ChatGPT launches
OpenAI releases ChatGPT as a research preview — a conversational model the general public could simply talk to. The chatbot era starts here.
Step Two — Weigh
Analysis
A compressed timeline from ChatGPT to Agent2Agent — the models that changed what software can do, the protocols that let agents work together, and the laws racing to catch up. Three tracks, one page, no hype: what happened, when, and where it is going.
Analysis does not begin from nothing. It begins under Standards: those terms decide how evidence is weighed and what a claim must survive. What survives becomes concrete results and recorded learning in Field Notes; the strongest of that learning is promoted into Research.
THE AGENT STACK
INSIDE A TRANSFORMER
Everything else on this page is about what AI systems do. This section opens one up. A transformer reads text as tokens, turns each token into a vector, and passes those vectors through twelve layers of attention and arithmetic before it bets on the next token. The explorer below walks that pipeline on seven short sentences, stage by stage.
1 · Tokens and embeddings
Tokenization cuts text into pieces from a fixed vocabulary of 50,257 — common words whole, rare words in fragments. Each piece is looked up in a table of 768-number vectors, its embedding; a second table adds a vector for the piece's position. What travels through the model is the sum: meaning plus place. Pick a sentence, then a token. The scatter is a PCA projection of those 768 numbers down to two dimensions — distances are indicative, not exact — and the strip shows the first 48 numbers of each token's vector, exactly as recorded.
2 · Attention
Attention is how tokens talk to each other. Every token emits a query, a key, and a value; its next representation is a weighted blend of the values of the tokens it attends to, and the weights are scaled dot products of queries against keys, normalized to sum to one. GPT-2 runs twelve heads of this in parallel, twelve layers deep. Rows are the listening token, columns the listened-to token. The empty upper triangle is the causal mask: the model is never allowed to peek at the future it is trying to predict. Touch a cell to read the exact weight.
Layer 1 of 12
Layer 0 is the embedding itself — attention starts at layer 1.
3 · What each layer thinks comes next
Apply the model's output head to every layer's representation of the final token — a probe called the logit lens — and a draft answer forms early, then sharpens. The probe is not a component; the model itself only ever reads the last layer. Below the strip, the heat row shows each token's residual-stream norm per layer: watch the stream grow as layer after layer writes into it.
4 · The output, and a dial
At the end, the final representation is scored against all 50,257 token embeddings and a softmax turns scores into probabilities. Temperature rescales the scores first: below 1 the favourite pulls away, above 1 the field flattens. These bars are recomputed in your browser from the model's recorded top-50 scores; the other 50,207 tokens share one combined bucket, so the probabilities stay honest.
1.00
Honesty note: every number in this section is a real output of GPT-2 (small), recorded by Qompass for these seven sentences — attention weights, embeddings, and logits alike. Nothing is simulated, and the model does not run in your browser: the data file is static, the arithmetic happens on your device, and nothing is sent anywhere.
The idea of an explorable transformer — the pipeline laid out stage by stage, attention as a matrix you can interrogate — owes a debt to Georgia Tech's Transformer Explainer, which runs a live model in the browser. This is an independent implementation with recorded activations, built for this page.
THE TIMELINE
2022
2023
GPT-4
OpenAI releases GPT-4, a multimodal model accepting text and image input, first to ChatGPT Plus subscribers and an API waitlist.
Claude 1
The same day, Anthropic introduces Claude and Claude Instant via API and chat — a frontier competitor built with Constitutional AI and positioned on safety.
Llama 2 — open weights go mainstream
Meta releases the Llama 2 family (7B, 13B, 70B) with weights available for research and commercial use. A capable open-weights model is suddenly something any developer can run.
China's generative-AI measures take effect
The Interim Measures for the Management of Generative AI Services — issued jointly by seven bodies led by the Cyberspace Administration of China and promulgated Jul 13 — take effect: the first national rules for public-facing generative AI services.
EO 14110 — the US AI executive order
President Biden signs Executive Order 14110, “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” — a whole-of-government order including reporting duties for powerful dual-use foundation models (88 FR 75191).
Bletchley Park AI Safety Summit
The UK hosts the first global AI Safety Summit; 28 countries plus the EU agree the Bletchley Declaration — a shared account of frontier-AI opportunity and risk, and of international safety research.
The OpenAI board crisis
OpenAI's nonprofit board removes CEO Sam Altman on Nov 17; five days of employee and investor pressure later, he is reinstated on Nov 22 with a reconstituted board — the governance tension inside the frontier labs, in public.
Gemini 1.0
Google and Google DeepMind announce Gemini 1.0 — Ultra, Pro, and Nano — a natively multimodal family; Gemini Pro reaches Bard the same day.
2024
Claude 3
Anthropic launches the Claude 3 family — Haiku, Sonnet, Opus — its first multimodal generation, available on claude.ai and the API at launch.
GPT-4o
OpenAI announces GPT-4o (“o” for omni): one natively multimodal flagship across text, audio, and vision, rolled out to free and paid ChatGPT users alike.
Colorado AI Act — first US state AI law
Gov. Jared Polis signs SB24-205, “Consumer Protections in Interactions with Artificial Intelligence Systems”: the first comprehensive US state AI law, imposing reasonable-care, documentation, and disclosure duties around high-risk AI in consequential decisions. (Operative date later delayed to Jun 30, 2026.)
EU AI Act enters into force
Regulation (EU) 2024/1689 — published in the Official Journal Jul 12 — enters into force, laying down risk-based rules for AI systems in the Union. Most obligations apply from Aug 2, 2026; the phase-in is the story (see the bar below).
o1-preview — the reasoning era
OpenAI releases o1-preview and o1-mini, the first o-series models, designed to spend more time thinking before answering. Test-time compute becomes a product.
Anthropic computer use
Alongside an upgraded Claude 3.5 Sonnet, Anthropic puts computer use into public beta: Claude can look at a screen, move a cursor, click, and type. The agent era gets hands.
MCP spec first published (v2024-11-05)
JSON-RPC 2.0 protocol with hosts, clients, and servers. Server features: tools, resources, prompts. Client features: roots, sampling. Transports: stdio and HTTP+SSE.
Anthropic open-sources MCP
The spec plus Python/TypeScript SDKs, local server support in Claude Desktop, and an open-source server repo (Google Drive, Slack, GitHub, Git, Postgres, Puppeteer). Created by David Soria Parra and Justin Spahr-Summers. Early adopters: Block, Apollo.
[secondary] date per corroborating sources
Anthropic publishes “Building effective agents”
The workflows-vs-agents taxonomy (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, autonomous agents). Core advice: simplest solution first, sandboxed testing, human feedback at checkpoints.
2025
OpenAI publishes “A Practical Guide to Building Agents”
Agent = model + tools + instructions. Single- vs multi-agent orchestration, tool design, guardrails, and the line everyone quotes: set up evals to establish a performance baseline.
[secondary] PDF undated; Jan 2025 per coverage
DeepSeek releases R1
China's DeepSeek releases R1, an open reasoning model matching the o-series pattern at a fraction of the presumed cost.
EO 14110 revoked
On inauguration day, EO 14148 (“Initial Rescissions of Harmful Executive Orders and Actions”) revokes EO 14110 among 78 Biden-era actions. The revocation is often misattributed to EO 14179; the instruments are distinct.
OpenAI Operator
OpenAI previews Operator for US ChatGPT Pro users: an agent on its Computer-Using Agent model that drives its own browser to perform web tasks.
EO 14179 — “Removing Barriers to American Leadership in AI”
Three days later, EO 14179 sets US policy to sustain and enhance global AI dominance and directs an AI Action Plan within 180 days.
The DeepSeek market shock
A week after R1, investors reprice the US AI build-out: a global tech selloff in which Nvidia loses $593 billion of market value in a day — a record one-day loss (Reuters/LSEG).
Paris AI Action Summit
France and India co-chair the third global summit, at the Grand Palais: public-interest AI, the future of work, trust, and governance. Its Statement on Inclusive and Sustainable AI is not signed by the US or the UK.
MCP spec 2025-03-26
OAuth 2.1 authorization framework. HTTP+SSE replaced by the new Streamable HTTP transport. JSON-RPC batching (removed again in June). Tool annotations — read-only vs destructive hints.
Google announces Agent2Agent (A2A)
Open protocol for agents from different vendors to collaborate. 50+ launch partners. Explicitly “complements Anthropic's Model Context Protocol.” Core concepts: Agent Card discovery, task lifecycle, messages, UX negotiation via parts.
Claude 4
Anthropic introduces Claude Opus 4 and Sonnet 4 — hybrid models with near-instant and extended-thinking modes, aimed at coding, reasoning, and agent workflows; Claude Code reaches general availability.
A2A spec v0.2.1
Authenticated extended cards, referenceTaskIds. The v0.2 line announced around Google I/O.
MCP spec 2025-06-18
Batching removed. Structured tool output. Servers classified as OAuth Resource Servers; clients MUST implement Resource Indicators (RFC 8707) so malicious servers can't steal tokens. New Security Best Practices page. Elicitation. MCP-Protocol-Version header.
Texas TRAIGA signed
Gov. Greg Abbott signs HB 149, the Texas Responsible Artificial Intelligence Governance Act: disclosure duties for government and healthcare AI, prohibitions on certain harmful uses, civil penalties, a regulatory sandbox, and a state AI council. Effective Jan 1, 2026.
Google donates A2A to the Linux Foundation
The Agent2Agent project forms with AWS, Cisco, Google, Microsoft, Salesforce, SAP, ServiceNow. Google transfers the spec, SDKs, and tooling. 100+ companies supporting.
America's AI Action Plan
The White House releases “Winning the Race: America's AI Action Plan” — 90+ federal actions across innovation, infrastructure, and diplomacy — with three same-day orders: EO 14318 (data-center permitting), EO 14319 (“Preventing Woke AI in the Federal Government”), EO 14320 (exporting the American AI stack).
A2A spec v0.3.0 (breaking)
Agent Card well-known URI changes from agent.json to agent-card.json. Signed cards. mTLS added to security schemes.
GPT-5
OpenAI releases GPT-5 as a unified system — the GPT series' speed with the o-series' reasoning — and makes it the default model in ChatGPT for free and paid users.
IBM's ACP merges into A2A
IBM/BeeAI's Agent Communication Protocol winds down; the repo is archived and IBM Research joins the A2A Technical Steering Committee. A2A consolidates as the de-facto agent-to-agent standard.
[secondary]
MCP Registry launches (preview)
An open catalog and API for discovering publicly available MCP servers at registry.modelcontextprotocol.io — official MCP project, open source.
Google announces Agent Payments Protocol (AP2)
Open standard for agents to transact via cryptographically signed mandates. 60+ launch partners. Built on A2A + MCP — the commerce layer above the messaging protocols.
[disputed] some sources say 09-17 or 09-29
California SB 53 — the frontier-AI law
Gov. Gavin Newsom signs SB 53, the Transparency in Frontier Artificial Intelligence Act (Chapter 138): frontier developers must publish safety frameworks and transparency reports, report critical safety incidents to the Office of Emergency Services, and protect whistleblowers. Effective Jan 1, 2026.
MCP spec 2025-11-25 (“One Year of MCP”)
OIDC discovery for authorization servers. Tool/resource icons. Incremental scope consent. OAuth Client ID Metadata Documents. Experimental Tasks — durable async requests.
Anthropic donates MCP to the Agentic AI Foundation
AAIF, a Linux Foundation directed fund co-founded by Anthropic, Block, and OpenAI. MCP joins Block's goose and OpenAI's AGENTS.md as founding projects. Claimed: 10,000+ public servers, 97M+/month SDK downloads.
[secondary] date per MCP blog
2026
A2A v1.0.0 — first stable spec
Breaking vs 0.x. Protocol separated from transport bindings (JSON-RPC, REST, gRPC). Modernized OAuth 2.0. Native multi-tenancy. Signed Agent Cards. tasks/list with filtering and pagination.
Linux Foundation: A2A one-year press release
150+ organizations, 22,000+ GitHub stars, SDKs in five languages, native integration in Azure AI Foundry, Copilot Studio, Amazon Bedrock AgentCore. Restates the complementarity: A2A for agent-to-agent across boundaries, MCP for agent-to-tools-and-data.
A2A v1.0.1 (patch)
Prefer application/a2a+json in the HTTP binding; error-mapping and TaskStatus fixes. Latest 1.x as of research date.
MCP spec 2026-07-28 — “largest revision since launch”
Stateless core: sessions and the initialize handshake removed; every request carries version + capabilities in _meta. New server/discover RPC. ping, logging/setLevel removed. Tasks moved to an official extension. MRTR replaces sampling + elicitation. SSE resumability removed.
A2A joins the Agentic AI Foundation
A2A accepted as a Growth Stage project at AAIF. Both protocols now share a governance home — specs remain separate, no merger.
REGULATION, PHASED IN
The EU AI Act is a phase-in, not a switch: obligations land in stages from 2025 to 2027. The US picture is executive action at the federal level and statutes in the states.
WHERE IT'S GOING
Trajectory, not prediction — each item is tied to a dated fact above.
- Protocols consolidate under foundations. MCP (Dec 2025) and A2A (Aug 2026) both sit at the Agentic AI Foundation; the specs stay separate, the governance converges.
- Agents get wallets. AP2 (Sep 2025) layers signed payment mandates on A2A + MCP — commerce follows messaging, as it did on the web.
- The EU phase-in reaches high-risk systems. GPAI obligations landed in 2025; high-risk obligations phase in across 2026–27.
- US state laws switch on. Colorado, Texas, and California statutes take effect through 2026 — compliance becomes a state-by-state map.
- Agents move from demos to infrastructure. A public registry (Sep 2025), a stable A2A 1.0 (Mar 2026), and a stateless MCP core (Jul 2026) are what infrastructure looks like while it is still being built.
SOURCES
Expanded from the Daelin timeline — the Protocols track is ported from it, honesty markers and all. Dates on the other tracks were verified against primary sources (the Federal Register, EUR-Lex, state legislature records, and company announcements); each entry links its source. Entries marked [secondary] rest on corroborating coverage rather than a primary page; [disputed] marks a date sources disagreed on.