤️ Safe open source ❤️❤️

Today we're launching the Open Defense Initiative: up to $5 million in @depthfirstlabs credits for critical open source projects to find and fix real, exploitable vulnerabilities. We’ve been investing heavily in the entire agentic security technology stack: a harness that allows models to efficiently reason through code the way a security engineer would, the context understanding to validate exploitability and produce ready-to-merge fixes, and recently, post-trained models optimized for vulnerability discovery. Today we shared a datapoint that highlights the value of that approach. depthfirst autonomously identified 12 vulnerabilities in @FFmpeg , after Anthropic did several hundred runs with Mythos, for 1/10th of the compute cost reported by Anthropic. As a former @GoogleDeepMind researcher, I am grateful to foundation labs for pushing the boundaries of what AI can do, as every advance there gives us a stronger base to build on. But running general purpose models without the proper harness can be too expensive to be economically viable. That's what makes today's launch possible. The efficiency we've built lets us put frontier-level security in the hands of the maintainers who need it most, before open source models catch up and bad actors wield these capabilities without filters or guardrails. If you maintain a critical open source project, apply for Open Defense credits through the form in the comments.

@Sam_Duffield @HoldijkLars @iclr_conf @mmbronstein I am still thinking about that key equation too!

I think we’ll all agree I dressed appropriately too 🇧🇷⚽️

It was super fun to present my SDE decomposition at the @iclr_conf DeLTa workshop And thanks for stopping by @wellingmax @mmbronstein, really enjoyed chatting and lots to think about!

At @iclr_conf in Rio? Join us for the Normal Silicon Social on Sunday, April 26 from 6:30 to 8:30 PM at the Copacabana Palace. Request an invite: luma.com/3ioi64wv #ICLR2026

Did @wellingmax’s keynote at #iclr2026 leave you wanting more stochastic thermodynamics and generative AI? Come join Max and me at poster #1023 in Pavilion 3 in the afternoon to chat about how we can learn escorted protocols for multistate free-energy estimation!

Normal Computing is sponsoring @iclr_conf 2026 in Rio de Janeiro. Our research team will be on-site presenting papers spanning thermodynamic computing, Normal's physics-based computing architecture, and the AI methods underpinning Normal EDA, our platform for silicon engineering. Come talk to the researchers behind the work. We'll share a preview of what we're launching live at the conference. Follow along for real-time updates in the ICLR app. Meet us at the booth: Thursday, April 23 – Saturday, April 25 9:30 AM – 5:30 PM daily

I'll be at ICLR this week 🇧🇷 presenting my complete SDE decomposition at the DeLTA workshop on Monday. Reach out if you want to meet!

cuthbert is turning into a really awesome tool that's made our experience in writing dynestyx (github.com/BasisResearch/…) much more pleasant! Super happy to be seeing symbiosis between the two!

There's a load of fun examples in the cuthbert docs and we're always looking to add more! 🐛 state-space-models.github.io/cuthbert/examp…

In this highly nonlinear example it's more accurate and faster than both extended and particle filters. state-space-models.github.io/cuthbert/examp…

The ensemble Kalman filter has been added to cuthbert by @DanWaxmanML @BasisOrg 🔥 The EnKF is one of those algorithms that "just works" - oftentimes in settings it has no right to github.com/state-space-mo…

Love this, let's have the LLMs protect our code for us

depthfirst has raised an $80M Series B at a $580M valuation. Attackers are using AI to break into systems faster than ever before. depthfirst is on a mission to stop this. RT + Comment “depthfirst” and I’ll send you a FREE vibe coding security agent.

Parallelizing nonlinear RNNs is gaining traction! More efficient than transformers; more expressive than linear RNNs. My PhD thesis provides an intro guide to the math (Newton's method) behind the parallelization. Great as a quick-start if you want to explore this new field!

Blog Post - Optimizers and ODEs A continuous-time view of gradient-based optimization: starting from the observation that integrator choice matters in physics simulation, and transferring that insight to understand modern optimizers.

@cthorrez Oh this totally sounds like a challenge 🤩 would love to see if we can beat it with cuthbert 🐛 Input information = timestamps + {win/lose/draw} ? Performance metric = result prediction accuracy/NLL ?

even though it's old as hell it hasn't been beaten for general-purpose use by any other system which uses the same input information

Glicko 2 is still the GOAT for online rating systems

@octonion General bivariate Poisson family with some free parameters which you then fit to data doi.org/10.1111/1467-9…

For a baseline soccer model in a top club league, which model do you prefer? Poisson Dixon-Coles Bivariate Poisson with covariance Diagonal-inflated Poisson Conway-Maxwell Other?

Struggling with minibatch noise in Stochastic Gradient Bayesian Inference? Want your chains to naturally run on stochastic hardware? Introducing — Stochastic Gradient Lattice Random Walk! New work from @NormalComputing in collab with @zierhjmensch, @adnarim066, @wellingmax, and the dream team at normal computing, @Sam_Duffield, @MaxAifer and @ColesThermoAI.