The way to protect our software and computer systems from attack is not to prevent people from having access to AI models that can find security holes. That is just a way of shooting the messenger. It is to find the security holes as quickly as possible and get rid of them.

Effective Altruism is a cult masquerading as a charity. Their main concern isn't bednets for people in countries with malaria, it's seizing control of artificial intelligence research and development. If you ask an AI about this, it will just spit back at you stuff it read from EA websites, so please do your own research on this instead.

There is anti-intellectualism here, but it is embedded in the words "tech bros". The "tech bros" he denigrates are responsible for CT scanning, MRI, all the synthetic techniques used for making all the drugs that get prescribed, autoclaves, surgical robots, and I could go on and on. And yes, this is really true. CT scanning, for example, was invented by an electrical engineer and a physicist, not by a practicing physician. MRI was invented by three people, one of whom (Damadian) had a medical degree but was much more of a "tech bro" and inventor; the other two had no medical degree. I remember seeing presentations 36 years ago about early surgical robots, and the people pioneering them were... "tech bros".

Moments ago, Valar Atomics took Ward 250 critical for the first time. This fulfills President Trump’s EO 14301, which called for 3 advanced reactors to go critical by July 4th. This is our second criticality as a company, and an important step toward our goal of power by July 4.

I’ve seen this trope posted by a lot of doctors in the last 24 hours. The claim from doctors is that Midjourney’s cheap full body scanner is bad, because people would get treated unnecessarily for harmless conditions, and that would increase death rates. Now, I have a lot of skepticism about Midjourney’s announcement (a lot of skepticism) but I think that the doctors are failing to look in the mirror. What they are saying is, in effect, that they cannot be trusted with having more information about patients. That, if you give doctors too much information about a patient’s condition, they may kill the patient by treating them unnecessarily. This isn’t a problem with inexpensive full body scanners. This is a problem with doctors and medical practices.

If America does not build data centers, what parts of modern life are we prepared to slow down? Medical research? Fraud prevention? Cybersecurity? Navigation? Small business tools? Emergency response? AI services? Data centers are not just buildings. They are the infrastructure behind the tools Americans already use every day.

A rice that prevents childhood blindness has been ready since the mid-2000s. It grows nowhere. Vitamin A deficiency blinds up to 500,000 children a year. Half die within twelve months. To help, scientists added beta-carotene, the pigment that turns carrots orange, to rice. It cooks and tastes the same, though the grain is yellow. They called it Golden Rice and licensed it free to any farmer earning under $10,000. Greenpeace spent twenty years fighting it country by country. Activists destroyed test fields. Lawyers got approvals revoked. A hundred Nobel laureates asked them to stop. They did not stop. I spent weeks calculating what the delay cost. My estimate is 110,000 dead children. Roughly 15 a day, every day, for two decades. The same technology did reach American grocery stores. It makes a pink pineapple that sells for up to $50. The pineapple has a corporation rolling out a new fruit, with the profit margins to take on critics and market the science. Golden Rice had no profit margin to defend itself.

I don’t like raining on the parade, but I see no pictures of actual hardware here, no serious technical explanation of how the reconstruction occurs, and no actual scans. I’m not going to say that it isn’t real, but I am going to say that I don’t yet have enough evidence to know that it is real.

Over the last week we’ve audited all our GPU code written by Opus 4.8 and have found a lot of perf goofs We have no way of knowing whether what we found was honest goof or intentional nerf, either way you should audit all your performance critical Opus 4.8 code with Codex 5.5

“To conclude, modern diff-in-diff methods yield the same old answer: minimum wages reduce jobs. The opposite finding in CDLZ and DL is the consequence of discretionary, narrow, and unsound researcher choices that favor that result.” In even less kind words: if you torture the data long enough, you can extract a false confession.

The US government already spends around seven trillion dollars every year, vastly more than the value of Elon Musk's companies, vastly more than he has created in his entire lifetime. Why don't we live already in paradise? At what level of taxation will Utopia suddenly appear?

If you’re not a computer scientist, skip this post. The biggest impediment to formal verification of systems code is currently this: we have at least three extant formal semantics for C, no viable formal semantics for Rust, and no other viable systems languages. (Yes, I am ignoring C++.) If we want to formally verify systems code at all, never mind at scale, we need either a new systems language with a formal semantics, or we need a formal semantics for Rust and a reasonable feeling that the semantics reflects the actual implementation of the language and can be maintained going forward as the language evolves. Unfortunately, as good as Rust is, and it is very good by historical standards, it is a complicated language, and there is a strong suspicion that the type system is not actually sound. It is also probably too large a language for which to build a formally verified compiler given the current state of the art (though that might change), and it is not clear that the official compiler would not rapidly evolve away from a version of the language that a formally verified compiler implemented. Even given a formally verified compiler and a semantics, neither of which exist, the current proof assistant landscape, tools like Rocq and Lean, are hardly ideal for this sort of work, even given AI assistance. So, we are in a position in which we know that there is a substantial gap between what we need in order to move forward to formally verified systems software, and the tooling that we have currently. This is a hard problem. It is also one that needs a solution, or more likely, a set of solutions. (One irony of the current situation is that most experiments in formally verified systems software still reuse one of CompCert’s intermediate representations as a target. And thus, C is still being perpetuated, even in high assurance programming work.)

GPT-5.4 helped drive a medicinal chemistry project from literature review to a validated experimental result. Paired with Maria AI and specialized lab, the model proposed an unexpected way to improve a widely used reaction in drug discovery.

If GitHub gives up its preeminent place, I would prefer to see a federated system where one identity was shared across a distributed forge network rather than having to have logins on 50 different forges, having to check notifications on all of them, having no integration between them, etc.

For those of you who claim that this is all a bubble: does that mean you are shorting SpaceX, Nvidia, etc.? And if not, doesn’t that contradict your claim? I sometimes hear people say in response that it’s too hard to short things you only expect to drop a couple of years hence, but you can do it with long term derivatives contracts. So why aren’t you all doing it? My guess is it’s because talk is cheap.

What I would like is to formally verify that code works exactly as expected. “Formal verification” means mathematically proving the correctness of software. This will require AI automation because it’s usually too labor intensive for human developers.