Eight questions
The Builder
You ship, then you argue
You meet a new technology by taking it apart. A humanoid robot is a bill of materials and a control loop to you, a token is a smart contract with an interface, and a market is something other people worry about while you finish the prototype. You trust what runs on your bench and distrust every slide deck.
Worth knowing: Your blind spot is the price. The best hardware in a category has lost to the cheapest more than once, and the protocol with the nicest code is rarely the one with the liquidity. Read the market essays once a month, even if they bore you.
For your type
Where to go next
Robotics
Robot foundation models
The brain moved before the body did
The interesting thing about the current robot wave is not the hardware. It is that a single trained model can now drive many different bodies, and that it learns from video of people rather than from code written for machines.
Robotics
Dexterity
The hand is the hard part
A robot can run, jump and carry a crate. Picking up a strawberry without crushing it is still a research problem, and it is the one standing between humanoids and most of the work they are sold for.
AI & Computing
Agents
What an AI agent actually does all day
An agent is a language model in a loop with tools and permission to act. That sentence contains both the reason it works and every reason it fails.
Tools
Tools for you
When a robot pays for itself
Automation Payback
The payback period and five-year saving of a robot cell, from purchase and integration cost, shifts, labour cost and the hours a person is freed for other work..
OpenTokens per month to money
AI Inference Cost
What an AI feature costs per month at scale: requests, tokens in and out, model tier, cache hit rate, and the comparison with renting a GPU for a self-hosted model..
OpenGPUs to megawatts
Data Center Power
How much power a cluster of accelerators draws, what it costs a year, how much heat it rejects and what a hyperscale site needs from the grid, from GPU count, chip power, overhead and electricity price..
OpenBasics
Start with these
How large language models work
A very large function that predicts the next chunk of text. Everything these models can do is a consequence of doing that extremely well.
02Why AI needs GPUs
A neural network is millions of multiplications that can all happen at once. A graphics chip was already built to do exactly that.