My readme Kanban

Explore how to measure taste and how to communicate it to an agent. This is hard even for humans. How do we know a painting is good? How do we express what we feel when we see one? Conveying that is already hard between people. How do we get an agent to understand it? Like Taste Labs: make taste measurable, starting with design.

Models perform extremely well on short term tasks. Can they combine those small tasks and keep them going long enough to run a business? That takes long running work, coordination, and a memory of what already happened. As Andon Labs describes it: an AI run store, a café and a radio.

Trying to understand what happens inside LLMs. Problems I care about:

• faithfulness: do explanations reflect what the model actually computes?
• scaling feature discovery to frontier size models
• evaluating explanations without a human in the loop
• interpretability as a tool for alignment

Projects: NLASmith, a framework for experimenting with Natural Language Autoencoders. More to come.

Finishing my Systems Engineering degree at UTN.

Shipped at Rebill (YC W22): led the SDK behind one click checkouts and built PCI compliant infra.

Implement an original and hilarious readme view. Now the cards actually move between columns.