David Pfau
@davidpfau.com
So far I have not found the science, but the numbers keep on circling me.
Views my own, unfortunately.
Views my own, unfortunately.
The original reply guys.
November 2, 2025 at 8:19 PM
The original reply guys.
Meanwhile Commonwealth Fusion Systems is building SPARC, which will be the highest magnetic field tokamak in the world when it comes online in a few years. It is also likely to be the first magnetic confinement device to generate more energy than it consumes. (Image: Hsu and Wurzel 2025)
October 16, 2025 at 1:58 PM
Meanwhile Commonwealth Fusion Systems is building SPARC, which will be the highest magnetic field tokamak in the world when it comes online in a few years. It is also likely to be the first magnetic confinement device to generate more energy than it consumes. (Image: Hsu and Wurzel 2025)
Trump's approval rating has been fairly steady, but his favorables have been falling off a cliff. Not really sure what to make of this - people are just happy their 401(k) is going up but increasingly hate Trump himself? (Source: Real Clear Polling)
September 16, 2025 at 10:55 PM
Trump's approval rating has been fairly steady, but his favorables have been falling off a cliff. Not really sure what to make of this - people are just happy their 401(k) is going up but increasingly hate Trump himself? (Source: Real Clear Polling)
Can't wait for DARPA to rebrand as WARPA, relocate to London and start releasing critically-acclaimed electronic music.
September 6, 2025 at 3:10 PM
Can't wait for DARPA to rebrand as WARPA, relocate to London and start releasing critically-acclaimed electronic music.
I'll be at ICML this week, presenting our paper on Wasserstein Policy Optimization on Tuesday! If you're in Vancouver, come say hi!
July 14, 2025 at 8:16 AM
I'll be at ICML this week, presenting our paper on Wasserstein Policy Optimization on Tuesday! If you're in Vancouver, come say hi!
I'm afraid it's not working at the moment, this site is shrinking anyway. And it's shrinking in large part because people find it too much of a monoculture.
June 5, 2025 at 8:27 AM
I'm afraid it's not working at the moment, this site is shrinking anyway. And it's shrinking in large part because people find it too much of a monoculture.
I don't often beat up on the NYTimes, because I think the alternatives are mostly worse, but every once in a while they are so egregious my jaw literally drops.
Left: NYTimes. Right: The Telegraph (a paper so right-wing it's often called "the Torygraph")
Left: NYTimes. Right: The Telegraph (a paper so right-wing it's often called "the Torygraph")
June 1, 2025 at 11:49 AM
I don't often beat up on the NYTimes, because I think the alternatives are mostly worse, but every once in a while they are so egregious my jaw literally drops.
Left: NYTimes. Right: The Telegraph (a paper so right-wing it's often called "the Torygraph")
Left: NYTimes. Right: The Telegraph (a paper so right-wing it's often called "the Torygraph")
The war on science in the US is already having an effect on private sector research like AlphaFold. Bears repeating but the private sector builds on top of things created by academic research for the public good. This hurts everyone.
May 28, 2025 at 10:18 AM
The war on science in the US is already having an effect on private sector research like AlphaFold. Bears repeating but the private sector builds on top of things created by academic research for the public good. This hurts everyone.
I'm enjoying Andor season 2 but every couple of scenes I see something that looks obviously AI-generated and I feel like I'm going insane.
May 25, 2025 at 11:56 PM
I'm enjoying Andor season 2 but every couple of scenes I see something that looks obviously AI-generated and I feel like I'm going insane.
In some sense, there’s nothing in this paper that we couldn’t have done in 2018 (and I wish we had! I’d be famous!) But the inspiration for this paper actually came from the fantastic recent work on Wasserstein QMC by @k-neklyudov.bsky.social and others. Good research should be timeless.
May 2, 2025 at 8:12 AM
In some sense, there’s nothing in this paper that we couldn’t have done in 2018 (and I wish we had! I’d be famous!) But the inspiration for this paper actually came from the fantastic recent work on Wasserstein QMC by @k-neklyudov.bsky.social and others. Good research should be timeless.
Tellingly, we have some preliminary evidence that, as the task dimension increases, WPO gets *better* relative to other methods. But it’s surprisingly hard to find truly high dimensional (>100) RL benchmarks. Something for the community to work on!
May 2, 2025 at 8:11 AM
Tellingly, we have some preliminary evidence that, as the task dimension increases, WPO gets *better* relative to other methods. But it’s surprisingly hard to find truly high dimensional (>100) RL benchmarks. Something for the community to work on!
This update can be used as a drop-in replacement in almost any actor-critic method in continuous action spaces, and we find that a KL-regularized version of WPO is competitive with other methods on dm_control and a simple tokamak magnetic control task.
May 2, 2025 at 8:11 AM
This update can be used as a drop-in replacement in almost any actor-critic method in continuous action spaces, and we find that a KL-regularized version of WPO is competitive with other methods on dm_control and a simple tokamak magnetic control task.
Previous work using Wasserstein gradient flows for continuous control arrived at methods similar to SAC which use the reparameterization trick, but we found that there is a simple and general closed form update that works for *any* stochastic policy.
May 2, 2025 at 8:10 AM
Previous work using Wasserstein gradient flows for continuous control arrived at methods similar to SAC which use the reparameterization trick, but we found that there is a simple and general closed form update that works for *any* stochastic policy.
Most RL algorithms for continuous actions fall into two classes. PPO, MPO, etc... work for arbitrary policies but only use scalar value information. DDPG, SAC, TD3, etc… use gradients of the value but are limited to policies that can be represented by reparameterization.
May 2, 2025 at 8:10 AM
Most RL algorithms for continuous actions fall into two classes. PPO, MPO, etc... work for arbitrary policies but only use scalar value information. DDPG, SAC, TD3, etc… use gradients of the value but are limited to policies that can be represented by reparameterization.
New paper accepted to ICML! We present a novel policy optimization algorithm for continuous control with a simple closed form which generalizes DDPG, SAC etc. to generic stochastic policies: Wasserstein Policy Optimization (WPO).
May 2, 2025 at 8:09 AM
New paper accepted to ICML! We present a novel policy optimization algorithm for continuous control with a simple closed form which generalizes DDPG, SAC etc. to generic stochastic policies: Wasserstein Policy Optimization (WPO).
Spotted at a pancake restaurant in London. Sign of the times.
April 21, 2025 at 12:13 PM
Spotted at a pancake restaurant in London. Sign of the times.
I was basically right there! We were on Lake Salem in Derby.
April 8, 2025 at 6:38 PM
I was basically right there! We were on Lake Salem in Derby.
$NVDA below $100 today. It was about $140 when this photo was taken the day after the inauguration. I think in a generation we'll look back on this as the peak of the generative AI bubble.
April 4, 2025 at 2:59 PM
$NVDA below $100 today. It was about $140 when this photo was taken the day after the inauguration. I think in a generation we'll look back on this as the peak of the generative AI bubble.
Ok this is getting out of hand...
March 19, 2025 at 10:04 PM
Ok this is getting out of hand...
It was masked somewhat by increased immigration from other countries due to the new points-based system, but overall European migration is net negative. People are leaving.
February 26, 2025 at 12:37 PM
It was masked somewhat by increased immigration from other countries due to the new points-based system, but overall European migration is net negative. People are leaving.
I'm not sure if we can survive another 47 months of this, but at least the honeymoon appears to be over.
February 20, 2025 at 11:35 PM
I'm not sure if we can survive another 47 months of this, but at least the honeymoon appears to be over.
Imagine telling someone in the '90s that "President Trump might shut down The X-Files" is a real actual thing in the year 2025.
February 3, 2025 at 7:42 PM
Imagine telling someone in the '90s that "President Trump might shut down The X-Files" is a real actual thing in the year 2025.