An Anthropic co-founder just put a number on it. Most people read it as a doom warning. I read it as the most useful calendar anyone learning AI has ever been handed.
In May 2026, Jack Clark, co-founder of Anthropic (the company that makes Claude) and its Head of Public Benefit, wrote this in his personal newsletter, Import AI:
Jack Clark, Import AI #455
"I reluctantly come to the view that there’s a likely chance (60%+) that no-human-involved AI R&D - an AI system powerful enough that it could plausibly autonomously build its own successor - happens by the end of 2028."
In plain English: he thinks it’s more likely than not that within about three years, AI will be able to build the next version of itself with no human doing the research. He put the chance for 2027 at 30%. This is his own forecast, not an official Anthropic prediction, and he hasn’t changed it since.
The key measure is how long a task AI can finish on its own, timed by how long it would take a skilled person. Researchers at METR track this. Here’s the timeline Clark laid out:
01
2022 · 30 seconds
GPT-3.5 could handle tasks that take a person about 30 seconds.
02
2023 · 4 minutes
GPT-4.
03
2024 · 40 minutes
OpenAI’s o1.
04
2025 · 6 hours
GPT-5.2.
05
2026 · 12+ hours
Claude Opus 4.6 hit about 12 hours. METR’s newest measurement, Claude Mythos Preview, is at least 16 hours, which METR says is past what its current tests can reliably measure.
Ajeya Cotra, a forecaster at METR, thinks tasks of around 100 hours by the end of 2026 aren’t unreasonable. That’s her view, not a promise, but the curve has kept climbing.
Most big tech shifts blindside people. The internet, smartphones and social media didn’t come with a three-year heads-up from the people building them. This one did.
The Real Question
Not "will this happen?" It’s "what do I want to know how to do by then?" That’s what the next few years are for.
01
Build a real agent, not just a prompt
The basic unit of AI work is becoming the agent, which knows your context, uses your tools and remembers. Build one this month so it’s second nature. Here’s how I build mine.
02
Get good at judging AI work
As AI writes better, the skill that matters is knowing whether it’s actually good. Learn to write a checklist for "good" and check output against it. Anthropic’s free AI Fluency course teaches this as Discernment. Links here.
03
Connect AI to the tools you use
The value is moving into what AI is plugged into. Pick one set of tools you care about, like Gmail, Calendar and Drive, or Shopify and your ad accounts, and learn to connect Claude to them.
04
Ship something small, publicly
With Claude you can build a working app, a Skill or an agent in a weekend without being an engineer. Put it where someone can click it. That proof will matter more than any resume line.
05
Follow the safety side
What AI is allowed to do will shape what it does. People who understand AI safety and policy, like Anthropic’s Responsible Scaling Policy and the EU AI Act, will be rare and valuable.
The Right Response
A 60% forecast isn’t a prophecy, and it isn’t a reason to panic. It’s a reason to start the small kind of practice that compounds. Three years of that and you won’t be catching up to the curve. You’ll be on it.