Learn

Anthropic courses (GitHub)

Where I start everyone - five structured notebooks taking you from zero to tool use and multi-agent systems, free and runnable in Colab.

Anthropic Claude certifications

The structured route into building properly with Claude - I am working through the CCA Foundations exam now and it forces you to think in the right frame.

Claude Code

Not a course - the tool to learn by building with. The single biggest unlock. Nothing teaches you AI faster than shipping something real.

Become an AI-Native Builder

Colin Matthews' hands-on course, created with Lenny Rachitsky - the one paid AI course I would point a non-technical builder to, built around real work in Claude Code, Cursor and MCP.

The paper that built the modern LLM - the transformer architecture everything since runs on. Dense, but this is the source.

The 2014 idea that you could map one sequence to another with a neural net - the seed of everything that became translation, then chat.

Where 'attention' first appears - letting the model decide which words matter to which. The mechanism the transformer later built the whole architecture around.

The paper that showed scale alone turns a language model into a general-purpose tool you prompt rather than train - why 'just make it bigger' worked.

People have argued about what counts as an agent since long before LLMs - this 1996 paper is the academic answer, a taxonomy that separates a real agent from 'just a program', and the modern debate keeps rediscovering it.

The textbook definition every AI student learns - the agent as a perception-action loop that senses its environment and acts on it. Where the modern idea comes from.

The clearest academic answer to 'what is an agent' - a model in a loop that reasons, acts with tools, observes, and repeats. The foundation under every agent framework.

The most ambitious internal AI memo any company has published - what rethinking an org around AI actually looks like at scale.

Matches exactly what I saw inside a PE firm this summer - every rollout splits into a barbell and the metrics hide it.

The clearest map of where the AI capex cycle sits - if you want to understand what deployment phase means, start here.

The best framing for why AI adoption numbers lie - the people doing the most are the least visible, and that gap shapes every strategy decision.

Short and sharp - the best answer to "should I start now or wait for a better model". Waiting is a decision too, and it usually costs more.

The essay that started it - Karpathy's 2017 argument that neural nets are a new kind of software you train rather than write. Read it before Software 3.0 to see how early he called it.

The clearest reframe of where all this goes - code went from hand-written, to trained weights, to English prompts, and Karpathy shows why plain language is now a programming layer anyone can use.

The most ambitious attempt to picture what a decade of powerful AI actually does to biology, health and government - not doom, not hype, a concrete reimagining from someone building it.

Evans' sharpest historical analogy - we have the models but almost none of the products, so the money is years away up the stack, not in the models themselves.

The instinct to sand AI down into normal IT automation is exactly how firms throw away the upside - defaulting to cutting staff instead of asking what becomes possible.

The reframe that AI stops selling software seats and starts selling the finished work itself - a genuinely different business model, and the clearest map of where value migrates next.

YC's new default - spend tokens not hires, keep teams tiny and let compute do the scaling. The Lightcone worldview in one crisp rule.

Intelligence gets cheap, so the scarce expensive thing flips to the context and data you feed it. Reframes where the money actually goes.

The most memorable jobs reframe of the year - not which roles get automated, but which archetypes survive when a generalist plus agents replaces the org chart.

The rebuttal to the whole AI-takes-the-jobs story - the roles everyone said AI would kill first are the ones that grew. A useful counterweight.

Karpathy's bet that one giant do-everything model breaks into specialised species of model, like biology - a fresh way to think about where the frontier goes next.

The next giants will not sell software to service firms - they will be the service firm, with agents doing the delivery. The Lightcone worldview made concrete.

The commoditisation argument made vivid - if compute ends up as cheap and ubiquitous as a high-street kebab shop, the value has to live somewhere else.

A sharp reframe of human value in the agent era - the durable work is precisely the stuff you cannot write a grader for, judgement and taste.

The open-source-everything bet and why he thinks open models reshape the whole power structure of AI - the canonical Meta worldview from the source.

Why he writes agent plans in HTML, not Markdown - models read structured HTML far more reliably past a couple of hundred lines, and it turns a plan into something you can watch tick off.

His 2023 call that AI is as fundamental as the graphical interface or the internet - a fifty-year computing veteran landing firmly on the 'this changes everything' side.

The clearest short case that agents - software you talk to that does the task, not apps you operate - become the main way we all use a computer. Where the platform shift actually lands.