Who is using deep RL in the Age of Transformers? Obviously Deepmind are miles ahead in the research, although I’m not sure how/if they’re productising it. I kinda suspect it’s secretly widely used in quantitative finance. But who else?

Sad to hear that Jim Simons passed away today. A code breaker, co-inventor of the Chern–Simons form, former math department chair who founded Renaissance Technologies. A philanthropist who was one of UC Berkeley's & MIT's biggest supporters. May his memory be a blessing.

Hot take: ML researchers are underestimating how quickly recent scaling laws may flatten out - it’s quite likely what people see as an exponential function is a sigmoid, and the harsh reality of the physics of high energy costs and power plant construction restrict the expected…

#ChatGPT doesn't simply answer prompts. It simulates the statistically most plausible answer. Which means it lies in a very convincing way, in the direction we'd expect the answer to be if we didn't know any better: in the direction that harmonizes with our prior beliefs.

The default setting of #ChatGPT is *checks notes* vacuous aspirational corporate piffle. chat.openai.com/share/95cde8ce…

Have been working on network flows which led me down a rabbit hole tracing the history of the max flow problem. The original formulation of this problem wasn’t declassified until 1999 (!), almost 45 years after its formulation

A simple economy in four layers.

———————
1. Belief system
———————
2. Enforcement system
———————
3. Allocation system
———————
4. Transaction system
———————

A First Course in Causal Inference: these >400 page lecture notes only require basic knowledge of probability theory, statistical inference, and linear and logistic regressions. From the Yule-Simpson paradox through potential outcomes, mediation and more: arxiv.org/abs/2305.18793

Wayne Thiebaud, Ripley Ridge, 1977

Debating organization theory with #ChatGPT reads uncannily like debating with with a Twitter rando. Reports of the demise of the confused undergrad essay strike me as greatly exaggerated.

Flow chart for sales pricing decision, from Richard Cyert, James March & C.G. Moore, 'A specific price and output model', ch. 7 in A Behavioral Theory of the Firm (1963). archive.org/details/behavi…

Prof. Solow destroys exponentialist bullshittery in three sentences. The first derivative of an exponential curve is still an exponential curve, and if you try to flatten an exponential curve it's also still an exponential curve. Careers should end over this.

It's fun to watch how everyone is suddenly discovering the Boltzmann distribution...

Diffusion models can be reformulated as equilibrium thermodynamical systems.

Just been published this morning! For buying or freely downloading see the url in the comment.

Language and chess are two of the best-known domains for combinatorial reasoning, and they share the fundamental property of all combinatorial domains: their discrete building blocks differ in quality, which means there is no in-between.

Ailsa Land and Alison Doig's original paper introducing the branch-and-bound algorithm for discrete optimization problems appeared in Econometrica, but almost immediately all follow-up papers appeared in OR and MS journals. jstor.org/stable/1910129…