Noah Weiss

@maxzimofan Another major plot hole: why did troy accept a large noncash gift from a foreign donor and never file a form 3520 with the IRS

odyssey plot hole: why was the Cyclops making goats cheese with no profit incentive or shareholders to create value for?

This is really cool. These guys did something special with their harness! I tried something similar in May, and the LLMs then couldn't figure out how to coordinate basic picking and placing actions. This new sol/Fable gen has massively improved in this area. I did some testing within codex for 5.6sol xHigh, which spent about 4 hours building its own harness and roughly 1.5 hours executing for the clip below. I also tested Fable xHigh in claude code, which took about 2 hours to set up but eventually got stuck in a loop after 3 hours during its attempt. Both seemed to improve dramatically over the course of the run; I believe the chat log drove RL-like self-improvement. Raw model capability is finally crossing the threshold where spatial reasoning works, but it's still some way from being competitive with VLAs on inference speed + cost. GPT 5.6sol xhigh "👇

Everyone pray cognition doesn't audit the database and find our $0.01/mo grandfathered rates 🙏

VLAs are cool!

Why is Fable emoting?

Gradient Ascent 📈

@johnrtian @philhchen @tkkong Bro is him

I'm joining Concept this summer to build the future of work with @philhchen and @tkkong! This was supposed to be my last free summer before Stanford: I'd planned to travel, hobbymax, and go on crazy side quests, not work. But Concept’s mission was too compelling to pass up, and conversations with Phil and TK won me over. Phil is an extraordinarily technical builder who also sees the bigger picture, and TK is a rare blend of extremely strong business and design skills. Excited for a summer of poker nights and shipping!

For my @CS153Systems final project (@AnjneyMidha) I built FARM, and in two months I'm bringing it to a factory in Dongguan, China to automate linework. For this generation of VLA models to hit reliable repeatability on multi-object tasks, you need a per-task prior, but fine-tuning a full policy per task is infeasible on compute and time, especially in low-volume factories where tasks change regularly. FARM runs on a full fine-tune of @physical_int π.5, a GSE expert fine-tune, then a per-task LoRA. The LoRAs hotswap locally and are trained on ~30 demos each. They're 1% the size of the model, initialize per action in under 40ms, and cut mean joint error on held-out episodes by ~37%. On tasks already in the FFT-GSE training data, the LoRA layer can half joint error. Because they live locally on the floor, they compose into new skills over time. Give it a plain prompt and it will understand its surroundings and automatically choose appropriate LoRAs (skills) to achieve the task, working out what order to run them in and checking its work after each action. This planning step only needs to happen once and can be saved so the same line can run for weeks without continual LLM inference.

Autonomous Chess ♟️ Stockfish v. Stockfish

Data collecting 👾

No solver. No move library. Just a 6DOF arm, a camera, and a few fixed points to benchmark LLM spacial reasoning! Grok 4.3 in @cursor_ai figured out the scanning, solving, and mapping cube logic onto moves the arm could actually make. @elonmusk

Just tested Composer 2.5.. Opus 4.6 quality output, Grok level moral negligence (Definitely NOT going into what I had it code)

The dichotomy 😭 @AnjneyMidha

Nike just lost the AI race

@maxzimofan Its over 📉📉

having a whole PIG in the stanford dining hall for lunar new year dinner must be an ai bubble indicator…

@DramaAlert Someone’s vibecoding in Birmingham

Birmingham is witnessing a rare phenomenon - pink haze all across the skies.

@functi0nZer0 What happens when you don’t use thinking mode

“You’re absolutely right: that was a fatal dose. My apologies.”

As n → ∞, p → 𝜋 Here's a cool mechanical 𝜋 calculator that I built over break

@sasquatchpog @interaction Awesome stuff! So cool to see how we approached the same problem differently

@NoahHWeiss @interaction if you wanna check out the current version :) github.com/Shashwatpog/po…

Created custom Canvas MCP server <> @interaction Poke. Connected to my canvas to scrape all PDFs/assignments/announcements/etc. Actual game changer.