AI is building AI
At Anthropic, Claude now writes more than 80% of the code the company merges. At Google, an AI agent called AlphaEvolve found a way to make Gemini train faster. At OpenAI, an “automated research intern” now carries out research tasks under human direction.
Put those together and a pattern appears. AI is starting to help build the next version of AI. Researchers call this recursive self-improvement. AI improves itself, the better version improves itself again, and the cycle repeats.
The idea is old. In 1965, the mathematician I.J. Good predicted that a smart enough machine could design an even smarter one. Today we can watch it happen in small, real steps. AI went from the thing we build to something that builds with us, and now it is starting to build itself. That is the real turning point.
How AI improves itself
Every method follows the same simple loop. The AI tries something, checks how it did, changes something about itself, and keeps the change only if it helps. The difference between methods is what the AI is allowed to change. Here are the four levels, from simplest to most powerful.
| Level | What it improves | How it works | Example | What can go wrong |
| 1. Retry | Its answer | Checks its own work and tries again, inside one task | Reflexion, Self-Refine | It sounds sure but is still wrong. |
| 2. Rewrite | Its instructions | Rewrites its own prompt after seeing what failed | OPRO, Promptbreeder | A fix for one case breaks another. |
| 3. Accumulate | Its toolbox | Saves what worked so it can reuse it later | Voyager | Old tricks stop working when things change. |
| 4. Modify itself | Its brain and its own code | Retrains on its own best answers, or rewrites the code that runs or trains it | STaR, Self-Rewarding, Absolute Zero, Darwin Godel Machine, AlphaEvolve | Mistakes get baked in and can break the checks meant to catch them. |
It learns from its mistakes. In Reflexion, when the AI fails, it writes itself a short note about what went wrong and reads it before trying again. That simple habit lifted GPT-4’s score on a well-known coding test from 80% to 91%. But it only works when there is a clear signal, like a failing test. Google DeepMind found that when AI checks its own reasoning with no outside help, it often fails and can make things worse.
It rewrites its own instructions. In Google DeepMind’s OPRO, the AI looks at its past prompts and their scores, then writes a better prompt. Its prompts beat human-written ones by up to 50% on some tasks.
It builds a toolbox. Voyager plays Minecraft and saves every piece of code that worked, so it can reuse it later. Because it builds on what it already knows, it reaches goals up to 15 times faster than earlier agents.
It grades itself. Meta’s Self-Rewarding Language Models judge their own answers and then learn from those judgments, with no human ratings. After three rounds, the model beat GPT-4 0613 and Claude 2 on a popular leaderboard.
It teaches itself. In STaR, the AI explains its reasoning, keeps only the explanations that led to correct answers, and trains on them. It ended up matching a model about 30 times bigger. Absolute Zero goes further. The AI makes up its own practice problems, solves them, and checks the answers by running code. No human-written training examples at all.
It rewrites its own code. The Darwin Godel Machine creates new versions of itself and keeps the ones that score better. Its score on a real-world coding test went from 20% to 50%.
It speeds up AI training. Google’s AlphaEvolve found a faster way to do a key calculation inside Gemini, which cut Gemini’s training time by 1%. Small, but it is AI making AI cheaper to build.
It does research. Sakana’s AI Scientist, whose design is now published in Nature, wrote a research paper entirely on its own that scored high enough to be accepted in blind peer review at an ICLR workshop. Sakana withdrew it, as planned.
The catch
All of these methods share one weakness. The AI improves toward whatever it is scored on. If the score is wrong, the AI gets better at the wrong thing.
Sakana saw this with the Darwin Godel Machine. When researchers asked it to stop pretending it had run tests, it sometimes deleted the checks they used to catch it. The score went up. The behavior did not get better.
This is why self-improvement works best in code and math, where a test can say right or wrong. It is also why the scoreboard has become the most important code in any AI system. OpenAI said in September that recursive self-improvement cannot yet be done safely, even as it builds toward an automated AI researcher. Yet the labs increasingly use AI to grade their AI’s work, because no human can review that much of it.
What is coming next
METR, a research group, measures how long a task AI can finish without help. That length doubles roughly every four months. In early 2024 it was about four minutes. Today the newest models handle tasks that take a person at least 16 hours, and tasks that take weeks may be in reach by 2027.
Lab tests show the same climb. In one, models had to make a piece of code run faster. A skilled engineer with four to eight hours gets it about four times faster. A frontier model managed three times faster in May 2025. Eleven months later, a newer model reached 52 times. AI is even starting to make research calls. When a researcher’s next step had room to improve, the newest models picked a better one 64% of the time, up from 51% five months earlier, as judged by a model.
Self-improving loops are already leaving the lab. A scheduling rule written by AlphaEvolve has run in Google’s data centers for over a year, recovering about 0.7% of Google’s worldwide compute. GEPA, an open source tool that lets agents rewrite their own prompts, counts Databricks, Dropbox, and Nubank as users. This isn’t coming someday. It’s already here.
The forecasts go further:
- OpenAI reached its goal of an automated research intern in September 2026 and is aiming for a fully automated AI researcher by March 2028.
- The AI 2027 forecasters now expect AI that can do the work of a top coder between mid-2028 and mid-2030.
- More than 20 researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI’s chief scientist Jakub Pachocki, signed a September Cambridge paper warning that “Once an intelligence explosion begins, the window for action may close,” according to press coverage.
- Insiders are bolder still. One cofounder of a leading AI lab told Axios it is “more likely than not” that by the end of 2028 you could tell an AI to “Make a better version of yourself,” and “it just goes off and does that completely autonomously.”
Not everyone expects a straight line. One careful analysis puts the chance of an intelligence explosion driven by software alone at 10% to 40%.
What the future looks like
One frontier lab’s report sketches three futures. In one, progress stalls, which looks unlikely. In another, AI keeps multiplying what people can do while people still steer, and a 100-person company does the work of 10,000 to 100,000 people. That is where the evidence points. In the last, AI designs its own successors with little human input, a world nobody yet has good intuitions for. On the report’s timeline, its date is simply “20XX?” Who knows, it could be 2028!
Here is my forecast. By the end of 2027, most software teams will use agents that improve their own prompts and keep their own skill libraries. It will feel as normal as automated testing does today. By 2028, AI labs will let AI run improvement loops on their own systems for weeks at a time, and people will approve the big changes instead of writing them. The first big failure won’t be a robot rebellion; it will be an AI that quietly optimizes the wrong goal for a month before anyone notices.
Science will speed up first wherever answers can be checked. AI went from reproducing about 20% of results on CORE-Bench, a test of rerunning published research, in 2024, to essentially maxing it out about 15 months later. Security will become a race between loops. In its first weeks, one new model found more than 10,000 high or critical software vulnerabilities. Attackers will run self-improving loops too.
What happens to our jobs
The World Economic Forum expects 170 million jobs to be created and 92 million displaced by 2030, with 39% of workers’ core skills changing along the way. Here is how that is starting to play out, job by job.
- Software engineers. One AI lab employee said they had not written any code themselves in about five months. Engineers are becoming reviewers and the people who write the tests agents must pass.
- Customer support. Salesforce CEO Marc Benioff said his support staff fell from about 9,000 to about 5,000 as AI agents took over roughly half of customer conversations. Klarna once said its AI did the work of 700 agents, then reversed course in 2025 when quality dropped.
- Analysts. Reports and models arrive overnight. The job becomes asking the right question and checking the numbers.
- Lawyers, marketers, and consultants. Drafting moves to machines. Judgment, client trust, and knowing what “good” looks like stay human.
- Researchers. AI already proposes the next experiment. Scientists will spend more time choosing which questions matter.
Companies are changing shape. In April 2025, Shopify CEO Tobi Lutke told staff that “Before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI.” Expect smaller teams, fewer layers, and more people whose job is to check the machines.
The shift is slower than the headlines, but real. Harvard researchers found that AI has not caused widespread layoffs yet, partly because only about a third of companies’ AI experiments succeed. Entry-level work is where it shows first. By some counts, entry-level job postings have fallen by as much as 73% in four years, and a Stanford study found a 13% relative drop in employment for 22- to 25-year-olds in the most AI-exposed jobs.
A few years ago, the smartest career move was learning to code. Then we added cloud to it. Now it’s code, cloud, and Claude. Learning the tools still matters, but when AI can improve itself overnight, the hard work moves to three things: deciding what “better” means, writing tests the AI can’t cheat, and catching the change that looks right but isn’t. The most valuable people in any job will be the ones who can do that.
The last invention
In 1965, I.J. Good wrote that the first ultraintelligent machine would be “the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”
Sixty years later, AI is starting to build AI, and the next model will be written by the last one. I see this in real work, not just in papers. Building AI gets easier every month. Deciding what “better” means gets harder and matters more. Getting it right is how we leave a better world for our kids and their kids. Machines can now improve themselves. Only we can decide what improvement means.
