The maths for the back propagation algorithm in neural nets uses the chain rule calculus attributed to Leibniz.

Marvin Minsky - built a learning neural network at Harvard. Gave up. Would have to be ridiculously large to learn anything interesting. Frank Rosenblat in his PhD came up with a device for tabulating data and created the perceptron with a single layer of adjustable weights.

Inductive bias and generalisability are the two man differences between human and machine learning

Metalearning to train neural networks with less training by including inductive biases. For example initial weights to help make language learning possible with 5 years of data rather than 5000. Outer loop and inner loop. Optimise for initial weights that allow the model to learn better from limited data. What biases are needed to improve language learning?

The amount of data it takes to train an LLM to undertand a language supports Chomsky’s views of how complex an object language is.

Kant's "synthetic a priori" judgments. Our minds are hard-wired with a framework to process reality.

The big difference between human minds and brains and LLMs is inductive bias. Human children acquire language with about 5 years of data. LLM training requires the equivalent of 5000 years of data. What explains the 4995 years of priors found in a human child?

Bayesian reasoning - you have your priors, you get new data, you calculate a likelihood function and you update your priors. Where do your priors come from? We have rough feelings. Intuitions. How do we learn language in the first place.

The ability to break down problems and set goals is a product of a resource constraint. It’s a tool that emerges when you can’t reason endlessly.

Algorithms to Live By - Tom Griffiths and Brian Christian. Argues that many human dilemmas—like deciding when to stop looking for an apartment or how to organise a closet—are essentially computational problems that computer scientists have already solved

Computer science is a better guide to rationality than economics

Karl Friston - The Free Energy Principle. Proposes that all living systems, from cells to brains, minimize a mathematical quantity called "free energy" to maintain their integrity and resist disorder. It acts as a unified theory of brain function, suggesting that action and perception work together to minimize sensory prediction errors, effectively making the brain a proactive prediction machine.

The Rational Use of Cognitive Resources - Tom Griffiths, Feb 2026

Don’t use haiku or local models with personal agents if security is a concern. They are gullible and susceptible to prompt injection

The latest generation of models have a lot of post training to avoid prompt injection attacks.

Toxicity in the crypto world

There’s a subculture of people who follow viral growth and try to quickly snaffle the name on socials etc.

Peter clearly has very relaxed permissions for his own personal agent!

A lot has changed recently. Four months development of the frontier AI models is all it took to go from miserable attempts to convert a codebase from one language to another to a completely successful conversion (in the case in question: into the slightly obscure Zig).

The Age of the Lobster

It doesn’t have to replace you as a person. Just find the corners of your work and let it replace those

First step for new adopters should be to reproduce your own work with an agent.

"Rejecting AI before spending serious time with it ...is like trying git for an hour and deciding you’re not more productive with it."

Always have an agent running in the background because there is always something to do. Disable its ability to notify you: remain in control of interruptions. You choose when to interrupt the agent.

Identify the tasks that require thinking and the tasks that don’t, and delegate the only the tasks that don’t require thinking

On “harness engineering”. When you see AI do a bad thing, try to build tooling it could have called out to avoid or course correct away from that bad thing.

"GitHub and forges [in general] do not work with agentic workflows today"

Some projects are now allowing PRs only from vouched-for contributors. Those who break trust are denounced and gated.

The flood of Claude-generated PRs is causing a lot of pain in the open source community and an inevitable backlash reaction. Where does this lead? Blanket rejection of AI contributions to open source projects is also not the answer.

There is an unspoken compact in open source that effort in the submission is met with effort in the review. The surge in low-effort, and often low-value AI generated open source contributions breaks the compact.

Interesting parallel between housing permitting and AI regulation -- both suffer from a bias toward inaction dressed up as caution.

Klein argues the problem isn't left vs right but vetocracy vs action. The permitting system has become a tool for blocking rather than building.

"Half of all entry level white collar jobs could be replaced in the next few years. Can maybe see the hints of graduate job losses and hints of the productivity boom but it's very early."

On government regulation. There has been close to no successful movement on AI regulation. Open question if the testing can actually be effective.