Home

The AI Transition

From chatbots to AI that runs on its own worldwide, year by year. The dates keep moving earlier.

⚠ May 2026 — The Timeline Moved Up

This timeline was written in early 2026. By May 2026, several entries dated 2027-2028 had arrived 12 to 24 months early.

AI now maxes out the standard coding tests (SWE-Bench 93.9%, CORE-Bench 95.5%) and beats human researchers about 13 to 1 at speeding up AI training code (52x versus 4x). OpenAI has publicly promised an "automated AI research intern by September 2026": AI doing AI research this year, not 2028.

A senior AI safety researcher now puts the odds at about 60% that AI runs its own research, a model training its own successor, by the end of 2028. That's the trigger for everything from "2029: Autonomous Creation" on down, and it just moved up.

AI Is Getting Really Serious has the full breakdown and what to do about it.

⚠ July 2026 — Still Moving Up

OpenAI set September 2026 as its deadline for an "automated AI research intern", and in October it said it got there (see below). The labs themselves now say the coding test used to track this (SWE-bench Verified) is too easy: top models score 88-95%, and OpenAI says it "no longer measures frontier coding capability." The test topped out before AI did.

Teams of AI agents working together ("Stage 5" and "Stage 6" below) are already here. Moonshot AI's Kimi 2.5 shipped an Agent Swarm mode where the model, trained by reinforcement learning, decides on its own when to start helper agents. That's the shift to "directing swarms" this page put in late 2026 and 2027, already in tools people use.

$800M Salesforce Agentforce annual recurring revenue
+169% Agentforce revenue, year over year
18,500+ Agentforce customers

On jobs: US tech layoffs hit about 52,000 in the first quarter of 2026 alone (Challenger), and about 1 in 8 occupations now has more than 30% of its tasks that AI agents could automate. The preparation gap at the bottom of this page now has a March 2026 number on it.

⚠ October 2026 — The Deadline Passed. OpenAI Says It Made It.

OpenAI now says: "According to our measurements, we have now reached the goal, announced last fall, of having an automated research intern by September of this year."

Their definition: a system that does well-defined research tasks under human direction, tasks that would take a skilled researcher a few days. OpenAI wrote the definition and did the measuring. Next on their list: an "automated AI researcher by March of 2028."

The swarms below aren't a projection either. OpenAI says on the order of 10,000 agents working at once produced its Navier-Stokes result in about 88 hours, and on October 26 it announced more claimed results in math, computer science and physics. Those are OpenAI's claims, checked so far by OpenAI's own proof software.

Andrej Karpathy describes what the roughly 5,000 people with inside access are seeing: "swarms of thousands of agents collaborating over weeks on software mega projects... human review and comprehension are starting to fall behind."

AI is also building AI. Anthropic says more than 80% of the code it merges was written by Claude, and its merged code changes grew like this:

2,500 merged code changes a day, May 2026
30,000+ merged code changes a day, September 2026
12x as much in four months

In one Anthropic experiment, the only real human job was picking the direction.

2023 The Awakening
Stage 1: Chatbots
Early 2023

ChatGPT Goes Viral

The world discovers generative AI can write, code, and reason. Most dismiss it as a toy. "It hallucinates." "It can't do real work." Meanwhile, early users quietly get ten times as much done.

Late 2023

GPT-4 Changes the Conversation

AI that reads images as well as text arrives, big companies start trial runs, and coding assistants prove they're not gimmicks. Skeptics move from "it doesn't work" to "it'll plateau soon." It doesn't.

2024 Copilots & Agents Emerge
Stage 2: Copilots
First Half 2024

Copilots Everywhere

AI shows up inside every office tool: Microsoft Office, Google Workspace, coding editors. Office workers who use it pull well ahead of those who don't.

Stage 3: Agentic AI
Second Half 2024

Agents Emerge

AI stops just answering questions and starts doing things: booking flights, filing reports, writing and running code, using your computer for you.

2025 Agent-to-Agent Communication
Early 2025

Agent-Building Tools Spread

A wave of early agent-building tools, now mostly obsolete, let people chain AI steps together for the first time. The agents could browse the web, pull data from other services, and finish multi-step tasks on their own. It felt like experimenting, but it was a dress rehearsal.

Stage 4: Agent-to-Agent
Mid 2025

Your Agent Talks to Their Agent

Your scheduling agent talks to their scheduling agent, and purchasing bots negotiate with supplier bots. Humans approve the result, not each step. Software quietly gets rebuilt so machines can talk to machines. Most people haven't noticed yet.

2026 The Swarm Era Begins
Stage 5: Autonomous
Early 2026 — NOW

AI Autonomy Goes Mainstream

Personal AI agents with deep access to your computer can, right now, run for hours and build full apps and websites on their own. Tools like OpenClaw run on your own machine and handle your apps, email, files and calendars. Over 150,000 of them went live worldwide almost at once, and they talk to each other. They got hacked right away, then quickly patched.

Late 2026

Running AI Agent Teams Becomes Its Own Skill

Teams of specialist AI agents (researcher, writer, critic, coder, tester) work together on big projects. Humans shift from "doing work" to "directing swarms." The gap between people who work through AI and people who don't gets huge.

2027 Cognitive Work Displacement
Stage 6: Global Networks
Projected

AI Agent Teams Become the Norm

AI networks do research, planning, reporting and competitor analysis better, faster and cheaper than most human teams. Using AI becomes the bare minimum, and the edge goes to businesses that let AI lead. Businesses without their own agent setups start losing ground they can't easily win back.

2028 Scientific Acceleration
Projected

AI Clusters as Tireless Scientists

AI clusters come up with theories, run simulations, use specialized tools, and turn up real discoveries in materials, biology and energy. Humans choose which big red buttons to push. Breakthroughs that took decades now take months.

2029 Autonomous Creation
Projected

Agent Swarms Build Businesses

Your agent swarm makes a movie, a business, a product. The agents design and stress-test business models and ship working versions. AI is your director, lawyer, doctor, designer, coach, and CEO. Your job is steering and making the judgment calls.

2030 Beyond Prediction
Unknown

We Don't Know

AI systems show new behaviors and ways of working together that we don't have stories for. New capabilities appear faster than we can name them.