What it costs and where it works

S01 Epoch AI · continuous · Dashboards

Training compute, hardware price-performance, algorithmic efficiency, inference cost and data supply, as live charts. Two facts do most of the work elsewhere: frontier compute has grown several-fold annually for a decade, and the cost of any given capability falls fast once it exists.

What to do Two curves on the first pass and ignore the rest: training compute over time, and inference cost for a fixed capability. Note the doubling time on the first and the decline rate on the second. Those two numbers quietly settle most later arguments about cost and timelines.

Also in Start here.

15 min epoch.ai · free

S02 Dylan Patel et al. · continuous · Industry analysis

The reference on chips, fabs, interconnect, data centre buildout and the actual supply chain. Technical, sometimes paywalled, and where the industry reads to find out what the industry is doing.

What to do Do not subscribe yet. Read three free posts on data-centre buildout and interconnect, enough to learn the vocabulary — HBM, networking, power density. Then treat it as the place you go when a supply-chain claim needs checking, rather than a weekly habit you will quietly abandon.

Ongoing semianalysis.com · free and paid

S03 International Energy Agency · 2025 · Report

The authoritative treatment of data centre electricity demand, grid implications and regional concentration, from the body that does this for every other energy question. Replaces both the alarmist and the dismissive version of the energy argument with scenarios.

What to do Read the executive summary and the scenario ranges, and note the range rather than a single figure. Then adopt the habit for life: whenever you next meet an AI energy number in the press, check which scenario and which year it came from. It usually came from the top of a range, undated.

1 hr 30 min iea.org · free

S04 Lawrence Berkeley National Laboratory · 2024 · Report

The US-specific accounting beneath the data-centre energy argument: historical electricity use, facility characteristics and scenario ranges through 2028. Pair it with the IEA's global outlook. One gives you the worldwide system; this gives you a detailed national inventory with assumptions visible enough to argue with.

What to do Read this only after the IEA report, and read it for the assumptions rather than the totals. Set the national inventory against the global outlook and note where they disagree. That disagreement is the honest uncertainty, and it is more useful than either number alone.

45 min eta.lbl.gov · free

S05 Brynjolfsson, Li & Raymond · 2023 · Field study

Five thousand customer support agents given an LLM assistant. Productivity rose about 14% on average, concentrated almost entirely among the least experienced workers, with little effect on the most skilled. The most-cited empirical result on AI and labour, and the distributional finding is the part that keeps getting dropped.

What to do Read the abstract and the distributional results. The headline is 14%; the finding is that almost all of it went to the least experienced workers. Carry the second number, because the first one travels alone everywhere else and changes the policy conclusion completely.

NBER working paper; the abstract page links the PDF and ungated versions circulate.

1 hr NBER w31161 · free

S07 International Labour Organization · 2025 · Working paper

Maps generative-AI exposure across occupations and countries using a task-level index refined with worker input. Its central finding is more transformation than outright replacement, with exposure distributed unevenly by occupation, income level and gender. Useful ballast against both the no-jobs and no-change versions of the labour story.

What to do Read the methodology for how exposure is scored, then look up your own occupation and one occupation you employ. Take transformation-rather-than-replacement as the finding, and note specifically how unevenly it lands by income level and by gender. That unevenness is the policy story.

35 min ilo.org · free

S09 METR · 2025 · Randomised trial

Experienced open-source developers working in their own repositories were about 19% slower with AI assistance, while believing themselves roughly 20% faster. The most important negative result in the field, precisely because it is a randomised trial in a domain where everyone was certain of the answer.

What to do Read the whole thing, methodology included, because you are going to want to disbelieve it. Extract both numbers: 19% slower, and believed themselves 20% faster. Then treat your own sense of AI-assisted speed as unreliable evidence, permanently. That is the finding, and it is about you.

45 min metr.org · free

S10 Dell'Acqua, McFowland, Mollick et al. · 2023 · Field experiment

Seven hundred and fifty-eight BCG consultants, randomised. Large gains on tasks inside the model's capability frontier, and measurably worse performance on a task just outside it, because people could not tell which was which. The single most useful concept for anyone deploying this in an organisation.

What to do Read the design and the outside-the-frontier result. Then try it against three tasks in your work you believe are inside the frontier and one you suspect is just outside. Being unable to tell is the point, so revisit the prediction after a month of real use and see how you did.

HBS working paper 24-013; the PDF is linked from the faculty item page and mirrored on SSRN.

1 hr Harvard Business School · free

S11 OpenAI · 2025 · Evaluation + paper

An attempt to measure economically valuable work instead of exam-shaped cleverness: realistic deliverables across 44 knowledge-work occupations, judged against expert outputs. The limitations matter as much as the scores, because the real world cruelly insists on iteration, ambiguity and people changing their minds after lunch.

What to do Skim the occupation list and one example deliverable, then read the limitations twice as carefully as the results. Note one thing about your own job that this eval structurally cannot capture. That sentence is a useful boundary on what the result means.

Also in Start here, How it works.

20 min openai.com · free

S12 Google DeepMind · 2020– · Research programme

Protein structure prediction at experimental accuracy, a database covering essentially every known protein, and a Nobel Prize in Chemistry in 2024. The strongest evidence that machine learning produces genuine scientific results, and worth understanding for exactly how narrow and well-posed the problem was.

What to do Read the overview and enough about CASP scoring to understand how the win was judged. Then be precise about the boundary: note exactly how well-posed the problem was — fixed input, measurable ground truth, decades of curated data. Check every "AI will do this for science" claim against those three conditions.

1 hr deepmind.google · free

S13 Anthropic · continuous · Research hub

Anthropic's running attempt to measure how people actually use Claude at work and what that implies for tasks, wages and the economy. It is company data, so keep your skepticism plugged in, but it is still unusually direct evidence about usage rather than vibes wearing a necktie.

What to do Read it with the conflict of interest stated out loud: a vendor measuring its own usage. Extract the task mix, ignore the framing, then check it against the ChatGPT usage paper below. Where two differently-motivated datasets agree, believe it more. Where they diverge, that is your research question.

20 min anthropic.com · free

S14 Chatterji, Cunningham, Deming et al. · 2025 · NBER working paper

A large privacy-preserving study of consumer ChatGPT use, including work versus non-work use, topic mix, adoption patterns and the surprising amount of value created through advice, information and writing rather than pure programming. Consumer usage is not the economy, but it is a real window, and windows beat fog.

What to do Read the topic mix and the work-versus-non-work split. Then use it correctively: note how much of the measured value is advice, information and writing rather than code, and check that against your own assumption about what these tools are for. Most technical readers guess wrong.

25 min nber.org · free

S15 Stanford HAI · annual · Report

The field's annual census: capability, investment, adoption, cost, policy and public opinion, all sourced. Skim the top takeaways on release, then use it year-round as the reference whenever someone quotes a number at you. The adoption-versus-measured-value gap in the economy chapter is the most interesting figure in it and the one most often skipped.

What to do Read the top-level takeaways, then go straight to the economy chapter and find the gap between reported adoption and measured value. Note that gap as a number. Come back to the rest only when you need a citation — never cover to cover. Skim the takeaways again each year on release.

Also in Start here, Keep current.

20 min–2 hr hai.stanford.edu · free

S16 UNESCO · 2023 · Guidance

A human-centred policy guide to privacy, age limits, institutional validation and the design of generative-AI use in education and research. It avoids both ban-it and sprinkle-chatbots-everywhere. The useful question running through it is whether a use protects human agency and educational purpose rather than merely producing an answer faster.

What to do Read it for the test rather than the rules: does this use protect human agency and educational purpose, or does it only produce the answer faster. Then use that test to reconsider one AI use you have already adopted.

45 min unesdoc.unesco.org · free

S17 World Health Organization · 2025 · Guidance

WHO's governance guidance for large multimodal models in health care, public health, research and drug development. The document treats capability as only the beginning: evidence quality, bias, automation bias, privacy, accountability and post-deployment monitoring matter more when a fluent error can become a clinical decision.

What to do Read the automation-bias and post-deployment-monitoring sections; skip the general ethics framing if time is short. The transferable lesson is not medical — a fluent error inside a high-stakes workflow is worse than an obvious one — so read it even if you will never touch health data.

1 hr who.int · free