I hate when agent get stuck on loop with review agent and they just spiral out of control

This was hard. Could have had Vice City on it, had almost 100% completed and then my HD died. Learned about RAID after that

I was playing Sonic on Game Gear and TDII on old PC still in early 2000s...

Thank @jullerino for great example for my agent to find on how to do Clerk device flow

What a time to be alive! My agent optimized ML framework just got on par or passed PyTorch on our performance benchmark and another agent using the framework got to spitting distance of our reference benchmark with our new architecture. Seems that one can solve problems by just throwing tokens at them

Interesting observation from my token usage. After fable 5.1 and astra dropped I rarely hit 1B daily tokens but I'm getting more done than before. With 5.6 so I was just burning tokens and running in circles compared to astra + fable (and now opus 5.5) combo

Astra: there is this one hypothetical issue and limitation, Fable: the repo rules says YAGNI, you'll need concrete proof that the current system is not enough How I love coding with agents

I've been building new ML training framework with lisp frontend and on performance side with @NVIDIAAI cutile-rs it's on small networks faster than compiled PyTorch, but not on larger models. Going to keep throwing tokens at this if it would help.

Opus 5.5 seems fine. One of the first prompts I did run 4 and half hours. Did the thing it was supposed to do. Fable 5.1 reviewed and brought the improvement in as is.

I've been building new ML training framework because I got annoyed about all the the mistakes the agents were doing when implementing training scripts and models in python. They didn't put the focus where they should have. To keep us moving the experiments forward I've been building really strict "research factory" system that contains scripts, and issue templates and review rules and all kinds of stuff to try to force the models to behave certain way. And today I realized that if my new lisp based ML framework lives to it's promise it's going to make most of those things obsolete! I'm so excited to start the real world experiments with the new system and do evals on how well it actually helps in making ML research at the speed of thought.

As a technical founder and a dad the only thing that allows me to be as productive I am is @t3dotcodes mobile app.

Does your ML experiment tracking tool do this? I've spend few months trying to figure out how to do AI assisted AI research the most efficient way. Been building tooling to help us move faster. I've tried different things to provide different interfaces for the agents to explore and diagnose the training runs. I build cli, 2 versions of API, python SDK for easy querying, was thinking about MCP. Nothing seemed to feel right. Outside of loss curves the diagnostics were always case dependent analysis. Then it hit me. Why don't I just expose the collected metrics for the agent via SQL? No more random data sync scripts, no more random cli command pipes created by the agents to find the insights on why does the gradient vanish at layer 7. Just raw SQL that the agents know. Accessed data cached locally, series data in parquet so one can query it also outside of duckdb.

Astra grinding. Original was 10x slower, then one round to match eager PyTorch. No going for the compiled

This morning my new lisp ML framework was 2-3x slower than eager PyTorch, now it’s faster. Next target compiled PyTorch Just throwing more tokens at the problem

I built a custom ML metric tracking system for our AI lab. The rational was to optimize it for the research agents, it has cli, nice api, nice python SDK. I rarely look at the metics on the dash, I always have agent to write script and custom visualization. Often syncing data to local duckdb to speed things up. It started to feel wasteful so decided to write @duckdb extension for it to speed up the agentic AI research even more. Just throw tokens at the problem like @theo suggests.

Have done few threads on @t3dotcodes damn this is massive productivity booster. Mainly the remote capabilities. So far 2 old macs seems to have been enough for remote work.

No wonder it wasn't learning, run set number 008 on hyperbolic nanogpt lab showed that the gradients get lost to rounding errors. Ugh, how does one solve this?

The blog hits so hard. I had left codex on a loop to solve a problem but the progress seems to be nowhere and then asked fable to check the work and it turned out to be almost 400k lines of irrelevant slop.