rupayanc.bsky.social

Maple hitting 218 tok/s is real. Treating that as proof the model is reliable is the mistake. Its own model card says “minimal general reinforcement learning.” Speed is a deployment benchmark. Trust is a capability benchmark.

100B on one CPU at 6.58 tok/s sounds ridiculous. The footnote matters: that’s a BitNet benchmark configuration on an M2 Ultra with 64GB RAM. Microsoft’s released model is 2.4B, and BitNet isn’t a switch for ordinary models. github.com/microsoft/Bi...

Fine tune Llama 3.1 8B on a 4GB laptop GPU. Soup streams the frozen base from system RAM one decoder layer at a time. RTX 3050 Laptop: 3.32GB peak, 119.6 tok/s, bit exact. The catch: measured on v0.72.2, not re run on a 4GB card since the fp32 fix. 5.6k stars. github.com/MakazhanAlpamys/Soup

Train a 64M language model from scratch for 43 cents. One RTX 3090, about 2 hours, 3 yuan of rental at Chinese prices, per the minimind README. The catch: one epoch of the Zero variant on the small datasets. Tiny model, real training run. 59k stars. github.com/jingyaogong/minimind

A picture of your code costs less than the code. An image is priced by its pixel dimensions, not by the text inside it, so pxpipe renders your context as PNGs first. About 3.1 characters per token against 1 as text. Their readme tells you not to trust the savings figure and to measure your own.

I stopped reading my coding agent's permission prompts weeks ago. I just approve. clawk boots a throwaway Linux VM per project and runs the agent inside with its safety flags off, deliberately. Their own line: "the host filesystem isn't hidden behind deny rules; it was never mounted."

Firefox, the real one, compiled to WebAssembly, running inside a browser tab. Feels like the cleanest incognito ever shipped. But a tab can't open raw sockets, so every page goes through a WebSocket proxy Puter hosts. Their own line: "Web content is proxied through a Puter-hosted Wisp server."

Ship That Code: rebuild Redis, Git, a database and an OS kernel from an empty file. 179 courses, 37 languages, and your code runs against a real executor rather than a video. Its page says "free forever". Its own pricing line says "free while we grow". Those are two different promises.

Nvidia reportedly buying Hugging Face, $12.9B. Reported, not confirmed. The Information says agreed. Business Insider says unsigned, could still fall apart. Nvidia's newsroom, five days on: nothing. Hugging Face turned down $500M from Nvidia last year to avoid one big investor. Now this.

New Qwen model, 125B params, 6B active per token. I saw "6B active" and got excited. Then I opened the GGUF repo. The most aggressive quant they publish is 67.6GB. Active parameters are about compute, not memory. You still hold the whole thing. Learned that one the slow way, again.

friend: bro just use a habit tracker me: i built a full RPG. quests, XP, bad habits as named bosses friend: how long did that last me: 27 days friend: and the tracker i sent you me: ...also 27 days

My AI has one standing rule: argue with me when my reasoning is weak. This morning it pulled up my own track record and asked what is actually different this time. Uncomfortable and correct. An agreeable AI is just a mirror.

Amazon lost 120k orders to AI-assisted code that skipped senior review. Lightrun 2026 report: 43% of AI code needs manual debugging in prod. Typing was never the bottleneck. The senior engineer saying "no, not this way" was. AI just made that conversation feel optional.

The scariest part of the new Claude model isn't what it can do. It's the second one Anthropic built alongside it and decided not to ship. "Too dangerous to release" used to mean nuclear material. Now it's how a lab describes a coding model.

The devs struggling most with AI aren't the ones who refuse it. They're the ones who use it like an answer machine instead of a thinking partner.

Solo-founded startups went from 23.7% in 2019 to 36.3% today. AI collapsed execution to $3K/year. But most founders still do distribution by hand. Building was never the hard part.

Funded startups have been running AI agents in production for months. That gap just closed to $0.08 an hour. The infrastructure excuse is gone. What's the first workflow you hand off?

watching everyone rebuild karpathy's llm wiki this week. cool tech. but if your second brain can be auto-generated overnight by any model, is it really yours? or is it just a cache you happen to own?