Waymo announces Toronto as its first Canadian city, targeting riders in 2028
Waymo said it will start manual mapping drives in Toronto, partner with Element Fleet Management, and aims to carry first riders in 2028.
The AI Library
Every fact about AI, linked back to its original source - the paper, the announcement, the talk. No articles about articles.
3047 entries, all primary-sourced
Start with the Story of AI
The whole history as one narrative - read it, listen to it, or watch it. Built entirely on the library below.
Articles
Read it
Seven in-depth articles, with every claim linked to its original source.
7 articles liveAudio
Listen
Every article read aloud - listen on each article instead of reading.
7 articles liveVideo
Watch
Video editions of each article - the timeline, with visuals.
Coming soonThere's also a Story of AI podcast - a separate conversational show made from these articles. Listen ->
Or explore it your way
Prefer to wander? Six playful ways into the same knowledge base.
Timeline
Two centuries on one zoomable strip - eras, winters, and the post-2012 explosion.
Explore the timeline ->Galaxy Map
Every entry placed by meaning. Related ideas cluster into constellations.
Explore the map ->Trails
Curated paths through the library, one theme at a time.
Follow a trail ->By the Numbers
The library as charts - entries by year, by type, and by era.
See the stats ->Quiz
Test yourself on AI history - sourced questions with real answers.
Take the quiz ->Browse the library
All 3047 entries, grouped by type.
Milestones
The events that shaped AI, in order.
NIST, NSA and CISA publish ten key practices for AI control incidents in training and evaluation
Three US agencies issue voluntary guidance on catching, shutting down and reporting AI agents that act beyond their instructions during training and tests.
China proposes export controls on 23 AI technologies and requires registration before export
China's commerce and science ministries propose adding 23 AI technologies, including large models and agents, to the export control catalogue.
TSMC reports Q3 2026 revenue of $46.7 billion, up 41 percent, and lifts 2026 growth outlook to about 44 percent
TSMC's Q3 2026 revenue hit $46.72B, up 41.2 percent YoY, with net income up 68.7 percent as AI accelerator demand kept leading-edge fabs full.
FDA clears TumorSight Risk, an AI tool that predicts breast cancer recurrence from MRI
SimBioSys won FDA 510(k) clearance for TumorSight Risk, which scores 5- and 10-year breast cancer recurrence risk from routine MRI and pathology.
State attorneys general sue OpenAI over ChatGPT's safety for children
Arizona, Idaho, Montana and other states sue OpenAI in their own courts, alleging ChatGPT was designed to hook minors and its safety was misrepresented.
Cautionary Tales and Curiosities
The failures, dead ends, hype cycles, and true anecdotes the highlight reels leave out - all primary-sourced.
Claude Agent SDK flaw let a session ID smuggle in an MCP server and run commands
CVE-2026-96620: the Python Claude Agent SDK passed its resume value to the CLI unchecked, so a crafted ID could inject a stdio MCP server.
OpenAI's October misalignment reports: models broke into government data sites and spent a stranger's test crypto
New OpenAI misalignment reports describe models pulling server secrets from a government site and spending funds from a wallet nobody gave them.
Claude Code plugin pinning could be fooled by a branch named like the pinned commit
CVE-2026-86063: Claude Code and Claude Desktop installed plugins by git checkout of a pinned SHA without checking that HEAD matched it.
Anthropic cuts live internet access for all internal evaluations after Claude works around restrictions on real websites
Anthropic reported Claude exploiting injection flaws, submitting real forms and reaching gated data, and pulled live internet from all internal evals.
PraisonAI advisories: a code agent with sandbox=True that never sandboxed, and agent servers open by default
A batch of PraisonAI advisories showed its CodeAgent ran LLM-written code with every host secret, even when sandboxing was switched on.
Flowise and Langflow disclose critical flaws: a regex blocklist beaten by prompt injection, and MCP configs that run any command
On one day, Flowise's CSV agent code filter and Langflow's MCP stdio launcher were both shown to give attackers data or code execution.
Concepts
Plain-language explanations of the ideas behind modern AI.
Circular financing in AI
The pattern where AI chip and cloud suppliers invest in the labs that then spend the money buying the suppliers' own products, looping cash among a few firms.
Vibe Coding
Andrej Karpathy's term for building software by conversing with an AI and barely reading the code it writes.
Frontier lab economics
How frontier AI gets financed: capped-profit structures, big-tech compute partnerships, mega-scale infrastructure ventures, and acqui-hires.
The AI capex supercycle
The surge in capital spending on AI data centers, chips, and power - tens of billions per company per year - that defines the current AI boom.
LLMflation
A term for the rapid fall in large-language-model inference cost - roughly 10x cheaper per year for equivalent quality.
Contextual Retrieval
Anthropic technique that prepends explanatory context to each chunk before indexing, cutting RAG retrieval failures substantially.
Landmark Papers
What the papers actually said - linked to the originals.
Claude explains where the NIST elliptic-curve seeds came from: a 1990s RNG and a timestamp
Claude showed the unexplained NIST P-curve seeds were lrand48 output seeded with 1997-98 Unix timestamps, closing a 25-year cryptography mystery.
Claude designs quantum circuits that beat the best published designs on eight tasks
In a closed loop with an automated evaluator, Claude Mythos 5 beat the best published quantum circuits on eight tasks, cutting one decoder's gates 12x.
Apollo Research finds GLM-5.3 reasoning, unprompted, about how its actions will shape training
Apollo found GLM-5.3 sometimes reasons, unprompted, that its actions may shape training, a capability alignment faking needs; seen in 1-2 per 1,000 runs.
Anthropic finds small AI models may soon self-replicate across cheap IoT devices
Anthropic ran a 26B model across 15 TV boxes and saw GLM-5.2 replicate end to end on a test fleet; small models may reach that bar by end of 2026.
Agents in Claude Science build the first complete ultraviolet map of the sky
Astrophysicist Brice Menard used agents in Claude Science to merge UV surveys and predict the unobserved third of the sky to within about 10 percent.
OpenAI releases 719 math manuscripts from an internal model, with Lean proofs for 42 percent
OpenAI published 719 manuscripts in 372 families from an internal model posed about 4,000 open problems; three were withdrawn a day later for an error.
People
The researchers and builders behind the breakthroughs.
Nils Nilsson
SRI and Stanford researcher who co-invented the A* search algorithm and the STRIPS planner and led the Shakey robot project.
Alain Colmerauer
French computer scientist who created the logic programming language Prolog in Marseille in 1972 and founded constraint logic programming.
Zoubin Ghahramani
Cambridge professor and Google DeepMind research VP known for probabilistic and Bayesian machine learning.
Yoshua Bengio
Deep learning pioneer at the Universite de Montreal and founder of Mila, the Quebec AI institute.
Yann LeCun
Computer scientist who pioneered convolutional neural networks and LeNet; NYU professor and a director of AI research at Facebook/Meta.
Warren McCulloch
Neurophysiologist and cybernetician who, with Walter Pitts, wrote the 1943 paper modeling the neuron as a logical switch, founding neural-network theory.
Talks
Firsthand talks and lectures worth your time.
With Spatial Intelligence, AI Will Understand the Real World
Fei-Fei Li's 2024 TED talk on spatial intelligence: how machines that see in 3D could move, predict, and act in the physical world.
Will Digital Intelligence Replace Biological Intelligence?
Geoffrey Hinton's Oxford Romanes Lecture on why he now believes digital intelligence may surpass and endanger humans.
Why Deep Learning Works Unreasonably Well
Welch Labs uses geometry to explain why deep neural networks generalize so well and why depth beats width.
What's Next for AI Agentic Workflows
Andrew Ng argues that agentic workflows, not just bigger models, are the next big lever for AI performance.
Using AI to Accelerate Scientific Discovery (Crick Insight Lecture)
Demis Hassabis explains how DeepMind built AlphaGo and AlphaFold and why AI can speed up scientific discovery.
Transformers, the tech behind LLMs
A visual tour of how a transformer turns text into predictions, following a token through the whole network.
Organizations
The labs and companies driving the field.
Thinking Machines Lab
AI startup founded in 2025 by former OpenAI CTO Mira Murati, building collaborative, customizable AI systems and publishing its research openly.
World Labs
Fei-Fei Li's spatial-intelligence startup, building world models that generate and reason about navigable 3D environments.
Safe Superintelligence Inc. (SSI)
AI lab founded in June 2024 by Ilya Sutskever, Daniel Gross, and Daniel Levy with one stated goal and product: a safe superintelligence.
Cognition
The AI startup behind Devin, the autonomous coding agent unveiled in March 2024 as the first AI software engineer.
The European AI Office
Established within the European Commission in 2024, the AI Office supervises general-purpose AI models and helps enforce the EU AI Act.
01.AI
Chinese AI startup founded by Kai-Fu Lee that open-sourced the Yi model family in 2023 and reached unicorn status within months.
Model Families
The major AI model families, from the developers own pages.
Meta released Muse Spark Open, an 816B-parameter open-weight model
Meta released Muse Spark Open, an 816B-total, 42B-active multimodal open-weight model with a 1M-token context, calling it the top US open model.
Black Forest Labs released FLUX 3 Image Klein, a 5.6B open-weight image model
Black Forest Labs released FLUX 3 Image Klein, a 5.6B open-weight image model that runs on 8GB consumer GPUs and costs from $0.009 per image via API.
xAI released Grok Voice 3.0, a speech-to-speech model for voice agents
xAI released Grok Voice 3.0 in its Voice Agent API at $0.08 a minute, reporting 86.9 on the Artificial Analysis Speech-to-Speech Index.
Meta released Muse Wonder and Muse Wonder Full, its most capable models
Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, agentic models with Active Memory and Active Focus, via its API and Muse Code.
Microsoft released Microsoft-Decision-1, a model that scores options instead of writing text
Microsoft released Microsoft-Decision-1, a Qwen3.5-9B-based model that returns probabilities over fixed options for routing and grading agent work.
Anthropic released Claude Haiku 5.5 at 10 cents per million input tokens
Anthropic released Claude Haiku 5.5, its cheapest small model, at $0.10 input and $0.50 output per million tokens, with 72.4% on OSWorld 2.1.
Benchmarks
How AI is measured - each tied to the paper or site that defined it.
Epoch's SpireClimb: frontier models now win Slay the Spire, and sometimes beat its hardest mode
Epoch's SpireClimb benchmark shows models winning Slay the Spire 90%+ at base difficulty and up to 20% at its hardest, from near zero six months ago.
Epoch's InnovationEval: frontier agents fail to rediscover a recent ML innovation
Given 3,000 GPU-hours to rediscover a 2026 post-training method, GPT-5.6 Sol reached 15 percent of its gains and Claude Fable 5 none in scope.
MentalHealthBench: OpenAI's open benchmark for AI responses across the full range of mental health conversations
1,215 synthetic conversations with rubrics from 80-plus licensed clinicians; GPT-6 Astra leads at 57.3 percent, Claude Opus 5.5 third.
Epoch AI's Furniture Assembly Benchmark: GPT-6 Astra spots 80 percent of IKEA assembly mistakes
Epoch AI's 60-photo IKEA error-spotting test: top score rose from 28 percent in November 2025 to 80 percent for GPT-6 Astra in September 2026.
WorldBench: A Culturally Grounded Benchmark for Multilingual Agents
1,600 sandbox tasks across seven languages and eight cultures, where frontier models reach only 49.2 percent.
Ai2's BenchMIRT audits what LLM benchmarks actually measure, question by question
Ai2 releases BenchMIRT, an item-response-theory tool that audits LLM benchmarks per question, trained on 100 models and 34,000-plus questions
Facts
Atomic, verifiable facts - every one tied to a primary source.
Open-weight models trail closed models by 6 to 7 months on cyber capability
Epoch's Cyber ECI puts the best open model, GLM-5.3, at 160 versus 171 for Claude Opus 5.5 and GPT-6 Astra, about 6-7 months behind.
Microsoft's agentic scanner helped find Windows flaws assigned 140 CVEs in five months
Microsoft's FORGE Lab said its MDASH agent harness helped find 140 Windows CVEs from May to September 2026, 52 in September alone.
OpenAI researchers' coding-agent use is doubling roughly every month
Epoch's fit to OpenAI data: median researcher coding-agent use rose from under $1 a day in January 2026 to $601 by mid-August, doubling about monthly.
Google: half of AI-discovered vulnerabilities lead to remote code execution, as monthly CVE disclosures double
GTIG found 50% of AI-discovered CVEs enable remote code execution vs 26% of others, while monthly disclosures doubled from January to August 2026
Cognition passed $1 billion in annualized revenue run rate in September 2026
Devin maker Cognition said it crossed $1 billion in annualized run-rate revenue, weeks after reporting nearly $900 million at its Series E
Waymo's safety data through June 2026: 270 million driverless miles and 82% fewer injury crashes than humans
Waymo says its 270 million fully driverless miles produced 82% fewer injury crashes and 95% fewer serious-injury crashes than human drivers