31 pages on the parts of AI that are not a single paper: how the model families evolved, what the benchmarks actually measure, what training and serving cost, the rules being written, and the questions nobody has settled. Each one links down into the paper summaries for the detail. Generated by scripts/build_manifest.py - do not edit by hand.
Most of these describe a moving target, so every page states when it was last reviewed.
AI has swung between two ideas - write the knowledge in by hand, or learn it from data - and this page walks through each era of that swing, with dated milestones that link to the paper summaries in this collection.
Claude is a closed-weight family sold in size tiers (Haiku, Sonnet, Opus, and since 2026 Fable and Mythos above them), shaped by a written constitution and a scaling policy that gates each release on safety testing - read a name as "tier, then generation".
DeepSeek is a Hangzhou lab whose models are defined by one habit - make the same capability cheaper to train and to serve through architecture changes (sparse experts, compressed attention, low precision), then publish the weights under MIT - so read each release as "which cost did they cut this time".
Gemini is Google's closed, natively multimodal model line that replaced PaLM and absorbed the Bard brand, sold as Pro / Flash / Flash-Lite tiers with a frequent-update Flash line, while Gemma is its smaller open-weight sibling - read a name as "generation, then tier".
OpenAI's models are one line of decoder-only transformers scaled up, then taught to follow instructions, then taught to think before answering, and since 2025 sold as a single product family with named capability tiers - read a model name as "generation, then tier".
Llama made "download a strong model and run it yourself" normal between 2023 and 2025, but its last release was Llama 4 (April 2025); in 2026 Meta moved its frontier work to a closed Muse line and returned to open weights only with the smaller Muse Glimmer - so "Meta = open" is no longer a safe assumption.
Mistral is the Paris lab that made small, efficient open-weight models its identity, wandered toward closed and restricted licences in 2024, came back to Apache 2.0 in 2025, and in 2026 sells "sovereign" European AI built on a mix of fully open and revenue-capped models - so check the licence of the exact model before you build on it.
Qwen is Alibaba's model family and the widest open-weight catalogue in the field - every size from under 1B to trillions of parameters, text, vision, audio, code and images - with the top model usually kept hosted-only and, since mid-2026, the biggest open releases carrying commercial carve-outs, so always check which Qwen and which licence you are looking at.
Agent benchmarks score whether a model can finish a multi-step job in a live environment (a desktop, a website, a customer-service conversation), and the most useful single trend line is METR's "time horizon": the length of human task an AI can complete half the time.
Benchmark scores drift upward for reasons other than capability (leaked test data, training aimed at the test, better harnesses, broken graders, selective reporting), and every public benchmark has a shelf life, so weight fresh, held-out and maintained tests above famous static ones.
Arena leaderboards rank models by which answer people prefer in blind side-by-side votes, which captures "feels better to use" better than any exam, but preference is not correctness, style sways votes, and the 2025 "Leaderboard Illusion" dispute showed how the rules of the arena shape who wins.
MMLU and its successors measure how much a model knows and how well it reasons on exam-style questions, and the useful ones in 2026 are the ones that are still hard: treat any score above about 90 percent as a solved test, not a ranking.
Maths and coding benchmarks are the easiest to grade automatically and the fastest to go stale, so trust the fresh, held-out and actively maintained ones (competition problems released after training, FrontierMath's holdout, versioned agent suites) over the famous static ones.
A context window is how many tokens a model can take in at once, and the number on the spec sheet is an upper bound on what fits, not a promise of how well the model uses it, so treat long context as a tool to test rather than a replacement for choosing what goes in.
"Open" in AI is a spectrum, from API-only access, through downloadable weights under restrictive or permissive licences, to fully open source with code and data, so ask exactly what is released and under which licence before deciding what you can build, and read the safety debate as a disagreement about marginal risk rather than about openness in general.
"Reasoning" or "thinking" models are language models trained with reinforcement learning to write out a long chain of working before they answer, which trades extra tokens (time and money) for accuracy on hard problems, so you should turn the thinking up for maths, code and multi-step tasks and down for everything else.
A language model only outputs a probability for every possible next token; the decoding strategy (greedy, temperature, top-k, top-p, min-p, beam search, constrained decoding) is the separate rule that turns those probabilities into actual text, and it changes the output as much as many model upgrades do.
When a language model generates text it keeps a running memory of every earlier token (the key-value, or KV, cache), and the size of that memory, not the model's weights, is often what limits context length, batch size and serving cost.
Language models read and write in tokens (chunks of text, usually pieces of words), not words or letters, and that one design choice explains a surprising share of their costs, limits and odd mistakes.
The price of a given level of AI capability has been falling by roughly 10x or more per year, but your actual bill depends on which tokens you buy (input, output, cached, reasoning), so read a price sheet as four prices, not one.
Frontier AI runs on a handful of accelerator families (NVIDIA GPUs, Google TPUs, AWS Trainium, AMD Instinct, and a few inference specialists), and the real limits are memory bandwidth, interconnect and chip supply, not raw arithmetic.
Training compute can be estimated with one formula (about 6 x parameters x tokens), and by that measure the cost of a frontier training run has grown from about $2 million (GPT-3, 2020) to roughly $500 million (Grok 4, 2025), so treat any single headline cost, including DeepSeek's "$5.6 million", as one slice of a much larger bill.
The big labs publish voluntary "if-then" policies - if a model reaches a dangerous capability level, then specific safeguards must be in place before it is deployed - and the live debates are whether those policies bind anyone, and whether they are being weakened as competition heats up.
The EU AI Act sorts AI by risk, bans a short list of practices outright, puts heavy duties on "high-risk" uses and on the biggest general-purpose models, and phases in over 2025-2028 - with the high-risk deadlines pushed back by the 2026 "digital omnibus".
The US has no comprehensive federal AI law; policy is set by presidential executive orders that have swung sharply between administrations, by a growing set of state laws, and by an unresolved fight over whether Washington should override those state laws.
As of September 2026, AI agents connect to the world through a small set of layered standards (a model API to talk to the model, MCP to reach tools and data, A2A to reach other agents, and Agent Skills plus AGENTS.md to carry instructions), and most of them now sit under neutral Linux Foundation governance.
As of September 2026 the frontier is set by a handful of US labs that keep their best weights closed (OpenAI, Anthropic, Google DeepMind, xAI), while the strongest downloadable models come mostly from Chinese labs (DeepSeek, Alibaba Qwen, Moonshot, Z.ai) and Europe's Mistral - so when you pick a model, the lab's weights policy and licence matter as much as its benchmark scores.
Open AI tooling sorts into five layers (get a model, train it, serve it at scale, run it locally, build an app on it), and knowing which layer a tool lives in tells you what it is for and what it replaces.
Whether language models "really reason" depends heavily on what you mean by reasoning; the evidence shows real multi-step internal computation and striking problem-solving, alongside brittleness that looks like pattern-matching - and both camps can point to strong results.
"Scaling" used to mean one thing - bigger models on more text - and that recipe is running into data and cost limits; the open question is whether newer ways of spending compute (reinforcement learning and thinking longer at inference) keep progress going, and for how long.
Training models on model-generated data is now standard practice and often works very well - when it is filtered, verified, or mixed with real data - but recursively training on unfiltered model output, in place of real data, degrades models over generations.