Last updated: 11 September 2026. I will keep updating this post as the conversation and the technology change.
This post is essentially my evolving notebook. It brings together the writers, researchers, podcasters, bloggers, YouTubers and workshop leaders whose work has helped me understand AI over the past two years.
I put it together by going back through my YouTube, X and Spotify histories and revisiting the blogs and newsletters I read. The people here often disagree, sometimes fundamentally. I am not presenting a curriculum, a ranking or an argument for or against AI—I am trying to understand who these people are, what expertise they bring and how their views fit into the wider debate.
I pay for the top-tier plans on both ChatGPT and Claude and constantly test the latest models on my own work. I know I am still only scratching the surface, so if you are deeper into any of these areas, please suggest the people I should be reading, watching or listening to.
1. Where is AI now?
The easiest way I have found to understand the market is to separate frontier models from everything else. Frontier models are the strongest general-purpose systems available at a given moment; the rest of the market includes smaller, cheaper and more specialised models that may be better for a particular job. The Stanford AI Index is the clearest annual overview I have found of capabilities, investment, adoption and policy.
What real big problems of humanity could AI solve?
Science is where the promise becomes more concrete. AlphaFold predicts the three-dimensional shape of proteins—the tiny biological machines whose shape helps determine what they do. That can give researchers a much faster starting point for understanding disease and designing experiments. AlphaFold 3 extends the work to interactions among proteins, DNA, RNA, drugs and other molecules.
Malaria: researchers are using predicted structures to explore which parts of a protein could go into a vaccine.
Neglected diseases: DNDi has used AlphaFold in work on Chagas disease and leishmaniasis.
Pollution: protein prediction can help scientists design enzymes that break down plastic.
Antibiotic resistance: structural information can reveal ways to block bacterial resistance mechanisms.
Cancer: AI is being studied across imaging, diagnosis, drug discovery and molecular interaction prediction. This review of AI in oncology is a useful reality check on both the possibilities and the limitations.
Where I would go to learn about AI and science
Demis Hassabis — DeepMind’s work on frontier AI, AlphaFold, robotics and scientific discovery.
Pushmeet Kohli — AI for science and the research behind AlphaFold.
Regina Barzilay — machine learning for cancer detection and drug discovery.
Daphne Koller — machine learning applied to biology and medicines.
Eric Topol and Ground Truths — medicine, health research and AI, written for a wide audience.
Google DeepMind: The Podcast — Hannah Fry’s accessible conversations with the scientists doing the work.
Hugging Face Daily Papers and The Batch — for the geeks- regular ways to see new research without relying only on social-media summaries.
2. What does “open” versus “closed” AI actually mean?
A closed model is accessed through a company’s app or API while its weights and much of its training process remain private. An open-weight model makes its learned parameters available so others can run or adapt it. That is not always the same as fully open-source AI; the Open Source Initiative explains the stronger definition.
I find the politics striking. The United States built much of the modern open-source software movement, yet its leading AI labs chose closed frontier models. Chinese companies have become major suppliers of capable open-weight models. Open models widen access and local control; closed models can make safety, support and rapid product development easier. Both also concentrate different kinds of power.
Which models currently lead?
These rankings change quickly. In the Artificial Analysis Intelligence Index v4.3 published on 7 September 2026, the leading closed models included Claude Fable 5.1, GPT-6 Astra and Claude Opus 5. Astra appears to be leading much of the current conversation, although it was tied rather than alone at the top of that index.
The leading open-weight entries included GLM-5.3, Kimi K3 and GLM-5.3 Flash. Qwen3.8 is another important family to watch.
Where I go to learn about open AI
Nathan Lambert’s Interconnects — open models, post-training, research culture and policy.
Hugging Face — model pages, technical details, licences and a large open-model community.
Goose — Block’s open-source AI agent, useful for seeing what an open ecosystem looks like in practice.
Artificial Analysis — capability, speed and price comparisons across open and closed models.
3. How do we actually measure which model is best?
There is no single “best” model. One may lead at maths, another at coding, another in speed or price. Benchmarks can also be gamed, contaminated or disconnected from everyday work. I therefore look at several independent sources and then repeat the same real tasks on the models I pay for.
The benchmarking platforms I watch
Artificial Analysis — a composite capability index alongside speed and price.
LMArena — blind human preferences between model outputs; useful, but preference is not the same as factual reliability.
Stanford HELM — transparent evaluation across many scenarios and metrics.
LiveBench — frequently refreshed questions intended to reduce benchmark contamination.
Epoch AI’s FrontierMath — difficult mathematical problems designed for frontier systems.
Hugging Face Open LLM Leaderboard — comparisons focused on open models.
My own test is simpler: can the model handle the documents, research, writing, product work and follow-through I actually need, and can it do so reliably enough to trust?
4. What could the next ten years hold?
No one knows, so I find it more useful to compare named views than to offer one confident forecast.
If you ask Elon Musk
The direction is abundance built on intelligence, robotics and energy. Tesla’s “sustainable abundance” master plan connects autonomous systems and humanoid robots to a future in which goods and services become dramatically cheaper.
If you ask Marc Andreessen
AI is a broad amplifier of human intelligence that could improve education, healthcare, productivity and creativity. Why AI Will Save the World is the clearest statement of his optimistic, pro-acceleration case.
If you ask Andrew Ng
The near-term opportunity is not one magical superintelligence but millions of useful applications. His talks on agentic workflows focus on what teams can build now, while warning against treating every task as suitable for automation.
If you ask Demis Hassabis
We may be in the “foothills of the singularity”: systems approaching more general intelligence, with major consequences for science, medicine and society. His view combines ambition with a strong emphasis on safety and scientific institutions.
Where I go to compare future views
The Economist’s interview with Yuval Noah Harari — power, institutions, democracy and the historical significance of AI.
Dwarkesh Patel — long-form interviews where lab leaders, economists and sceptics have room to explain their assumptions.
5. Why do some people think AI could end the world?
The catastrophic-risk argument is that a future system more capable than its operators could pursue the wrong objective, deceive people or gain access to critical systems before we know how to control it. The critics question how likely that is, whether today’s technology justifies the analogy, and whether the focus distracts from harms already happening.
The catastrophic-risk thinkers and their critics
Eliezer Yudkowsky — the most severe case that sufficiently advanced, misaligned AI could be uncontrollable and fatal.
Yoshua Bengio — catastrophic-risk concerns combined with technical safety research and governance proposals.
Geoffrey Hinton — Nobel Prize-winning pioneer of neural networks. He warns that systems smarter than humans may become difficult to control and could pose an existential risk, while also stressing AI’s potential benefits.
Center for AI Safety — a short statement showing how widely extreme-risk concerns are shared.
AI as Normal Technology by Arvind Narayanan and Sayash Kapoor — a critique of treating AI as an autonomous historical force instead of technology deployed by institutions.
Melanie Mitchell and Gary Marcus — sceptical examinations of claims about reasoning, reliability and rapid progress toward AGI.
International AI Safety Report — the best starting point I have found for separating evidence, disagreement and uncertainty.
6. Why I don’t think AI will replace people in every domain
This is my current conclusion, not a prediction that disruption will be mild. A job is a bundle of tasks, relationships, judgment, responsibility, trust, physical action and local knowledge. AI can automate parts of that bundle without becoming the person who accepts liability, understands an unspoken concern or lives with the consequence.
The ILO–NASK global index estimates that one in four jobs has some exposure to generative AI, while finding transformation more likely than wholesale replacement. That still leaves serious questions about eliminated roles, entry-level work, bargaining power and who keeps the productivity gains.
Andrew Ng’s human context advantage
In Andrew Ng’s conversation with Silicon Valley Girl, he argues that judgment and taste often come from context the AI does not have: experience, relationships, organisational history and memories of what failed. He also raises the problem of cognitive offloading: AI can help us complete a task without helping us learn it.
People studying AI and work
Erik Brynjolfsson — augmentation, productivity and how organisations capture value from technology.
Daron Acemoglu — the economics of automation and the difference between genuinely useful tasks and cost-cutting.
David Autor — expertise, job quality and how technology can restore or erode middle-class work.
7. Why are people protesting AI—and what do data centres and job losses have to do with it?
People are rarely objecting to one thing called “AI.” They are responding to particular costs: job loss and weaker bargaining power; creative work used without permission; surveillance; misinformation; the loss of entry-level learning; concentrated corporate power; and the electricity, water, land and noise required by data centres.
A data centre is a building full of computers. Frontier AI requires large numbers of specialised chips, plus electricity, cooling, networking and backup power. The International Energy Agency explains the global energy picture. A study in Nature Sustainability examines possible US water and carbon costs. The UK government’s public-engagement survey shows that concern also centres on inaccurate information, privacy, deepfakes, jobs and damage to creative or critical thinking.
Where I go to understand the opposition and the physical costs
SemiAnalysis — chips, power, networking and data-centre economics.
UN Environment Programme — the environmental impact of AI across its full life cycle.
Ed Newton-Rex and Fairly Trained — copyright, consent and licensed training data for generative AI.
Ada Lovelace Institute — public interest, regulation, biometrics, work and the social effects of data-driven systems.
International Labour Organization — evidence on which occupations and tasks are most exposed.
These sources do not agree on whether the costs are justified. That is useful: the central questions are who benefits, who pays and who gets a say.
8. Where I learn how people are using ChatGPT
This is one of the practical centres of the article. I pay for ChatGPT’s top-tier plan and use the newest models for research, writing, document analysis, coding and product work. But most of my imagination has come from watching other people share their screens: seeing the problem they start with, the context they provide, the mistakes the model makes and the workflow they eventually keep.
YouTube: the individuals I learn from
How I AI with Claire Vo — probably the most directly useful format I have found. Guests share their screens and show how they research, write, manage products, build agents and run parts of their companies.
Jeff Su — clear, edited walkthroughs of ChatGPT for research, projects, documents and everyday knowledge work. His ChatGPT Work walkthrough is a good introduction.
Every and AI & I — conversations in which writers, founders and researchers demonstrate how they use ChatGPT and other models to think, create and build.
Silicon Valley Girl — interviews about AI, careers and business. Her conversation with Andrew Ng broadened how I think about learning, judgment and human context.
Kevin Stratvert — patient, beginner-friendly walkthroughs that are useful when a feature is unfamiliar and I want to see every click.
Matt Wolfe — frequent experiments, comparisons and product news. I use him to discover what to test, rather than treating every first impression as a verdict.
Official reference: I use OpenAI’s channel to check release demonstrations and product announcements after seeing how independent teachers use the tools.
Substack, newsletters and written guides
One Useful Thing by Ethan Mollick — research-informed experiments with ChatGPT at work and in education, including what happens when ordinary people are given frontier models.
How I AI’s workflow library — searchable demonstrations organised by guest, company and tool. It is useful when I want an example for a particular kind of work rather than general advice.
Every — writing about AI-assisted thinking, writing, coding and organisational design, produced by a team that also builds AI-native products.
Simon Willison’s Weblog — detailed, reproducible testing of models, tools and security problems. It is one of the best places to understand what a new capability really does.
The AI Exchange — Rachel Woods on turning responsibilities and processes into AI playbooks that teams can repeatedly use.
Official reference: OpenAI Learn is useful for checking current features and documentation, but it is not where most of my practical ideas come from.
Spotify: long-form examples and ideas
How I AI — the audio version of Claire Vo’s practical workflow interviews. The video is better when the screen matters, but the conversations still generate ideas while walking or travelling.
AI & I with Dan Shipper — writers, founders and creative people explaining how they use ChatGPT and Claude to think, write, code and make decisions.
Practical AI — real-world implementations, architecture and the gap between an impressive model and a dependable system.
The Cognitive Revolution — longer conversations with founders, researchers and policy thinkers that broaden the picture beyond one product or workflow.
Official audio: The OpenAI Podcast adds conversations with the people developing the models and with builders using them.
The limitation running through all of these sources is that fluent output is not the same as correct output. The demonstrations are starting points for experiments; important work still needs sources, privacy controls and human review.
9. Where I learn how people are using Claude
I also pay for Claude’s top-tier plan. I use it most for long documents, careful writing, analysis and, increasingly, Claude Code and agentic work. Claude becomes easier to understand when I watch somebody build an entire workflow with it—not only a polished final result, but the planning, context, permissions, corrections and review that made it work.
YouTube: people teaching Claude and Claude Code
How I AI with Claire Vo — real Claude and Claude Code systems for product management, research, writing, company knowledge and agents, normally shown on screen rather than merely described.
Jeff Su — accessible demonstrations of Claude for knowledge work and design. His Claude Design video shows how reusable context and a design system change the output.
Brock Mesarich — explanations of Claude and Cowork for people who are not engineers. His Cowork concepts video covers context, memory, skills, connectors and subagents.
Sabrina Ramonov — end-to-end Claude Code builds, agents, automation and reusable skills, often aimed at people learning to build without a traditional software background.
IndyDevDan — opinionated, principle-led material on agentic engineering, planning, context files, subagents and production discipline.
Cole Medin — technical builds combining Claude Code with agents, retrieval, MCP servers and automation.
Every — interviews and demonstrations of how Claude Code is changing software, writing and small-team company building.
Official reference: After watching the independent teachers, I use Anthropic’s channel and the Claude Code team’s workflow video to compare those methods with the product team’s own approach.
Substack, newsletters and written guides
Every — one of the richest collections I have found on Claude Code, AI-native products, compound engineering and the effect of agents on knowledge work.
Simon Willison’s Weblog — careful testing of Claude releases, computer use, coding agents, security and prompt injection, usually with enough detail to reproduce the experiment.
Latent Space — deeper writing and interviews for people moving from using Claude to engineering products and agent systems around models.
Sabrina Ramonov — practical guides and reusable material for Claude, agent workflows and automation.
Official reference: Anthropic Engineering and Anthropic Academy are where I check the technical detail once somebody’s demonstration has given me an idea worth trying.
Spotify: Claude, agents and building
Claude Code for normal people from How I AI — Grace Clarke demonstrates skills, voice mode and service-business tools she built with Claude.
AI & I — Dan Shipper’s conversations with people using Claude and Claude Code for writing, software, research and running companies.
OCDevel Claude Code — a technical, hands-on podcast covering session design, context files, hooks, permissions, skills, subagents and MCP.
Latent Space: The AI Engineer Podcast — conversations about model infrastructure, agents and the engineering decisions behind AI products.
You Need to Learn Claude Code in 2026 by Sandeep Swadia — an accessible account of the shift from chatting with a model to delegating multi-step work.
Claude Code AI Subagents by Jake Morrison — an AI-narrated audiobook about specialised agent teams. I include it for discovery, while treating its performance and cost claims as ideas to verify.
Tutorials date quickly because Claude changes quickly. I use the independent sources for ideas and working methods, then return to Anthropic’s documentation and my own experiments before adopting a workflow.
10. Where do Microsoft Copilot and Google Gemini fit?
I have spent much less time with Copilot and Gemini, so I do not want to pretend to offer a deep personal comparison. Copilot is most relevant where work already lives inside Microsoft 365, Windows, GitHub or Azure; Microsoft’s Copilot documentation and the Phi model family are the useful primary sources.
Gemini sits across Google Search, Workspace, Android, Cloud and DeepMind’s model work. The Gemini overview shows the product family, while Sundar Pichai’s conversation with Rowan Cheung is a better source for Google’s wider strategy.
From learning about AI to building with it
Over the last few weeks my learning has shifted from reading and experimenting to building. These are the projects where I am testing what the technology can and cannot do:
CompanyBoard.org — a dashboard that brings UK company deadlines, directors, filings, documents and compliance information into one place.
Loop AI — a relationship layer for subscription and membership businesses that identifies customers needing attention and drafts a personal follow-up for human approval.
Buy Back Brixton and BrixtonCulture.com — connected work around community ownership and making Brixton’s culture more visible and economically powerful.
The simplest doorway to all of this work is gvdp.co.uk.
What I am learning now: deploying agents at scale
I am now looking at how several agents can use tools and company data, divide work, check one another and escalate consequential decisions to people. The difficult questions are permissions, security, evaluation, cost, monitoring and recovery when something fails.
Model Context Protocol specification — the authoritative description of the open standard that connects AI applications to tools and data.
Zapier MCP — a practical example of connecting services such as calendars, email and CRM to ChatGPT and Claude.
Anthropic on trustworthy agents — research on autonomy, safeguards, permissions and prompt-injection risk.
Building AI-native organisations
“AI-native” interests me as organisational design, not as adding a chatbot to an old process. These sources explore how companies might redesign work around context, feedback, evaluation, customer trust and human approval:
Garry Tan: New Rules for Founders — leverage, reusable instructions and building companies in the AI era.
Paul Graham on startups and great founders — why AI progress can be astonishing in one task and unreliable in the next.
Andrew Ng: Generative AI for Everyone — a non-technical course on capabilities, limitations, business opportunities and social consequences.
Marketing and sales
The useful question here is not only how to produce more content. It is how AI can help teams understand customers, analyse conversations, explore positioning, create variants and run better experiments without generating more spam.
Marketing AI Institute — structured education for marketing and business leaders.
Claire Vo’s How I AI — screen-shared demonstrations of real workflows.
Gong Labs — analysis of sales conversations and deal activity; useful company research that should be read with Gong’s commercial interest in mind.
Where I go to keep learning
These are the sources I return to, grouped by what they help me understand.
My current top three AI podcasts
If I had to reduce a very long list to three, these are the ones that currently give me the widest spread of perspectives:
Dwarkesh Podcast — my choice for long, deeply prepared conversations with the people building, studying and criticising frontier AI. It is where I go when I want an argument developed properly rather than compressed into a clip.
Google DeepMind: The Podcast — Hannah Fry makes frontier research, AlphaFold, agents, robotics and safety understandable without removing the science. It is produced by DeepMind, so I treat it as an excellent primary source rather than independent coverage.
How I AI — Claire Vo asks people to show the real workflows they use. It is the most practical of the three; the video edition is especially useful because the screen-sharing is part of the lesson.
Newsletters and podcasts for practical AI
One Useful Thing — Ethan Mollick’s research-informed writing about AI, work, education and everyday practice.
How I AI — Claire Vo’s screen-shared workflows. Its strength is seeing exactly how somebody does the work.
Sabrina Ramonov — tactical material on Claude, agents and automation, sometimes presented alongside bold commercial claims.
Elena’s Growth Scoop — AI through the lens of growth, product and marketing.
Lenny’s Newsletter — broader product and growth thinking with real implementation stories.
Newsletters and podcasts for understanding the frontier
Interconnects — Nathan Lambert on models, post-training, open AI and frontier research.
Understanding AI — Timothy B. Lee and colleagues explaining technical developments with limited jargon.
Import AI — Jack Clark’s detailed research and policy digest. Clark is an Anthropic co-founder, so it is insider analysis with an identifiable institutional perspective.
Dwarkesh Podcast — deeply researched interviews rather than quick news.
The Algorithmic Bridge — Alberto Romero on AI, culture, philosophy and society.
SemiAnalysis — chips, compute, data centres, networking and the economics beneath the models.
AI as Normal Technology — a rigorous counterweight to inevitability, hype and simplistic AGI narratives.
YouTube for using the products and building
Jeff Su — unusually clear, structured product walkthroughs. The two videos linked above cover ChatGPT Work and Claude Design.
Silicon Valley Girl — Marina Mogilko’s interviews on AI, business and careers. Her conversation with Andrew Ng is where I encountered the ideas about cognitive offloading and human context advantage. Some titles use aggressive “top 1%,” income and future-of-work framing, so I focus on what the guests actually say.
Brock Mesarich — accessible explanations of Claude and Cowork. I have linked his Claude Cowork concepts video in the Claude section.
Matt Wolfe — frequent AI news, tool testing and practical demonstrations. He also operates FutureTools, so I use him for discovery rather than treating a product mention as an independent verdict.
Y Combinator — founder interviews, startup advice and discussions about building AI-native companies. It is also an investor’s view of the market.
YouTube for frontier research and deeper conversations
AI Explained — independent commentary on model releases, research papers, benchmarks and the path towards more general systems.
Two Minute Papers — visual, enthusiastic explanations of research in AI, graphics and simulation. It is an accessible doorway into a paper, not a substitute for reading the paper.
Dwarkesh Patel — long, heavily prepared interviews with researchers, lab leaders, economists and critics.
Google DeepMind — research demonstrations, explainers and video editions of its podcast. It is a primary company source, not independent coverage.
Latent Space — models, agents, tools and infrastructure for people who want to move from general interest into AI engineering.
Machine Learning Street Talk — long technical conversations about machine learning, cognition and AI philosophy. This is the deep end rather than a beginner channel.
Spotify and audio on Claude Code and agents
You Need to Learn Claude Code in 2026 — Sandeep Swadia’s episode on the move from chat-based assistance towards agentic coding.
Claude Code AI Subagents — Jake Morrison’s independently published, AI-narrated audiobook. I have included it for discovery, but its performance and cost claims still need checking against primary sources and direct experiments.
A wider map of people I follow
Some of these people appear earlier in the article. I have brought them together here by the part of the AI conversation they help me understand. This is not a ranking, and the categories inevitably overlap.
Intellectuals and lab leaders
Dario Amodei — essays on frontier capabilities, interpretability, safety and the possible scientific and economic benefits of powerful AI.
Demis Hassabis — AGI, scientific discovery, AlphaFold, robotics and DeepMind’s long-term research direction.
Sam Altman — OpenAI’s case for abundant intelligence, infrastructure and broad access; his longer writing is at samaltman.com.
Geoffrey Hinton — the foundations of deep learning and his warnings about systems becoming more intelligent than their operators; the Nobel interview is a useful introduction.
Yann LeCun — open research, world models and a prominent critique of near-term AGI and extinction predictions.
Yoshua Bengio — deep learning, AI safety, governance and research on controlling increasingly capable systems.
Fei-Fei Li — computer vision, spatial intelligence and human-centred AI; she also co-founded Stanford HAI.
Mustafa Suleyman — consumer AI, the politics of powerful technology and how Microsoft is turning models into widely used products.
Yuval Noah Harari — how AI may reshape institutions, democracy, stories, power and human agency.
Erik Brynjolfsson — productivity, augmentation and how organisations capture—or fail to capture—the value of new technology.
Daron Acemoglu — automation, labour markets, inequality and the economic choices behind how AI is deployed.
Practitioners and educators
Ethan Mollick — research-informed experiments with AI at work, in education and as a creative collaborator.
Andrew Ng — accessible courses covering foundations, business use, machine learning and agentic systems.
Dror Poleg — AI’s effects on work, organisations, cities and markets, alongside practical workshops.
Claire Vo — screen-shared demonstrations of how people use AI in real professional workflows.
Allie K. Miller — workshops and AI-first operating approaches for leaders and teams.
Rachel Woods — AI operations and turning organisational processes into documented, repeatable playbooks.
Dan Shipper — experiments, writing and products created by an organisation trying to work AI-natively.
Paul Roetzer — marketing education, organisational adoption and practical AI workshops.
Sabrina Ramonov — tactical material on Claude, agents, automation and AI-assisted building.
Elena Verna — AI through the lens of growth, product, retention and marketing.
Builders and researchers
Andrej Karpathy — language models, AI coding and clear technical education; his YouTube lectures are especially useful.
Simon Willison — hands-on model testing, AI-assisted development, security and prompt injection.
Jeff Dean — large-scale systems, research infrastructure and Google’s work on machine learning.
Nathan Lambert — open models, reinforcement learning, post-training and research culture.
swyx — AI engineering, agents, developer tools and conversations with the people building them.
Jim Fan — agents, robotics, simulation and embodied AI at NVIDIA.
Dylan Patel — chips, compute, networking, data centres and the economics beneath frontier AI.
Ilya Sutskever — foundational model research and the pursuit of safe superintelligence through SSI.
Operators, founders and investors
Tobi Lütke — Shopify’s push to make reflexive AI use part of everyday company operations.
Jack Dorsey — open-source software, decentralisation and user-controlled agents such as Goose.
Garry Tan — AI-native startups, founder behaviour and what very small teams can now build.
Aaron Levie — enterprise agents, organisational data and the future of workplace software.
Sebastian Siemiatkowski — Klarna’s highly visible experiments with AI in customer service, operations and employment.
Farhan Thawar — how Shopify deploys AI tools, shared infrastructure and MCPs across an organisation; this Shopify case study is the best starting point.
Marc Andreessen — the accelerationist case for AI, the investment landscape and technology-led abundance.
Critics and counterweights
Arvind Narayanan and Sayash Kapoor — evidence-led scepticism and the argument that AI should be understood as a powerful but socially embedded technology.
Melanie Mitchell — the gap between impressive performance, reliable reasoning and general intelligence.
Gary Marcus — persistent criticism of language-model reliability, industry claims and the route to AGI.
Timnit Gebru — training data, bias, labour, concentrated power and the social costs often excluded from capability debates; she is active on Bluesky.
Emily M. Bender — computational linguistics, the limits of language models and resistance to anthropomorphic AI claims; she co-hosts Mystery AI Hype Theater 3000.
Ed Zitron — an intentionally combative critique of AI economics, corporate incentives and industry hype, also available through the Better Offline podcast.
Curators and journalists
Dwarkesh Patel — deeply prepared, long-form interviews with frontier researchers, lab leaders, economists and critics.
Jack Clark — a detailed research and policy digest written from inside the frontier-lab world.
Timothy B. Lee — technical and policy developments explained with limited jargon.
Alberto Romero — AI, culture, philosophy and society.
Azeem Azhar — the interaction between exponential technologies, economics, institutions and geopolitics.
Ben Thompson — the business strategy, platform economics and competitive structure of the AI industry.
Rowan Cheung — a fast-moving daily overview of launches and demonstrations, useful for discovery rather than final judgment.
Kevin Roose and Casey Newton — accessible technology journalism across the Hard Fork archive. They ended the New York Times show in August 2026 and are building a new independent programme; Newton continues to publish Platformer.
On my list—not yet read or watched
This is my queue, not a list of endorsements. I have collected these sources because they may add something missing from the picture, but I have not yet examined them closely enough to place them in the main sections.
Material already collected
The Moonshots roundtable with Peter Diamandis and others, for speculative conversations about agents, mathematics, scientific discovery and AGI.
Charlie Chang’s 12 Life-Changing Ways to Use ChatGPT, a broad survey of personal and business uses that also touches on automation.
Perspectives I want to add
African AI: Masakhane on open research for African languages; Deep Learning Indaba on African AI research communities; and Lelapa AI on resource-efficient language technology designed in Africa.
Britain and Europe: Azeem Azhar’s Exponential View; the Ada Lovelace Institute; the UK AI Security Institute; and the European Commission’s AI Act material.
China and technology policy: Jordan Schneider’s ChinaTalk.
Copyright and creative work: Ed Newton-Rex’s Fairly Trained, which advocates licensed training data and certifies qualifying AI companies.
An evolving record
Looking across these sources, I do not see one settled story about AI. I see overlapping arguments about capability, productivity, creativity, scientific progress, corporate power, employment, environmental cost, safety and control.
That is why this page will never really be finished. Models will change, new voices will emerge and some of today’s commentary will age quickly. I will keep returning to the post to add people, replace outdated material and record what I am reading and learning next.
What I’m still looking for
This list is still heavy on US labs and light on almost everything else. If you follow someone strong on AI in Africa, China or Europe; on copyright and creative work; on healthcare or government use; or on projects that did not work, reply to this email or leave a comment. I read everything and will add the best suggestions to the queue above, with credit.