I agree. I think it may be possible that we build AI some day that does, but the current local minimum (attention and transformers) is not it - and I don't see that trajectory changing any time soon.
These days I sprinkle less salt on Stallman's _software_ ideas, there are many things I thought he was being an idiot about, which very recently became true. But maybe it could Nostradamus-isms: predict enough an eventually something will stick.
If you want to imagine a "1.0 quality" which means "Is used in some software that isn't a toy" then all kinda of crap counts. Bill has given specific goals, I don't think he'll meet them or perhaps even understands how high those bars are†, but even by his understanding Odin hasn't reached those goals.
Unlike Jai you can just download Odin and see for yourself, it has the particular things Bill prioritized (swizzling, a very particular way to do generic programming) and it doesn't have things which Bill feels are a mistake (most obviously package management, but also closures, first class user-defined types, macros, I could go on). The resulting perf isn't very good, and to me it "feels" clumsy to use.
One of the striking things in the Handmade languages is that they're so often wedded to LLVM and so in that respect they're much worse than C which of course isn't even wedded to modern architectural choices like 8-bit bytes, much less LLVM. Zig is the most free of this peculiar curse, which is ironic because years ago Bill called out Zig as unable to escape this, while insisting Odin would not require LLVM - the reverse of what actually transpired.
† In particular Bill thinks he's going to completely specify the language. Anyone who works on this problem for WG14 (C), WG21 (C++) or Rust knows that's basically a rabbit hole made entirely of more rabbit holes. I think Oracle's Java has a complete specification, and maybe TC39 has one for "Javascript" neither of those were uh, cheap or easy.
Zig too. But this is far from being mainstream. Rust took around 5 years I think to become accepted in major companies, and those languages are harder to justify.
I've started distrusting anything good people have to say about Zig because of the wide variety of untrue claims made about it - unless those claims come from Andrew himself. Last time I was assured that compile-time memory guarantees were possible and were going to happen (they didn't: the recent announcement is runtime/debug assertions). Zig is still making massive breaking changes, and while that is not a bad thing, it makes it categorically _not_ production ready.
Also note that isn't much different from debug heap that MSCV was already having in 2000[0], assuming you didn't want to shell out some money to Insure++{1], BoundsChecker[2] and similar products.
Apps can have vulnerabilities (intentional or not). It's a contrived example, but maybe some weather app has a custom icon feature, which means you can then browse photos by opening it on the lock screen and going to that setting.
Widgets are just apps with limited scope and surface area. There’s no reason they can’t apply this to the weather app which already has (or should have) a limited scope.
> It also means that the "hey Claude vibeslop me a CUDA Anubis solver" route is on its way to being fundamentally dead.
Would there be any risk in making a dynamically built algorithm (i.e. the risk that we don't find N zero prefix bits)? Chain a set of transforms and build a unique WASM binary (including optimizations to deter pattern matching); possibly one a day or hour or something. That way you'd need to have Claude vibeslop the CUDA implementation at that cadence, per website.
The paper was posted on the 18th of June. I would guess that means they started writing the first draft in the winter of 2026, if not earlier. 2026 data obviously wouldn't be available then, or even now---we're only in September!
In fact, most data products take a few months to compile and distribute. The US will release its Q2 estimate of GDP at the end of September, for example; Statistics Canada is on a pretty similar schedule. Thus, 2025 data might not have been available when they started writing either. It's certainly possible to estimate things on your own, but now you need to show that a) your method of estimating GDP (energy prices, whatever) is valid and b) any effects you see aren't simply due to using different metrics.
In short, it just takes a while to do and write up research.
