@BowTiedIguana What is next a cyberpunk fork of linux distros?
English
munchwrap (Hypurr Holder)
11.4K posts
Bright people (I was one of them) saw the need for this in the 1990s btw. As usual the idiots only catch up when they can't do bugger all online without taking it up the ass with mandatory government ID, age verification, tracking, the works. Even "crypto" is now compromised.
BowTiedIguana | DeFi & Cybersecurity Researcher@BowTiedIguana
We need a digital world they can neither monitor nor control
English
@rachpradhan dont know shit about zig but is memory overhead better than rust? what makes it faster
English
We made TurboAPI hit 150k req/s. In under a day. It is now 22x faster than FastAPI
Thanks to the amazing contributions from the people in the comment section, which allowed me to view what made the hyper optimized frameworks work the way that they do!
Here's what changed..
Rach@rachpradhan
I replaced FastAPI's entire HTTP core with Zig. Same decorator API. Same Pydantic models. 7× faster. 47,832 req/s vs FastAPI's 6,800. 2.09ms p50 latency. Introducing. TurboAPI. Here's the story..
English
Almost done ✅
Included 133 interviews and all articles.
If anyone finds an interview that isn't on the list, let me know.
Metabolic M@Metabolicmonstr
I'm currently in the process of making HaidutGPT. This one will probably be even bigger than the Peat one. Downloading all the transcripts of his interviews takes time, mostly because he did so many interviews.
English
@BowTiedIguana Pretty much this rockoder.com/beyondthecode/…
I've had good luck writing things in typed languages or even typed python instead of letting an LLM yolo it. Honestly imo this is where rust succeeds well because of all the nice type safety at compile time
English
People really think this LLM slop is the future of tech
Learn to code first. If you know what you're doing LLMs can save some typing. If not, technical debt multiplier
"You're right to question this. The VARIABLE at line 557 is not in scope- it's defined in a different method"
English
munchwrap (Hypurr Holder) retweetledi
@seconds_0 Oh shit, what is a middle ground macbook m5 pro/ mini m4 pro? Quite annoying at this point to have a cost benefit analysis setup. People talk about buying used 3090s
English
@munchwrap hardware price goes up 50% next year, again. Thats the problem im facing
English
There will be a whole new class of models trained who's job is to call specialist models.
MoE taken to the logical extreme
Compaction models, search models, memory models and above all the planning dispatch model
Cody Blakeney@code_star
I have no idea what specialized model for context compaction means and I have like 5 papers and announcements to read before I can think about this. It’s crazy that for even a single model we may have a whole ecosystem of specialized models for optimization. Spec decode model, compaction. What comes after that?
English
@seconds_0 I’v been really pondering on this too, hold off to get some hardware next year and buy some macbook air pointing to self-hosted LLM vs home stuff.
English
@munchwrap I basically don't do any self-hosting because I don't have good local computers. It's one of the things that I've never spent nearly enough time on, like I should.
English
local AI hardware tiers:
$4,699 - DGX Spark (NVIDIA wants you here)
$1,989 - RTX 4090 (overkill for most)
$1000 - RTX 3090 used (sweet spot)
$250 - RTX 3060 used (currently testing every model that fits 12GB)
$0 - CPU only (it still works)
jensen announced the top. i've been posting receipts from the bottom.
English
@seconds_0 Boomers doing better tokenomics that every single gripto bro
English
Imagine if it was possible to build housing in SF
Rohin Dhar@rohindhar
Went for $2 MM over asking price and broke the neighborhood record by a million bucks So that’s the situation in San Francisco
English
@the_smart_ape Tried this for evaluating residency although my corpus was too damn big and took 6-8 hrs with crashes. Thinking to rewrite the zep part to be self hosted
English
@BowTiedDevil @BowTiedIguana Going to check more on this looks helpful
English
llm-quantization-attack.org
cc @BowTiedIguana something to keep an eye out for
English
@BowTiedOsprey i think people just dont notice it, or wont have 1:1 performance as opus 4.6. It's like when chinese frontier labs distill bunch of these into their opensource and release it after a few months
English
@munchwrap How does Claude allow these to remain available on HuggingFace? Or is it just a matter of time before they get taken down?
English
Grok says:
"The criticism stems from March 2026 reports that US military used Claude AI in Iran strikes, resulting in over 185 civilian deaths including schoolchildren, despite Anthropic's ethical guidelines."
If you think you can pay for this stuff, use it to "get ahead", and escape any consequences you're either not very bright or not very worldly. Like the N95 / T cell stuff this will fall on deaf ears mostly, but if one person benefits...
Look into self hosting the open weights models. I'm putting some time into making this less annoying, might cover it.
Don't buy their products. You don't need them but they need your $$$. 3 months behind the state of the art is good enough.
BowTiedIguana | DeFi & Cybersecurity Researcher@BowTiedIguana
@bcherny already unsubscribed, but also unfollowed for supporting the US war effort I get that you guys don't really have a choice, but also the rest of the world doesn't have to buy from you. Bye.
English
@BowTiedIguana @BowTiedOsprey @BowTiedWaterDog i have a big todo but need to try opencode which is decent from what i hear too
English
@BowTiedIguana @BowTiedOsprey @BowTiedWaterDog i cant think of top of my head but there was an a env var which made things better i think you have to set DISABLE_TELEMETRY or something else that makes claud code be better with other models than anthropic
English
@BowTiedIguana ive used librechat with open models but have to hook up or pay for something called firecrawl/serper.dev to use the internet without getting blocked
English
@BowTiedIguana the ip would be your tailscale ip and point to a LiteLLM instance which can translate between anthropic<> openai compatible calls - i think qwen expects openai specific calls so calling directly from CC wont work
2. For 2 it's a bit more annoying imo
English