Joseph Turian

Should I use superterm.dev, superset.sh, or supacode.sh? Because we're competing with open-source openclaw by openai.

"GPT-5.2 ... [is] slightly smug." - Opus 4.6

Last registrations open for AI Takes #0, happening tomorrow in SOMA SF. Saturday, 5pm-8pm Open bar (beer + wine) and food: idea-salon-ai-takes-0.secretparty.io

CONTRACT.md: The Naughty List for AI Coding Agents My 2026 resolution: stop letting LLMs overcomplexify greenfield work. AGENTS.md = ghost of christmas past CONTRACT.md = caps complexity until a human says otherwise discussdontcode.com/zettels/contra…

What if we trained AIs to believe that AI doesn't exist yet?

What have you unlearned recently? “The illiterate of the 21st century will not be those who cannot read and write, but those who cannot learn, unlearn, and relearn. ” -Alvin Toffler

@mutableai please auto wiki torchsynth/torchsynth

Have you ever mentored someone? If so, I want to ask you a few questions.

Learning asynchronous REST programming from chatgpt is like having a friend who's super enthusiastic to hang out but absent minded and stoned all the time.

Any sufficiently complicated Torch codebase contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of PyTorch Lightning @LightningAI

@mkinyugo If I could easily run stable diffusion and dreambooth on a 1080 this would be amazing

🤔What if you could run any model on any hardware? 💻Imagine running the 11B Flan-T5-XXL on a device with just 4GB of RAM! 🎉Excited to introduce #odewel, the tiny library that makes this possible. github.com/Kinyugo/odewel

@ryanobjc @ilyasut I'd like to ask it questions like: "Why do you think it output FOO when I expected it to output QUUX?"

@turian @ilyasut Those already exist, sort of. Too much data volume typically. What does GPT4 do for that though?

Can someone make a time-traveling debugger with GPT4 which watches the entire state of my program's execution, and allow me to retroactively debug it? @ilyasut

@csteinmetz1 I would be also interested in losses that determine if there are stereo level and phase issues in the estimate that are dissimilar from the target.

What loss function would you like to see in auraloss? Have there been any interesting or important developments in loss functions for audio processing and synthesis this past year that you would use in your ML workflows?

@mkinyugo Sashimi / S4

If you had sequences of 16k+ tokens for a generation task, where global coherence is important, and you are limited to a GPU with little memory. What would be your alternative for the scaled dot product attention and why (optional)?

Prediction: AI will cause the price of work that can happen in front of a computer to decrease much faster than the price of work that happens in the physical world. This is the opposite of what most people (including me) expected, and will have strange effects.

@tszzl It would be so deliciously ironic if we automated software engineeers out of jobs first, and we replaced all the Uber drivers with robots later.

there’s two views of AI products where one is like “automate a large number of low value jobs using AI” and led to eg autonomous driving debacle. the other is like “increase the productivity of the highest value workers” and seems much more fruitful

@tszzl Vision and driving AI are hard because millions of years of evolution make these hard tasks seem easy. The things humans find hard (playing chess, picking stocks, programming a computer) might actually be lower-hanging fruit for AI to solve.

@alexgraveley Stochastic meta-descent (Schraudolph 1999, Sutton 1992). The reason I called my first company Metaoptimize. Optimizing the process of optimizing the process of optimizing the process of....

Has anyone looked at metarecursive networks? E.g. output is a weights update.

Why are your fonts so small? #bigfontenergy