dnautics

21 ms·

601.
by dnautics 10mo ago
you understand that stack allocation can OOM too?
602.
by dnautics 10mo ago
i was a software guy, sorry, but those token rates are correct and what was flowing through my software. i believe there was a special deal on super special fpgas. there were dsps involved.
603.
by dnautics 10mo ago
> For example if after one branch you don't converge, but after two you do, do you accept that? you should refactor so that it's representable. > Add dynamically sized values (e.g. arrays) and loops/recursion and now yo
604.
by dnautics 10mo ago
i like fil-c, so i would say "fil c sprinkles magic fairy dust on C and makes it safe (at the cost of perf and elevated risk of crashing)
605.
by dnautics 10mo ago
not enforced for any given implementation is hardly "unknown". presumably the tin comes with a label saying what's inside
606.
by dnautics 10mo ago
i made a quip and realized that's not a bad description of what fil-c does
607.
by dnautics 10mo ago
ok? i worked on those devices, those numbers are real. theres a reason why they compare to h200 and not b200 > I have worked with FPGAs that outperform H200s in Llama3-class models a while and a half ago
608.
by dnautics 10mo ago
exactly! "guaranteeing the safety of C" sir what did you think that meant, sprinkling magic fairy dust to make it work!!?
609.
by dnautics 10mo ago
as someone building an analyzer for zig, if you sent something like this through the system im building, it would get handled by Zig's optional type tagging; but lets say you did "free" (so your result is half freed, half not
610.
by dnautics 10mo ago
pretty sure when obama murdered Abdulrahman al-Awlaki (nb: not talking about the more famous Anwar) that was unprecedented. Trump later murdered Abdulrahman's sister, but at that point, it was "precedented" by obama.
611.
by dnautics 10mo ago
eh, germany and japan seemed to go okay, grenada too. korea kind of a mixed bag (it took decades for it to not suck)
612.
by dnautics 10mo ago
i would feel better about that if the biden administration also prosecuted obama. they didn't. besides trump I (nor biden) didnt do any new foreign adventures AFAICT. we had a blissful 8 years of waning US imperialism
613.
by dnautics 10mo ago
probably not B200 level but better than you might expect: https://www.positron.ai/ i believe a B200 is ~3x the H200 at llama-3, so that puts the FPGAs at around 60% the speed of B200s?
614.
by dnautics 10mo ago
unlikely. trump didnt held obama accountable for all sorts of crazy things that happened during his administration (bombing libya, drone striking a us citizen minor, using USAID to mount a fake vaccination campaign for DNA surveillance in
615.
by dnautics 10mo ago
what you are proposing is functionally equivalent to "wrapping an mcp in a cli" which is what I mentioned in my root comment.
616.
by dnautics 10mo ago
I don't think this is correct. For inference, the bottleneck is memory bandwidth, so if you can hook up an FPGA with better memory, it has an outside shot at beating GPUs, at least in the short term. I mean, I have worked with FPGAs t
617.
by dnautics 10mo ago
I don't know if they'll reach $3B, but at least one company is using FPGA transformers (that perform well) to get revenue in before going to ASIC transformers: https://www.positron.ai/
618.
by dnautics 10mo ago
i hate python but if your bottleneck is that sqlite query, optimizing a handful of addition operations is a wash. thats why you need to at least have a feel for these tables
619.
by dnautics 10mo ago
that seems token inefficient. why have the llm do a full round trip. load the skill which contains the potentially hundreds of lines code then copy and paste the code back into the compiler when it could just run it? not that i care too t
620.
by dnautics 10mo ago
lets say i'm in erlang. you gonna include a script to unpack erlang bytecode across all active modules and look through them for a function call? oorrr... have that code running on localhost:4000 so that its a single invocation away,
621.
by dnautics 10mo ago
yeah but a skill without the mcp server is just going to be super inefficient at certain things. again going to my example, a skill to do a dependency graph would have to do a complex search. and in some languages the dependency might be h
622.
by dnautics 10mo ago
no that makes no sense. the skill doesn't do anything by itself, the mcp (can be) attached to a deterministic oracle that can return correct information.
623.
by dnautics 10mo ago
idk. its pretty easy to look over the diffs and catch the worst of stuff. a few things do slip through bug i dont find it hard to scour over the code, bad stuff stands out or is easy to grep. here's my current project, judge for your
624.
by dnautics 10mo ago
i don't understand this mcp/skill distinction? one of the mcps i use indexes the runtime dependency of code modules so that claude can refactor without just blindly grepping. how would that be a "skill"? just wrap the
625.
by dnautics 10mo ago
claude is really good at zig. amount of examples on the web is not a good predictor of llm capability, since we know you can poison an llm with ~250 examples. it doesn't take much.
626.
by dnautics 10mo ago
aw damn. thanks for clarifying.
627.
by dnautics 10mo ago
> LLMs take the heart out of what made me like programming: building stuff yourself with your near infinite lego box of parts and coming up with ideas yourself. i feel like that's all im doing with llms. just in the last hour i r
628.
by dnautics 10mo ago
out of curiosity, does that generally mean that (linux) OOM killer can't get you? IIRC the oom killer is only triggered on new page request, and only the requesting process is eligble for the murder?
629.
by dnautics 10mo ago
If we get REALLY nitpicky, zig currently (but not in the future) allows unbounded function recursion with "theoretically" assumes unlimited stack size, so it's potentially "still technically theoretically turing complete
630.
by dnautics 10mo ago
i think you mean "exactly not Turing complete"

More ›