Why a curated list is more useful than a long one
Every week we get asked some version of the same question: what should I be reading on AI? It is a reasonable question. It is also a trap.
The trap is that the honest answer ("more or less anything by these ten people, in roughly this order, while skipping these eight categories") feels too small. It looks lazy compared to the 80-item Notion databases circulating on LinkedIn, the "ultimate AI reading lists" with three hundred bookmarks, the YouTube playlists organised by month. So most founders go and get one of those instead, and then the reading list itself becomes the procrastination. Bookmarked, never opened.
What works is the opposite. A short list, sequenced. Read the first item properly. Use the tools for a week. Then read the next one. The compounding here is not in the volume of content consumed; it is in the gap between input and practice. People who read one book about cycling and then ride for a year understand cycling better than people who read twenty books and never get on the bike.
So this is the list we actually use. It is short on purpose. Every item earns its place by changing how you think or what you do, not by being the most recently published. We have ordered the items by when they are most useful to read, not by date or prestige. Some items are books, some are essays, some are short videos or podcast episodes. Where there is a free version we link to it. Where the only good version costs money, we say so.
Two principles ran through how we built this list. First, mental models compound and tactics expire. The 2024 prompt-engineering blog post telling you the exact magic words to use with GPT-4 is already obsolete; the 2017 essay on what software written by neural networks fundamentally looks like is still load-bearing. So we lean toward the former where it exists. Second, read what the operators read. The people who actually build with AI for a living read a particular set of writers. Most of those writers are not the loudest voices in the public AI conversation. We have listed them.
A third thing we will own up to: this is the list as it stands today, shaped by what we read in the margins between making vermouth in a South London workshop and shipping AI tools for our own businesses and our clients. The list will move. The principle (short, sequenced, paired with practice) will not.
First week (you have a tab open)
The foundational mental models. Read these before you write a single proper prompt.
You have used ChatGPT, or Claude, or Gemini. Maybe you have shipped a draft email, summarised a contract, or asked it to clean up a spreadsheet. You are not sure what you do not know yet. The right reading at this stage is foundational: mental models for what AI actually is, where it comes from, and how to think about its outputs without either over-trusting them or dismissing them.
The goal of this stage is not to make you technical. It is to give you the vocabulary to ask sharper questions and the calibration to spot when something the model says is suspicious.
Co-Intelligence: Living and Working with AI by Ethan Mollick (2024), The single most useful book for a non-technical founder who wants a working mental model of where AI fits in their day. Mollick is a Wharton professor who has spent more time than almost anyone writing about how knowledge workers actually use these tools. His "four rules" framework (always invite AI to the table; be the human in the loop; treat AI like a person; assume this is the worst AI you will ever use) is genuinely portable advice. We hand this book to every new Absolution Labs hire on day one, and we re-read the four rules at the start of each client engagement. Read it first.
The Wait But Why AI series by Tim Urban (2015, with updates), The most accessible explanation of why AI is a different kind of technological shift, written years before ChatGPT. Some of the specific predictions have aged variably, but the framing (the difference between artificial narrow, general, and super intelligence; the runaway-capability argument) is still the cleanest version most people will ever read. Allow two evenings.
Software 2.0 by Andrej Karpathy (2017), Short, technical-adjacent, but readable. Karpathy is one of the small number of people whose writing has stayed useful across the whole arc of the AI boom. This piece reframes what software actually is in the era of neural networks, and once you have read it, you will never quite look at "code" the same way. The most quoted essay in the field for a reason. This is the talk we played for our first non-technical hire on day one before they touched a single tool.
What is ChatGPT doing... and why does it work? by Stephen Wolfram (2023), If you have ever wanted a calm, mathematically-grounded explanation of what is happening inside the language model when you ask it a question, this is it. Wolfram is occasionally accused of using these essays to advertise his own products; ignore the last 20%, the first 80% is the clearest layperson-friendly explanation of LLM internals on the open web.
Generally Intelligent podcast, the Karpathy episode (2023), If you only listen to one piece of AI media, make it Karpathy talking about the history of the field on Generally Intelligent or on Lex Fridman. The Lex episode is longer; the Generally Intelligent one is denser. He has a gift for explaining ideas at exactly the right level for an interested non-specialist.
Attention Is All You Need by Vaswani et al. (2017), You will not understand the maths. You do not need to. Read the abstract and the introduction. Look at the diagrams. This is the paper that introduced the Transformer architecture, which is the foundation of every modern LLM. Knowing that the paper exists, what it claims, and roughly when it landed is useful context for every subsequent piece of AI commentary you will encounter.
The Bitter Lesson by Rich Sutton (2019), One page long. Argues that the lesson of seventy years of AI research is that general methods which scale with computation always beat hand-crafted methods built on human cleverness. Important because it explains, in advance, why the AI boom of the last three years happened: not because of a clever algorithm, but because computers got big enough to brute-force the problem.
One Useful Thing by Ethan Mollick (ongoing), Mollick again, this time in shorter-form. Subscribe. Read it weekly. It is the most reliable signal-to-noise ratio in the public AI conversation, and Mollick has a gift for finding the specific use case that will change how you work next week.
The State of AI Report by Nathan Benaich (annual, most recent 2025), A few hundred slides, released each October. Skim it. The point is not to memorise the content but to calibrate your sense of where the field is right now. Spend an hour with it once a year and you will be better-oriented than most people in your peer group.
That is your first week. Nine items, but realistically you will read three or four of them properly and skim the rest. That is fine. The goal of this stage is to be no longer surprised by basic AI vocabulary or basic capability claims.
First month (you've used it for real)
Operator-flavoured material for the moment you stop being a passenger and start being a driver.
You have now used AI tools for a few weeks. You have started to notice patterns: the tool is great at first drafts and brittle at edge cases; it confidently invents facts when it does not know them; it follows instructions in ways that sometimes surprise you. You have also probably tried a few different tools and started forming preferences.
The right reading at this stage gets more practical. Less "what is AI", more "how do operators actually use this". You are moving from passenger to driver.
Anthropic's Prompt Engineering Guide by Anthropic (ongoing), The single best free resource on writing prompts that work. It is written for developers but the principles are universal: be specific, give examples, ask for structured output, decompose complex tasks. Skim the whole guide; read carefully the sections on "system prompts" and "examples" (also called few-shot prompting). This will quietly upgrade every conversation you have with an AI tool. We treat the few-shot section as required reading for every Absolution Labs operator before they touch a production prompt.
Simon Willison's Weblog by Simon Willison (ongoing), Simon is a long-time programmer and one of the most prolific writers about practical LLM use on the open web. His blog is dense with concrete examples of what these tools can do, written from the perspective of someone who actually builds things with them. Subscribe via RSS. Read his "weeknotes" posts in particular; they are a running diary of what an expert practitioner is actually trying.
Let's build GPT: from scratch, in code, spelled out by Andrej Karpathy (2023), You will not write the code. You do not need to. Watch the first 30 minutes. Karpathy walks through, on a whiteboard, what a language model is actually doing when it predicts the next word. By the end of the first half hour you will have a working visual model of "tokens", "embeddings", and "attention" that will hold up for years. We use this as the first technical-depth reference for every operator joining our work, and it is the single most useful two-hour investment a non-technical operator can make.
The Latent Space podcast by swyx and Alessio (ongoing), The best AI podcast for builders. Episodes are long (often two hours) and dense. You do not need to listen to all of them. Pick the episodes that interview people whose work you have come across (Mollick, Willison, Karpathy when he appears as a guest, the heads of evals at Anthropic and OpenAI). You will absorb the operator vocabulary faster this way than any other.
Generative AI at Work by Brynjolfsson, Li, and Raymond (2023), The most-cited study of what happens when you actually deploy an AI assistant to a real workforce (in this case, customer support agents at a Fortune 500 company). The headline finding: productivity goes up 14% on average, but the gains are concentrated in the least-experienced workers. Important not because the specific number generalises but because it gave a generation of operators their first solid data point against which to calibrate hype.
The McKinsey State of AI report by McKinsey (annual), Skim, do not read. McKinsey's annual survey of AI adoption by function and by company size is the closest thing to a reliable benchmark for "where is the median company on this". Useful for sanity-checking your own pace and for arguing back at advisors who tell you "everyone is already doing this" (they are not).
How to use AI to do stuff: an opinionated guide by Ethan Mollick (rolling updates), Mollick's most-shared post. A pragmatic survey of which tools are best for which tasks, updated as the landscape shifts. Bookmark it. Re-read it every few months because the answers genuinely move.
The OpenAI Cookbook by OpenAI (ongoing), Worked examples of common patterns: extracting structured data from documents, building a Q&A bot over your own files, doing semantic search. You will not run the code yourself, probably. But reading the patterns gives you a working sense of what "yes, that's possible" looks like in practice, which is a useful intuition to have when you are scoping your own internal AI projects.
AI Snake Oil by Arvind Narayanan and Sayash Kapoor (2024), The most useful counter-weight to the AI hype literature, written by two Princeton academics who are sympathetic to the tech but allergic to the marketing. Particularly valuable for SME operators because much of the AI product being sold to small businesses is the category Narayanan calls "predictive AI", which has a vastly worse track record than the language models you have been using. Read this before you sign a contract for any AI-powered recruiting, lending, or scoring product.
That is your first month. Nine items. Again, realistically you will read four or five properly. By the end of this stage you should be able to explain to a peer, without notes, what an LLM is, why it sometimes invents facts, and how to write a prompt that gets better results than the average operator would.
Once you've shipped something (operator-level)
The craft layer. What you read once a tool with real stakes and thin margins is running in your business.
You have now built or commissioned at least one real piece of AI tooling for your business. A custom workflow in Make or Zapier with an AI step in it. A chatbot on your site. An assistant that triages your inbox. Something that actually runs and that you depend on.
At this stage the questions get sharper. How do you know it is still working? How do you tell when the model is drifting? What does "good" look like when you are evaluating a vendor's claim? What is the right way to think about cost as you scale? This is the operator-level material. It is also the material that pays for itself fastest, because the cost of a tool that quietly degrades for a quarter in a business with real stakes and thin margins is far higher than the cost of an hour spent reading Eugene Yan.
Ahead of AI by Sebastian Raschka (ongoing), Raschka is a researcher and educator who writes the clearest summaries of the most important new AI research papers, aimed at intelligent non-specialists. Subscribe. You will not read every issue. The ones you do read will quietly upgrade your understanding of why models behave the way they do.
Eugene Yan's blog by Eugene Yan (ongoing), Yan is a machine learning engineer at Amazon who writes the best practical guides to evaluating AI systems in production. His pieces on building evals, prompt design, and "RAG done right" (retrieval-augmented generation, the standard pattern for grounding an LLM in your own documents) are the canonical references. If you are commissioning AI work from a third party, his posts will give you the vocabulary to know whether their answers to your questions are serious.
When we first integrated AI into Asterley Bros's wholesale-enquiry handling, the Eugene Yan posts on evals were what stopped us from shipping a tool that scored 80% in our internal test set but would have hit roughly 40% on the actual messy inbox. We built a small eval harness against twelve months of real enquiries before flipping the tool live, caught two failure modes nobody on the team had thought to test for, and only then put it into production. The Yan reading was the difference between a tool that looked good in demo and a tool that held up when the inbox got weird.
Hamel Husain's blog by Hamel Husain (ongoing), Hamel writes about evaluating LLM applications from the practitioner's chair. His piece "Your AI Product Needs Evals" is the single most-shared link among teams that have actually shipped AI tools and discovered that they do not know if the tools are getting better or worse over time. Read it once. Refer back to it every time you are about to deploy a change.
The IBM and academic literature on model drift (various authors, 2023-2025), "Model drift" is the phenomenon where a model that worked well at launch quietly gets worse over time, either because its inputs change (data drift) or because what counts as a good answer changes (concept drift). You do not need to read the technical papers. You do need to know the phenomenon exists, why it happens, and what monitoring you would want to detect it. IBM's plain-English explainers and the various survey papers on this are a good way in.
The Latent Space Eval-Driven Development series by swyx (ongoing), The "eval-driven development" framing is the AI-era analogue of test-driven development. The idea that you write evaluation cases first, then build the AI system to pass them, then watch the eval pass-rate over time. Read enough of this material to understand why teams that ship AI tools without evals are flying blind.
Lilian Weng's blog (Lil'Log) by Lilian Weng (ongoing), Lilian writes extraordinarily clear long-form posts on AI topics, originally as an OpenAI researcher and now independently. Her posts on agent systems, hallucination, and prompt engineering are the closest thing the field has to a textbook. Operator-level depth but very approachable prose.
Anthropic's "Building Effective Agents" by Anthropic (2024), The clearest public articulation of when to use simple workflows versus when to reach for autonomous agents, written by the team building Claude. When we shipped the first proper background-worker agent for Asterley Bros, this was the canonical reference we re-read three times before locking the architecture. It saved us from buying complexity we did not need and gave us the vocabulary to explain to the rest of the team why a deterministic workflow with a single AI step was the right answer rather than a fleet of autonomous agents. If you are about to commission "an AI agent" to do something in your business, read this first.
The "Prompting Guide" by DAIR.AI by Elvis Saravia and contributors (ongoing), A reference-grade collection of prompt patterns: chain-of-thought, self-consistency, ReAct, tree-of-thoughts. You will not use most of these. Knowing what they are called, when they are used, and what problems they solve is part of operator literacy.
The "Models All The Way Down" essay by Jacob Andreas et al. (2024), The clearest framing of why composed AI systems (an LLM calling another LLM calling a tool) are the dominant shape of production AI work, and how to think about reliability across that composed surface. Useful when you are scoping a project that involves multiple AI steps chained together, which by stage 3 is most of them.
That is your operator-level reading. Nine items again. By the time you have read four or five properly, you will be able to interview a vendor or contractor about their AI capabilities and know within ten minutes whether they are serious. That alone is worth the time.
For founders specifically
The business-strategy layer. Where AI meets companies, not where AI meets the abstract.
A short bonus section. Where AI meets business specifically, not AI in the abstract.
Reid Hoffman on AI (various episodes of Masters of Scale and standalone talks, 2023-2025), Hoffman is the most coherent public voice on what AI means for the structure of companies. His framing of "Superagency" (the idea that AI gives existing people more agency rather than replacing them) is a useful counter-weight to the more apocalyptic public discourse. Pick any one substantial interview from the last 18 months.
The McKinsey GenAI economic-impact report by McKinsey (2023, updated 2024), The widely-cited estimate that generative AI could add $2.6-$4.4 trillion in annual value to the global economy. The number itself is contested. The breakdown by function (marketing, sales, software engineering, customer operations) is genuinely useful for thinking about where the leverage is in your own business.
Bain's "Generative AI" briefing series by Bain & Company (ongoing), Bain's reports are denser and more recent than McKinsey's, and tend to land closer to ground truth on what is actually happening at mid-market companies. Particularly useful for the consumer goods and retail sections if you are FMCG.
Stratechery by Ben Thompson (ongoing), Thompson writes the best business-strategy analysis of the AI industry, period. The paid newsletter is £150 a year and worth it if you are trying to understand competitive dynamics among the model providers, the implications for downstream sectors, and the strategic shape of where the value will accrue. His "aggregation theory" framework, developed for the previous tech wave, is being actively re-tested against AI in real time.
The Acquired podcast, episodes on Nvidia, OpenAI, and TSMC by Ben Gilbert and David Rosenthal (2022-2024), Multi-hour deep dives into the history of the companies that built the substrate the AI boom runs on. Less directly actionable than the operator literature, more useful than it looks for understanding why certain companies have the leverage they have and what that means for pricing and availability over the next five years.
The Andreessen Horowitz "16 Changes" memo and follow-ups by a16z partners (2023-2025), A16z's framing of how AI changes specific industries. Variable quality. The pieces on consumer, retail, and the creator economy are worth reading; the more technical pieces are largely redundant with the operator-level material above.
The Absolution Labs operating principles canonical we publish in this library is the synthesis of a year of reading-then-acting on the items in Stage 3 and Stage 4. Every principle in it was written down the first time we hit the failure mode it now prevents, while the operator-level reading was running in the background giving us the vocabulary to name what we were seeing. The reading was not the deliverable. The principles document was. The reading made the document possible.
That is six founder-specific items. Add three to four of these to the main reading list and you will have business-strategy context to match your technical literacy.
What to ignore
The categories we deliberately leave off the ladder, and why each one fails the bar.
The hardest part of curating this list was deciding what to leave off. Here is what we deliberately excluded, with reasons. This is the operational stuff nobody writes LinkedIn posts about, because saying "do not read this" gets less engagement than saying "here are 80 things you must read". The honest dispatches version of the curation is the negative list.
The "10 prompts that will change your business" content. Almost all of this is junk. The few useful prompt templates in it are already in the Anthropic guide above, in better form. Lifecycle of a prompt template: shared on LinkedIn, copied a hundred thousand times, obsolete in six months because models change. Skip.
The doomer-versus-accelerationist debate. The public conversation about AI risk has bifurcated into two camps that are mostly talking past each other. There are serious arguments in both camps. There are also enormous amounts of low-quality content from both camps that will absorb your attention without changing what you do on Monday morning. Read one good thing on each side (Stuart Russell's Human Compatible on one end; Marc Andreessen's "Techno-Optimist Manifesto" on the other) and then move on.
Tool-comparison content older than three months. "ChatGPT vs Claude vs Gemini" comparisons date faster than almost any other kind of content. The capability rankings flip every few months. Pick the tool that is most convenient for you, use it for a quarter, then re-evaluate. Do not read someone else's comparison from January.
LinkedIn-style "I asked GPT to do X and you won't believe the result" posts. Engagement-optimised, almost always misleading about what the model is actually doing. The screenshotted conversation is usually cherry-picked from many attempts. Treat as entertainment, not signal.
Vendor-funded reports about AI ROI. The Salesforce, HubSpot, Microsoft, and Google AI-impact reports have selection bias by construction. They ask current AI users how AI has impacted them. The people who tried it and gave up are not in the sample. Take the headline numbers with calibrated scepticism. The McKinsey and Bain reports above are not perfect but are less self-interested.
Speculative "AGI in 2027" timeline content. Everyone has a take. Almost none of the takes are predictive. Read Mollick's calmer "this is the worst AI you will ever use" framing instead and orient your business planning around capability ranges rather than dates.
How to use this list tomorrow
Reading the list is not the same as benefitting from it. Here is the approach we recommend.
The 12-week beginner-to-operator reading plan
- Week 1: Pick one Stage 1 item The default starting point is Co-Intelligence by Mollick if you read books, or the Karpathy Generally Intelligent podcast if you do not. Allow about three hours.
- Use the tools in production for a fortnight before Stage 2 The reading only sticks if you have ground experience to attach it to. Pick one workflow in your business this week, do it with AI for two weeks, then come back to the reading.
- From Stage 2 onward, read one item per fortnight Pace yourself. The compounding is in the gap between input and practice, not in the volume of input.
- Keep a single-page running notes document As you read, capture: the one or two ideas you would most like to remember, the one tool or pattern you want to try, and any open question the piece raised. This document, more than the reading itself, is what you will come back to in six months.
- Re-read your notes monthly, not the source material The notes will be more useful than re-reading the originals, and writing the notes is what actually consolidates the learning.
- At month three, share something publicly A short LinkedIn post, an internal team write-up, a blog post on your company site. The act of explaining what you have learned to someone else is the highest-leverage way to make sure it is actually stuck.
By the end of twelve weeks, you will have read roughly six items properly, skimmed another dozen, and produced a one-page notes document that captures your operating model for AI in your business. That document is the deliverable. The reading list was the scaffolding.
Most founders who actually do this report the same thing: by the end of the twelve weeks, they no longer ask "what should I be reading on AI?", because they have their own running shortlist of writers whose new work is worth their time. That is the goal of the exercise. We are trying to make the curated list obsolete in your head, by giving you enough vocabulary and calibration that you can curate it yourself.
Want help applying this to your specific business?
If you would like Absolution Labs to spend 30 minutes walking through where AI would actually be useful in your business, given where you are right now, book a free audit. We will talk through your current workflows, what is worth automating, what is not, and which two or three of the items on this list are most relevant to where you are heading. No pitch, no obligation.
Book your free audit →