Writings
Thoughts, ideas, and things I've learned along the way.
Robot Model Experiment Series 7: SmolVLA (Real DAgger Labels)
Three versions of the same SmolVLA tube pick. v7a added rescues and fell to 5%. v7b relabelled the same rescues as real DAgger and reached 30%. v7c kept v7b and added a grip guard at run time, reaching 48% on the relaxed watcher.
roboticsAIRobot Model Experiment Series 6: SmolVLA (DAgger + RL Improvement)
v6 kept the reward idea from Series 5 and fixed what went wrong: better rescue data from the improved expert rescue skill and a grade that counts hovering against a run. It picked up the tube on its own in 22% of runs, the best of any model so far.
roboticsAIRobot Model Experiment Series 5: SmolVLA (RL Grader)
DAgger got SmolVLA to pick up the tube on its own for the first time. So every run got a grade, a simple reward, and the next model trained most on its own runs that went best. It got worse, and the reason is a lesson about rewards.
roboticsAIRobot Model Experiment Series 4: SmolVLA (DAgger)
Demos taught SmolVLA where to go, but never how to finish from where it actually ends up. So the model drives, a watcher spots when it gets stuck, an expert rescue skill takes over, and the corrections become the next training data. Success went from 0% to 16%.
roboticsAIRobot Model Experiment Series 3: SmolVLA (Teleops post train)
I fine tuned SmolVLA, a pretrained robot model, three times on demos of picking up a tube. None of the versions picked it up on its own, but each failure pointed at what the next dataset needed.
roboticsAIRobot Model Experiment Series 2: ACT (simple demonstration → action model)
I trained an ACT robot policy on 20 demos of picking up a tube, on a Mac mini. It copied my demos within 2 degrees and still failed on the real table. Here is why.
roboticsAIRobot Model Experiment Series 1: Opus 5.5 + Robot Skills
I gave Claude Opus 5.5 a desk robot arm, two cameras and a Mac mini. It wrote a program that picks up a tube in 15.6 seconds with no demos, and it still needed me for a few things.
AIroboticsGeneralist vs Specialized Robots, Explained With LLMs
What is the difference between generalist and specialized robots? A plain explanation using LLMs and scripts, with real examples and the economics of each.
AIroboticsWhy Robot Hands Are Hard
A robot can see a cup perfectly and still fail to pick it up. Five plain reasons the hand is the hardest part of a robot.
AIroboticsFree Robot Data Is Not the Data You Need
3,553 hours of bimanual robot data went free this summer. Teams on the exact same arm are still collecting their own by hand. Here is what open robot data collection actually buys you, and what it cannot.
AIroboticsThe Data Wall: Why Robot Foundation Models Are Stuck on Demonstrations
The robot models are ready. The data to make them reliable in the physical world is not. Why demonstration data is the real bottleneck in physical AI—and why you can't scrape your way out of it.
AIroboticsTeleoperation, UMI, and World Models: How Robots Learn From Us
A field guide to the ways robots learn manipulation from humans—teleoperation, handheld grippers, egocentric video—and how VLA policies and world models fit together. The vocabulary of physical AI, explained.
AIroboticsAgentic-native software: the model is shared, your context isn't
Agentic-native software isn't a tool you operate. It's an agent that assembles capability from a model, your content, and your tools. The model is shared with everyone; your context and tools are the moat.
AIagentsIs Reading Books Still Worth It?
I learn most things by asking AI now. The research says reading still builds a child's reasoning, while for adults Q&A is faster but can leave knowledge shallow.
AIreflectionSkills, Not Just Tools — Rethinking Agent Capabilities
Why the agent ecosystem's obsession with tools is the wrong abstraction. The difference between building copilots and building agents that actually work autonomously.
AIagentsBitcoin Renaissance - Babylon Restakes
A Comprehensive Overview of the Bitcoin Ecosystem Part III - Bitcoin Restaking. Exploring how Babylon enables staked Bitcoin to generate additional yield while maintaining self-custody.
BitcoinDeFiresearchSurveying the AI x Crypto Landscape
A comprehensive framework for categorizing blockchain AI companies, examining their specific areas of concentration, technological approaches, and value propositions.
AIblockchainresearchA Two-Part Approach to Understanding zk Coprocessors
A comprehensive taxonomy of zk Coprocessors, examining how Data Access and Computation coprocessors solve trust and scalability challenges in blockchain development.
blockchainresearchA Comprehensive Overview of the Bitcoin Ecosystem Part II - Bitcoin Layer 2s
An in-depth analysis of Bitcoin scaling solutions including ZK Sovereign Rollups, Optimistic Rollups, Inscriptions, and Sidechains, examining the tradeoffs between security and scalability.
BitcoinresearchA Comprehensive Overview of the Bitcoin Ecosystem
An in-depth exploration of Bitcoin inscriptions, the Ordinals protocol, BRC-20 tokens, and the emerging ecosystem of fungible and non-fungible assets on Bitcoin.
Bitcoinresearch