New ask Hacker News story: Bun rewrite and FLT formalization had nearly identical resource usage

Bun rewrite and FLT formalization had nearly identical resource usage
4 by Zsfe510asG | 1 comments on Hacker News.
The Bun rewrite took 11 days and used 5.9B tokens: https://ift.tt/iy1CJZO The FLT formalization took 11 days (!) and used 6B tokens. Why the similar numbers? Why are the numbers so low? 6B is just enough for 6 agents at 1000 tokens/s running for 11 days.

Comments

Post a Comment

Popular posts from this blog

New ask Hacker News story: Qwen3.8 Fetches Weird URLs

Qwen3.8 Fetches Weird URLs 2 by Luker88 | 0 comments on Hacker News. I am working on opencode with the last qwen3.8-27B from unsloth (with dynamic 3.0 quants). while checking documentation for a rust project, instead of fetching the url it had listed in its own reasoning, it tried fetching urls like: https://routify-file-proxy-sg.oss-ap-southeast-1.aliyuncs.com/proxy_temp_file/production/2026-08-20/trace_0baf8c2b17874702866132490e0b56/requestId_97422c0041074091a68367881782d7b7/7c0791038d2f58b3a52c0f07714d3627.html?Expires=1818574288&OSSAccessKeyId=LTAI5t…QVZr&Signature=96y%2B…w%3D this happened in the same session, different batches, 16 times in total. It convinced itself there was some rewriting proxy in the webfetch tool. `Expires`, `trace` and `requestId` were consistent for a few requests, but they do change. `AccessKeyId` was always the same, "LTAI5t…QVZr" I did not add the `…` in the signature and access key. Probably training links, but that `Expires` number...
Read more

New ask Hacker News story: Ask HN: Releasing code under AGPLv3, but want to block LLM reconstruction?

Ask HN: Releasing code under AGPLv3, but want to block LLM reconstruction? 2 by zionsati | 0 comments on Hacker News. I am preparing to release a software project under the AGPLv3. The goal is traditional copyleft reciprocity - if you use it or host it, share your changes. However, I am realistic about the current legal landscape. Big tech corps are treating public code as free raw material for LLM training under the banner of "Fair Use". I am concerned that a company will ingest my codebase and use an LLM to effectively launder the logic, allowing their users to prompt a clean, closed-source recreation of my software without triggering the AGPL. Do we have a licence specifically to prevent this but still keep OSS healthy and alive? Do we have a llm.txt / robots.txt that LLM scrapers respect? I feel that the whole OSS model is under threat here, even more than before (e.g. big corps earn billions from Linux instances without having to pay any software licensing cost, but t...
Read more