nwah1

Simplifying and composing these proofs with other proofs should probably be the next goal. And formalizing within other computational proof paradigms.

So, a very inefficient database with massive footguns.

Then all software moats become shallow

There was some talk on here during Covid about the feasibility of making all structures up to BSL4 standards. Seems kind of nuts, but it would certainly contain a lot of biosafety threats. With AI it now seems borderline urgent.

This narrative is largely untrue. Scandinavia started to prosper only after they liberalized. Of course, there is a big difference between intelligent redistributionism which respects and leverages market forces, and the kind of socialists who implement rent control and government-run grocery stores.

Redistribution that enhances production rather than stifling it. Which provides justice rather than creating injustice. A great idea to bury under dust for centuries.

I'm sure we will. And, like with mutations, the default assumption for a random change isn't very good. In this case, the historical record points to worse-than-random outcomes.

They don't use ternary quantization. But, they could, if they wanted.

In simpler terms, the SRAM density improved by 2.9% (according to ChatGPT). Which is better than it sounds, because SRAM had reached a wall, in terms of density gains, and, chips can be dominated by SRAM in terms of surface area, so this can mean a lot more room for logic, even though that is also getting considerably denser.

Given that they paused signups for their 20x Max plan, your assessment seems overoptimistic.

Their new Broadcom chips seem cool, and they had a nice model release with Astra. But, that doesn't mean they've solved the profitability question during an era of rapidfire open-weight model releases, extreme memory shortages, and intense political pushback.

A practical solution for the here and now.

Seems like the next evolution for memory-safety will be in the hardware, with CHERI.

Language-level memory-safety is all fine and good, if you can afford the rewrite. But, you still have to trust the software supply chain (the compiler, build server, etc).

However, there's a third and broader category called Proof Carrying Code, which could be used to guarantee all kinds of security properties of binaries, including memory-safety in a way that requires no supply chain trust.

With both CHERI And PCC, your machine will simply never perform unsafe memory access, and it doesn't matter what language the binary was authored in.

Seems like language-level and PCC safety guarantees are the kind of thing that AI agents will make much more practical. With AI hacking, this only gets more important. They've already started trying to rewrite everything in Rust, but PCC and CHERI go further.

Redesigning the entire OS stack around zero virtual memory, and purely static allocation, seems like a direction that will need to be done. CHERI makes that much easier.