Sam Duffield

So happy to release posteriors 𝞡 We've carefully designed the package to be easily extensible, come tell us about your favourite method or application! Open source, scalable, uncertainty quantification FTW 🥳 github.com/normal-computi…

The era of Physics-based ASICs has arrived. On the surface, computing infrastructure seems unready to meet the demands of AI. But underneath, a Cambrian explosion of new computer architectures is rising to meet this challenge, with the goal of using physical processes as a…

VerilogEval is an odd coding dataset - it gives the agent access to the golden testbench, which you will use to evaluate the final design from the model. I'm honestly surprised how papers like VerilogCoder or ChipAgents only get ~95% accuracy in this setting... That's what…

Normalizing Flows (NFs) check all the boxes for RL: exact likelihoods (imitation learning), efficient sampling (real-time control), and variational inference (Q-learning)! Yet they are overlooked over more expensive and less flexible contemporaries like diffusion models. Are NFs…

You may think it’s “I use AI” but really it’s “AI uses me”

Hmm. To me it makes sense that first-order methods are ones that use first derivatives and an analytical model of second-order structure. Second-order methods should use first *and* second-derivatives (and ideally an analytical model of third-order structure)

Always a bit funny to me how the interpretation of "second-order optimisation" is not as narrowly-defined as one might expect. My experience is that operationally, it means something like 'a matrix is involved', which is a pretty wide net to cast!

posteriors 𝞡, our open source Python library for Bayesian computation, will be presented at #ICLR2025! posteriors provides essential tools for uncertainty quantification and reliable AI training. Join @Sam_Duffield at his poster session 'Scalable Bayesian Learning with…

Come out and join us, so we can talk about physical world AI and thermodynamic chips. Looking forward to hosting with @nathanbenaich.

Really exciting progress in silicon happening now at @NormalComputing. Check out Zach's blog post to read more about what we're building. 👇

during NYC deep tech week, i gave a talk about how @NormalComputing is scaling our thermodynamic computing paradigm to tackle large scale AI and scientific computing workloads today, we're releasing a blog post based on that talk! ⬇️

So simple! Normally, we order our minibatches like a, b, c, ...., [shuffle], new_a, new_b, new_c, .... but instead, if we do a, b, c, ...., [reverse], ...., c, b, a, [shuffle], new_a, new_b, .... The RMSE of stochastic gradient descent reduces from O(h) to O(h²)…

Check out our new paper on thermodynamic optimization! 🔥 We show how K-FAC, a powerful and scalable second-order optimizer, can be accelerated using thermodynamic hardware, bringing us closer to efficient GPU-thermo co-processing! This work pushes the boundaries of…

I think that Kalman filters are one of the few things that are genuinely much easier to understand from a Bayesian perspective.

I somehow lost the markdown for this so I asked GPT-4o to recover it from the image. Not only did it do a perfect job of the latex, formatting etc it also silently corrected the error/typo in the definition of V

An equally brilliant solution comes to us from @Sam_Duffield, who deploys a kind of statistical voodoo to reach the goal