Angehefteter Tweet
Todd Fisher
4.8K posts
Todd Fisher
@taf2
Living and working on https://t.co/eminYBe5tr
Maryland, US Beigetreten Ağustos 2008
463 Folgt329 Follower
@GregKamradt So I ran a llm fine tune over the weekend and it took many hours - codex monitors the long running process and can even send text message updates via CTM or any sms capable service - it’s really good at babysitting a long running process
English
When OAI employees say, “I let codex run all night…”
What framework they use?
How do you set up the task so it has enough work for 8 hours?
English
GitHub is a liability at this point, any open source alternative you know of?
@
We achieved Remote Code Execution on GitHub - and got access to millions of repositories belonging to other users and organizations 🤯 All it took was a single `git push` Here's how we did it (CVE-2026-3854) 🧵⬇️
English
Best usecase while outside waking for local ai on my phone - a personal pocket heater.
English
@vllm_project Cool. Does this version work well with Codex?
Codex cli was giving me errors connecting when using vllm to host Qwen 3.6 27B model with the responses api compatibility issues.
English
vLLM v0.20.0 is here! 752 commits from 320 contributors (123 new). 🎉
Highlights: DeepSeek V4, Hunyuan v3 preview support, CUDA 13 / PyTorch 2.11 / Transformers v5 baseline, FA4 as default MLA prefill, TurboQuant 2-bit KV (4× capacity), vLLM IR foundation.
Thread 👇
English
retweetet
As requested, I also asked GPT-5.5 (low) and Opus 4.7 (high) to fix the same bug.
GPT-5.5 (low) identified the correct root cause in 4m 14s and produced an almost identical fix in 2m 47s, using 164k tokens in total.
Opus 4.7 (high) “churned” for 6m 23s, using 87.7k tokens, and went down a completely wrong path.
It’s not even close, but that’s not a surprise.
@
I've completely changed my mind about 5.4 vs 5.5. Gave them the exact same task to investigate a fairly tricky bug. GPT-5.5 identified the bug and proposed a fix in 6m 59s using 117k tokens. GPT-5.4 took 8m 51s using 201k tokens, but it didn't find the bug and is asking for more information to investigate. Call me impressed.
English
qwen models are good but the censorship in these models can really cause problems with refusals on text that might be mis-taken as political. working on abliteration now to see if maybe these can be safer
English
If you're doing AI dev, you need to act like your system is rooted by North Korea. You cannot leave knives out in the kitchen, you cannot leave the passwords out on the counter. People are putting too much trust in alignment and not doing enough to "keep honest agents honest".
English
retweetet
PRO TIP
vLLM telling you to use `--enforce-eager` to avoid OOM because CUDA Graphs “don’t have enough VRAM”?
Don’t jump straight to eager mode
Try this first:
- lower `--max-model-len`, ex: 4k
- let CUDA Graph compile (which will be cached by torch.compile)
- restart, then raise context back up
You can keep the CUDA Graph performance gains without hitting OOM
English
Maybe not “every home” but by 1985, everyone in my circle had either a:
Commodore 64 (me)
TRS80 (Trash80)
an Apple II (the rich kids 🤣)
@
Yes! But in 1982, my high school didn't teach boys typing because it was for secretaries Schools didn’t anticipate computers in every home just 4 years later 🤓 In 2024, only 14% of US public K-12 schools taught appropriate AI use. History is repeating
English
Doing data privacy cleanup using new OpenAI model and set codex up with sms notify script via CTM to give me status updates … codex session been running for 20+ minutes just cooking while do other things
English
@TheAhmadOsman wait a second are you saying i can troll claude by mentioning HERMES.md ?
🤣
English