gliptic

The name for it _is_ abliterate. It's a portmanteau of ablate and obliterate.

Zig-native immediate-mode

dvui?

The borrow checker doesn't decide when things are dropped. It only checks reference uses and doesn't generate any code. This will work exactly the same as long as your program doesn't violate any borrowing rules.

Camping is like.. regulatory capture? Stretching the analogy thin here.

But that fine-tuning is done only on those 100-200 good samples. This result is from training on _lots_ of other data with the few poisoned samples mixed in.

This doesn't really make sense to me. Most cats I've known react to such reflections without ever having seen a laser pointer in their life, for the same reason they react to laser pointers.

Glyph advance or line spacing is not part of the bitmaps.

I think Sweden was deploying the Drakken (Dragon) and later the Vigen (Lightning).

The names are much less flashy, Draken (The kite, due to the shape) and Viggen (The tufted duck) :P.

`f(++i, ++i)` is/was indeed UB, but the example in munificent's comment was `foo(print(1), print(2))` which as far as I know is not even if both `print` calls read/write the same memory.

Yes, but that's just a subset of expressions where unspecified sequencing applied. For instance, the example with two `print()` as parameters would have a sequence point (in pre-C++11 terminology) separating any reads/writes inside the `print` due to the function calls. It would never be UB even though the order in which the prints are called is still unspecified.

The author is using the term in the way that everyone else understands it. They are not aware of your unusual definition.

The evaluation order is _unspecified_, not undefined behaviour.

Recent developments on BB(6) previously posted here: https://scottaaronson.blog/?p=8972

It's really not. Looks like something like JS with almost identical syntax.

Also, the compiler (and what else?) isn't even implemented [1].

[1] https://github.com/profullstack/smashlang/blob/master/src/co...

Dynamic Typing: Flexible type system with runtime type checking

Uhm, what does this have to do with C/Zig?

Ah, FastML is an extremely overloaded name.

It's strange that someone from FastML can be confused about this, unless it's supposed to be a bad joke.

I'm not seeing how this is pull in any sense. Calling recv on the channel doesn't cause any result to be computed. The push of the previous operators will cause the compution to continue.

EDIT: Ok, I guess because they are bounded to 1, the spcs will let the pushing computation continue first after the "puller" has read the result, but it's more like pushing with back-pressure.

For pixelation you can use another technique invented for astronomy: drizzling [1].

[1] https://en.wikipedia.org/wiki/Drizzle_(image_processing)

You're probably underestimating what you can do with these.

Any that can run 70B at >5 t/s are >$2k as far as I know.

I guess it will be a bigger issue the longer it's been since they stopped making them, but most I've heard (including me) haven't had any issue. Crypto rigs don't necessarily break GPUs faster because they care about power consumption and run the cards at a pretty even temperature. What probably breaks first is the fans. You might also have to open the card up and repaste/repad them to keep the cooling under control.

I arbitrarily chose $1k as the "cheap" cut-off. Two 3090 is definitely the most bang for the buck if you can fit them.

That's 1.2 t/s for the 14B Qwen finetune, not the real R1. Unless you go with the GPU with the extra cost, but hardly anyone but Jeff Geerling is going to run a dedicated GPU on a Pi.

Your best bet for 33B is already having a computer and buying a used RTX 3090 for <$1k. I don't think there's currently any cheap options for 70B that would give you >5. High memory bandwidth is just too expensive. Strix Halo might give you >5 once it comes out, but will probably be significantly more than $1k for 64 GB RAM.

So yes, it's a limitation of their own API at the moment, not a model limitation.

R1 is trained for a context length of 128K. Where are you getting 8K/32K? The model doesn't distinguish "thinking" tokens and "output" tokens, so this must be some specific API limitations.

while it's also fabricated in an older node

Is this 7 nm node really older than raspberry pi 5's 16 nm node?

Where are you seeing Hyperbolic offering DeepSeek V3 API? I'm only seeing DeepSeek V2.5.

Apparently you don't believe in reading titles either.