Investment in AI infrastructure now dwarfs historic spending on railroads and the Internet combined, yet investment in our civic and scientific institutions hasn't kept pace.
In @ScienceMagazine, @profjamesevans, visiting faculty in our Paradigms of Intelligence team, and @glenweyl explore how to close that gap: science.org/doi/10.1126/scie…
In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included).
Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs.
Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines.
This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!!
Enjoy: science.caseyhandmer.com/phy…
I'm becoming increasingly skeptical of the entire "AI for Science" narrative.
These days, almost anything can be called AI for Science.
Train a neural network on some scientific data, beat a baseline, improve prediction accuracy by 2%, and suddenly you're accelerating scientific discovery.
But what exactly have we discovered?
A better predictor isn't necessarily a better understanding of nature.
A model that predicts molecular properties more accurately hasn't necessarily taught us anything new about chemistry.
And a model that fits experimental data better hasn't necessarily uncovered the mechanism behind it.
I think we're confusing computational progress with scientific progress.
What's even more frustrating is that much of AI for Science is still evaluated like ordinary machine learning research.
New architecture. Better benchmark. Higher accuracy. Another paper.
But science isn't supposed to be a leaderboard.
The real question should be whether AI helps us formulate new hypotheses, discover previously unknown mechanisms, or design experiments that reveal something genuinely new about the world.
If it does, that's extraordinary.
But if all we're doing is applying increasingly sophisticated models to existing datasets and calling the resulting performance improvements scientific breakthroughs, I'm not particularly impressed.
AI for Science should be judged by the science it produces, not by the AI it uses.
Otherwise, we're just giving machine learning another fashionable name.
Ai was told to minimize the number of delayed flights per day. Ground stop all of them indefinitely is the only way to maximize this score. Zero delays from now onward. And thousands of non delayed flights today. Another success of optimization. Hurray!
If the Platonic Representation Hypothesis is true, brains and ANNs should converge in representation. In the 2010s, we found evidence of this in the visual cortex. Alignment has stalled since then. It's a huge opportunity. I break this down in this post
From a startup founder: “it’s now a red flag for me if I don’t see side projects on someone’s resume / application, because it signals lack of autonomy. Because it’s now so trivial to build anything on the side, vs before.”
How do we understand neural nets with so many interacting params that we can’t describe what they do in words?
Our take: objective functions and regularizers offer a concise description of "what models learn"
Can we *empirically* study implicit regularization in DL?🧵
Everyone in AI should hold put options for AI. I sure do. AI to the moon? You gain because your future incomes are a bet on AI. AI loses? Your portfolio gains.
Absolute slaughter.
Double digit losses for many AI companies. SpaceX 10% below IPO price, CoreWeave about to drop below its IPO price. Nvidia is down almost 10%
The AI industry’s day of reckoning may have finally come.
In canada the applicants to grants are allowed to use AI and the reviewers are not. That seems wrong. (chatgpt, which prompt did this author use? :) ) science.gc.ca/site/science/e…
Computer vision and connectomics jobs at mindspan just dropped. mindspan.org/jobs Build deep learning models for the one of the world's largest segmentation jobs. Build great and robust pipelines. Annotate or proofread. Mindspan is hiring.
How many times have you seen a mouse study claiming that a therapeutic works, and the survival curve looks like this? Broken axis, 6% difference in survival time, but the P value is < 0.05 and it's claimed as a positive result. It's not. In my 15 years of prion science, I have learned to look past such claims almost instantly. Yet sometimes these claims still lead to high-profile publications, millions of dollars in funding, and even to therapies being tested clinically. In this new blog post I release data on 301 untreated control animals from our last 9 years of in vivo studies, model out the various ways that P values can appear significant even when the result is null, and put into words the intuition I've built up as to which claims are believable and which are not.
cureffi.org/2026/11/03/how-t…