Gardner Bickford

The world is ready for a Ghost Dog sequel with Camille Winbush as the samurai.

@SIGKITTEN Use hooks to prevent it from running python directly: anthropic.mintlify.app/en/docs/claude…

this guy thinks one can just tell claude

@VictorTaelin Cerebras is pretty wild. The model isn’t as strong as for-5 but I do planning with larger models and implementation with Cerebras. It’s really fast.

I'm bottlenecked by Codex. Every day, I have to choose 2 or 3 things to work on, because it will spend a long time on each. My routine is like: - wake up - write a prompt - send it to Codex - go eat something - come back, it is still not done - take a shower - come back, still not done - scroll for a while - finally done, test it... it works (80% of the times) - write the next prompt - ... what do I do with my time now ... So why don't I code manually? Am I lazy? Well yes I am, thanks for asking, but also typing all the code would most likely take more time than just letting Codex do it... so I spend most of my time waiting. Also it feels like I don't even need the next model to "feel the AGI", at least in practical terms. The day we have something *equivalent* to gpt-5-high, but 50x faster, my own work will move 50x faster too, because I'm 100% bottlenecked on its inference speed. Also, I really wish OpenAI would let people SFT it ): almost every single time Codex fails, it is because of lack of familiarity with my domain. I'm so tired of it being like "oh I finally understand how to write linear programs, this is what you want: ..." for the 100th time, only to forget it again in the next conversation. Seems like Karpathy himself went through this when making nanogpt...

@__tinygrad__ @danielhanchen have you met tinygrad?

Here are the open bounties. If you know how to make things fast, there's a lot of money on the table. Also some driver work, MLPerf models, and two new backends on offer.

@simonw I use hooks to prevent it from running `python`. The error message includes examples of how to run `uv`

Anyone got a pattern they like for running Claude Code such that when it runs "python -c ..." it's actually running a specific Python from a dedicated virtual environment managed by uv? I've done this with "source venv/bin/activate" before running "claude", more elegant idioms?

@NVIDIAGeForce GeForce Day

🟢 GEFORCE DAY IS BACK 🟢 To celebrate, we're giving away TWO GeForce RTX 5080 Founders Edition GPUs, signed by NVIDIA CEO Jensen Huang. Want one? Comment "GeForce Day" for a chance to WIN & stay tuned for more!

AOC & Walz play #madden twitch.tv/aoc

𝗪𝗘𝗟𝗖𝗢𝗠𝗘 𝗧𝗢 𝗘𝗔𝗥𝗧𝗛, 𝗔𝗦𝗧𝗘𝗥𝗢𝗜𝗗 𝟮𝟬𝟮𝟰 𝗥𝗪𝟭! ☄️ Here's a clear shot of the much-awaited small asteroid 2024 RW1 (#CAQTDL2) burning bright into a greenish 'fireball' over Lal-lo, Cagayan around 12:39 AM PhST, 05 September 2024. Did you see it too? 😊 📸 Valiant Aniñon

@intrstllrninja Awesome! What are you using as the judge harness? FastChat?

DPO self-gen in progress w/ function call validation unlike plain text response SFT dataset, one can easily create DPO rejections w/o "LLM-as-a-judge" for function calls function calls are validated against ground truth by matching function names and arguments for rejections

"All My Friends" soundcloud.com/hereiseverybod…

@gtiso There is a Thai restaurant in Hataitai and it is not called "Hataitai Thai". #missedopportunities

There's a food truck in Taihape called Thai Happy

@emptywheel reports on a #DOJ audit of #FBI use of #ParallelConstruction emptywheel.net/2022/10/03/on-… #privacy #warrantless #spying

@Rq9jG @tenacioustek @ProtonMail ProtonMail can be accessed via tor: protonmail.com/tor That doesn't change the dangerous mismatch of asserted vs real values @ProtonMail seems to have.

@tenacioustek @ProtonMail What should I use instead?

Hey @rmannibucau, nice article at rmannibucau.metawerx.net/react-oauth2-p… I just released version 2.0.5 of react-oath2-pkce which now supports custom providers. Check out github.com/gardner/react-…

@OrionHealth helped make the NZ Algorithm Hub a reality: algorithmhub.co.nz The algorithms aren't made public. An algorithm can be run in or added to Hub if you email them to sign up. Can we run it locally? Is there a public repo for collab? What are the licensing terms?

@teh_aimee A small win: theguardian.com/cities/2016/ju…

That feeling when your new business idea is an article in The Onion: theonion.com/new-co-op-airl…

@HaydnIvan what can you share about Data Trust NZ? Thanks 🙏

@peabnuts123 @ryankurte I haven't used it yet. Your tweet just reminded me that it exists. 😂

@gar @ryankurte And does like… ANY existing tooling work on it?

Electron but it's built on Firefox