hansvm

30 ms·

451.
by hansvm 1y ago
That's what I assumed. I know languages have to handle cancellations in async code, but Rust has had a fair amount of drama over the years, and I assumed the title was accurate and reflected that some drama was happening.
452.
by hansvm 1y ago
The latter thing. One user might use features 1,2,3,4,5, another 1,3,5,7,9, another 2,4,6,8, and a fourth 1,2,3,8,9. Those usages weren't meaningfully disentangleable.
453.
by hansvm 1y ago
At one SaaS I worked at, I spent a day or two on a foolhardy initiative to analyze our users based on the subset of features they actively used (hoping to reign in the complexity of the product to a few core archetypes we could design for,
454.
by hansvm 1y ago
Even if you take Trump, Musk, etc at their words -- that the government is horribly bloated and that DOGE has saved $206B -- this still doesn't look in totality like a sound economic decision: 1. The deficit has increased with Trump in
455.
by hansvm 1y ago
"Allocated resources effectively" is doing a lot of heavy lifting in your description. The current system also rewards large-scale cons and rent extraction with the political power to do more of the same.
456.
by hansvm 1y ago
> must reserve With the current implementation exposed to the end customer, yes, that's required. Reserving specific seats isn't fundamental to the constraint that some people want to sit together. Plus, the current reservation
457.
by hansvm 1y ago
Relooper goes the other direction though. We don't need to produce high-level control flow. For a transpilation-to-Zig use case the algorithm can be extremely dumb. Naively translate every loop (or any other control flow containing non
458.
by hansvm 1y ago
Thanks :) My example is minimal, which is a little nice since I wind up re-deriving it in a hurry every time I need it. I do like the 1:2 to 2:1 symmetry though. Very elegant.
459.
by hansvm 1y ago
Yeah, it's fairly new. Maybe 0.14.1? Those cases are definitely supported. Here's a reasonably faithful re-interpretation of Duff's device which generates the right assembly. If you'll take my word for it, code and assem
460.
by hansvm 1y ago
> disagree with generalizations that you can judge the quality of code based on whether or not it uses a lot of Arc That wasn't really my point, but I disagree with your disagreement anyway ;) Yes, you don't want to over-genera
461.
by hansvm 1y ago
The constraint algorithm looks something like: 1. Choose the first token. If well-trained you have a 75% chance of choosing "a" and a 25% chance of choosing "b". Both are valid for that grammar. 2. Choose the second toke
462.
by hansvm 1y ago
> Regardless, code that doesn't use unsafe can't have two threads concurrently writing the same memory. It's a bit messier than that. Basically the only concurrency-related bug I ever actually want help with from the compi
463.
by hansvm 1y ago
That's fair. It's not really a good pattern though. You get all the runtime overhead of object-soup allocation patterns, syntactic noise making it harder to read than even a primitive GC language (including one using ARC by defau
464.
by hansvm 1y ago
That's definitely possible. As you know, (most current) LLMs build text autoregressively. This allows them to generate text with _exactly_ the same distribution as the training data. When you constrain LLM output at each token, that gi
465.
by hansvm 1y ago
Labeled switches come pretty close. Is there a particular case you have in mind where they don't suffice?
466.
by hansvm 1y ago
That's wild to me. I'm mostly not a fan of browser-based tools, and I was apprehensive of Discord calling things "servers" when they're clearly not (if they lie about that then what else?), but it's been rock-s
467.
by hansvm 1y ago
That's an interesting example. On the one hand, we really don't want actively suicidal pilots. On the other, are the current guidelines too stringent? I can't imagine you're saying that actively suicidal pilots should be
468.
by hansvm 1y ago
Even there it's risky. LLMs are good at subtly misstating the problem, so it's relatively easy to make them prove things which look like the thing you wanted but which are mostly unrelated.
469.
by hansvm 1y ago
I don't know which 1200 applications they saw, but IME they're a lot better at trying to con you than succeeding. LLMs aren't great for a lot of use cases (yet?), and this is one of those areas where reality doesn't matc
470.
by hansvm 1y ago
I must be missing something. 1200 real applications are hard to sort through. 1200 mostly fake applications are much easier. Hiring is a high-leverage activity, and it's absolutely worth spending a couple hours going through those by h
471.
by hansvm 1y ago
That's quite a neat trick. You do need to be careful if the same matrix appears more than once in your einsum call though.
472.
by hansvm 1y ago
Is that an argument against the quality, saying that AI cannot (or some weaker claim like that it does not usually) produce "art"? Else, is it an argument of provenance, akin to how copyright currently works, where the same visual
473.
by hansvm 1y ago
If HNSW were accurate enough (and if this DB were much faster) then I'd have a use case. I wound up going down a different route to create a differentiable database for ML shenanigans though.
474.
by hansvm 1y ago
Maybe. If we use our powers too capriciously then they'll deter behaviors other than criminal behaviors. Like that boat of alleged drug traffickers we recently blew up -- that looks more likely to discourage boating within 1000 miles o
475.
by hansvm 1y ago
And what do you call the Model object holding those collaborating objects? Is that not still one object? The article explicitly supports your position that models can be complicated (see the paragraph starting with "to support more com
476.
by hansvm 1y ago
The heavy lifting happens in the definitions of `option` and `integer`. Those will take in whatever arguments they take in and output some sort of `Stream -> Result<Tuple<T, Stream>>` function. That might sound messy but to t
477.
by hansvm 1y ago
Which is a bit shocking in its own way, even if the numbers were break-even instead of 20-50%. Public transit is widely touted as being more efficient than the alternatives, but for most trips it's cheaper (factoring in maintenance, de
478.
by hansvm 1y ago
[0] https://danluu.com/julialang/
479.
by hansvm 1y ago
Interestingly, that gives a different response distribution from simply regenerating while the output doesn't match the schema.
480.
by hansvm 1y ago
They're born with neural hardware whose architecture has been optimized by evolution. Any choice of architecture imparts some inductive bias, making some problems easier and some problems harder to learn, and humans have the advantage

More ›