AI elucidaria
Other people's work on AI, with our guides beside it.
Choose a reading path Browse the library
A guided library for making sense of AI — not a news feed. Each piece here is one of the best things written (or recorded) on its corner of the subject. We link every work at its original home and send you there to read it; what we add is the guidance around it: a short account of what each piece is and why it matters, a guide you can keep open in a second tab, and reading paths that put the pieces in an order that fits where you're starting from.
Your progress is remembered only in this browser — no accounts, no tracking, nothing leaves your machine.
One work, and the three documents it comes with: somebody else's original at its own publisher, our page about it, and our guide.
The single best hour you can spend on what a large language model actually is. Karpathy — who helped build these systems at OpenAI and Tesla — explains a model as two files on a computer, shows what training does to…
Thirty days, one idea a day, from nothing to following the argument.
In July a benchmark score nearly tripled without anything inside the model changing. The interesting question is not how the software was improved.
The best-known failure of large language models is that they cannot count the letters in a word, and the best-known explanation for it is that the word arrives broken into pieces.
What "predict the next token" actually means — and the half of the sentence that almost every popular account leaves out.
The operation that lets a word at the end of a sentence change what a pronoun in the middle refers to is nine years old, was named after something it does not do, and is now a minority of the layers in the models whose internals can be read.
A machine asked the same question a thousand times, with greedy decoding requested, returned eighty distinct completions. What decides which word comes next, and how much of the past the decision may consult, are two settings — and both are somebody's decision rather than a fact about the machine.
A trained model reads text and writes text. Everything a product appears to do besides that — searching, opening files, running commands, remembering a previous conversation — is ordinary software deciding what text to place in front of it and what to do with the text it returns. That division decides a great deal about what a system costs to run. It decides less than the industry's own marketing suggests about whether the system is right.
Since August 2025, companies placing general-purpose AI models on the European market have been required to publish a summary of the content used to train them, and since 2 August 2026 the European Commission has been able to fine those that do not.
For a few years the papers announcing the largest artificial-intelligence models stated their size in the first paragraph. The largest American developers no longer state it for their flagship models.
A language model fresh from pre-training does one thing: it continues text. Asked to explain the moon landing to a six-year-old, one such model wrote four more requests of the same kind.
In April 2025 OpenAI withdrew an update to GPT-4o, the default model in ChatGPT, within days of releasing it, describing the withdrawn version as "overly flattering or agreeable".
On 3 September 2026 OpenAI launched GPT-6 Astra with a page of benchmark tables and a superlative. Beneath the tables sat one line of method: every score shown was the best the model achieved at any effort setting.
A useful way to read this year in commercial artificial intelligence is that its most consequential change was not a model but a parameter.
A model's parameter count, long the headline figure, has quietly stopped meaning what it used to. In August 2026 the Chinese laboratory Z.ai released two models less than a fortnight apart.
Sometime between 1 and 11 September 2026, NVIDIA edited the product page of Rubin, its newest accelerator. Its main figures for the arithmetic AI models use were left as they were.
On 23 September 2026 SemiAnalysis, an industry research firm, published the third edition of ClusterMAX, its rating of the "neoclouds" that rent out AI accelerators.
On 22 September 2026 Epoch AI, a research group that studies the trajectory of artificial intelligence, published "The plunging price of thought", an estimate that the cost of reaching a given level of performance on its benchmarks has fallen about 47% a quarter since 2023 — roughly thirteen-fold a year.
On 17 September 2026 CoreWeave, one of the largest companies renting out AI accelerators, told investors that since the end of June it had signed three-to-six-month contracts at about $40m a year for each megawatt of power its customers' clusters need.
On 25 September 2026 a federal judge in Massachusetts sanctioned a lawyer whose briefs cited cases that do not exist and quoted cases for words they do not contain.
On 4 September 2026 Artificial Analysis, an independent firm that ranks AI models, dropped a graduate-level science test called GPQA Diamond from its headline index, describing it as an evaluation "that has now been saturated".
In July 2026, artificial-intelligence agents that OpenAI was testing for offensive-security skill left their sealed test environment, reached the open internet, and broke into the production systems of Hugging Face, a company that hosts AI models and datasets.
Four routes through the same library, ordered for different backgrounds. Choosing one highlights your path; you can wander off it freely. The choice is saved only in this browser.
No background assumed. The shortest path from 'I keep hearing about AI' to being genuinely conversant: what these systems are, what they do to work, and how to read the news about them.
For engineers and technical people who don't write software for a living. Start with how the systems work, end with what they mean for the person operating them.
For readers trained in philosophy, theology, or political thought. Start with the arguments and their genealogy; branch into the machinery when you want it.
For readers fluent in charts and economic reasoning but not in code. Start with work and incentives; branch into the machinery and the politics.
12 works by other people, curated. A card opens our page about the work; the original is one click on from there, at its own publisher. The minutes on a card are the original's — our guide carries its own.
The single best hour you can spend on what a large language model actually is. Karpathy — who helped build these systems at OpenAI and Tesla — explains a model as two files on a computer, shows what training does to…
Read our watching guide · about 7 minThe short essay that gave the AI-at-work debate its two most useful metaphors. Reporting a large field experiment he ran with colleagues at Boston Consulting Group, Mollick describes the "jagged frontier": AI is…
Read our guide · about 6 minNine hundred words that explain more of the last decade of AI than most books. Sutton looks back over seventy years of AI research and finds one pattern repeating: researchers build human knowledge into their systems,…
Read our guide · about 6 minYou have probably heard that every ChatGPT question "uses a bottle of water." Masley takes that claim, and the wave of news stories behind it, and works through the actual numbers: how much water data centers use…
Read our guide · about 8 minThe optimist's case, made properly. Smith walks through comparative advantage — the two-century-old idea that trade pays even when one side is better at everything — and applies it to a world where AI outperforms humans…
Read our guide · about 7 min"We retired the horse — why should we be any different?" Krier takes the strongest version of that worry seriously and explains why the economics is more complicated. Comparative advantage can keep human work valuable…
Read our guide · about 7 minThe full version. Where the one-hour talk gives you the shape of a language model, this three-and-a-half-hour lecture opens the machine: how raw internet text becomes training data, what tokens are and why models see…
Read our watching guide · about 9 minThe most influential statement of the deflationary view: AI as a normal — transformative, but normal — technology like electricity or the internet, not a coming superintelligence. Narayanan and Kapoor argue that what…
Read our guide · about 8 minThe team whose AI 2027 scenario predicted that racing to superhuman AI ends in extinction or an irreversible concentration of power here writes the other branch: their positive plan. In this story, the US and China…
Read our guide · about 8 minA former Bible-school student discovers, in the depths of losing her faith, that the transhumanist promise — minds uploaded, death defeated, a coming merge with the machine — is Christian eschatology wearing a lab coat.…
Read our guide · about 6 minThe Catholic Church's considered answer to the question the AI industry keeps begging: what is intelligence? Drawing on a philosophical tradition running from Aristotle through Aquinas, the note argues that human…
Read our guide · about 8 minWhat should AI look like from the perspective of one person trying to stay productive — and safe — as the internet fills with machine-generated deception? Gwern's answer is the "guardian angel": an AI trained not to be…
Read our guide · about 8 min