I recently read a newsletter that I found really interesting.
It was written by Jack Clark, co-founder of Anthropic. His newsletter is called Import AI, and it’s where serious AI researchers go to understand what’s happening at the frontier. I read it every week because it tends to be honest in a way that most AI coverage isn’t.
Issue 455 had a simple, alarming thesis: there’s a better than 60% chance that by the end of 2028, an AI system will be capable of building its own successor — end-to-end, no human required.
His words: “I now believe we are living in the time that AI research will be end-to-end automated. If that happens, we will cross a Rubicon into a nearly-impossible-to-forecast future.”
I thought about my students.
What “AI Building Itself” Actually Means
This isn’t science fiction framing.
The basic idea is this: AI systems have gotten so good at coding, so good at running experiments, so good at reading and reproducing scientific papers, that all the pieces are now in place for an AI to do the work that human AI researchers currently do. And if those pieces are in place, it’s only a matter of time — maybe a short time — before someone assembles them.
Here’s what the benchmarks show. In late 2023, the best AI systems could solve about 2% of real-world software engineering problems. Today, that number is over 93%. In 2022, AI could reliably handle tasks that would take a human about 30 seconds. By early 2026, that number had jumped to 12 hours of independent, unsupervised work. A benchmark that measures whether AI can reproduce the results of a scientific paper — reading it, running the code, checking the outputs — went from 21% success to over 95% in about 18 months.
Each of those numbers, in isolation, is impressive. Together, they tell a story about a system that is already capable of doing a significant chunk of what AI researchers do every day. Clark’s argument is that we’re not predicting this future — we’re watching it arrive.
What Changes When AI Builds Itself
If AI can learn to build a better version of itself — if it can read research, run experiments, optimize its own training, and produce a more capable successor — then the speed at which AI improves stops being constrained by how fast humans can do research. It becomes constrained only by compute and time. That’s a fundamentally different kind of acceleration than anything we’ve seen before.
We’ve already watched AI get good enough to write student essays, pass bar exams, generate lesson plans, and solve multi-step math problems. That happened while human researchers were still in the loop. What happens to that pace when the researchers are removed?
The question isn’t “is this real?” — the evidence he cites is credible — but “what are we actually preparing our students for?”
The Argument for Human Learning Gets Stronger, Not Weaker
Here’s the thing: I don’t think the answer is despair. I think it’s clarity.
For years, teachers have been having a slow, uncomfortable conversation about what school is for. If a student can Google the answer, why memorize it? If AI can write a passable essay, why teach five-paragraph structure? These questions have been uncomfortable because we haven’t had great answers — just gestures toward “critical thinking” and “real-world skills” that often feel more like slogans than roadmaps.
What Clark’s newsletter does, inadvertently, is sharpen the question. If AI can now do the work of researchers — not just the grunt work, but the actual scientific work of reading, hypothesizing, experimenting, and refining — then the human value proposition in learning isn’t about producing output. It never really was. It’s about something else entirely.
It’s about the judgment that comes from having actually struggled with a problem. The ethical instincts that develop when you’ve been held accountable for your conclusions. The wisdom that grows from being wrong in front of people who care about you. The curiosity that sparks not from a prompt but from genuine confusion about the world. The capacity to ask whether we should do something, not just whether we can.
AI systems can improve their own code. They cannot improve their own values. That’s not a knock on the technology — it’s just a description of what it is. Values aren’t weights you optimize. They’re the product of being a person in a world with other people, with real stakes, over time.
What This Means in a Classroom Right Now
I’ve been teaching AP World History for a while, and one of the things I love about the course is that it demands students grapple with real complexity. Not just “what happened” but “why did people make those choices, given what they knew and believed at the time?” That question doesn’t have a Google answer. It doesn’t have an AI answer either — or at least, not one that does the work for the student. You have to sit with it.
I think about that kind of learning differently now. Not as a relic of the pre-AI world, but as one of the most future-proof things I can help a student do.
The skills that matter most in a world where AI builds AI are not the skills that AI is best at. They’re the skills AI is structurally incapable of having: the ability to decide what’s worth doing in the first place, to weigh competing values when there’s no right answer, to sit with uncertainty without outsourcing it, to maintain your sense of self when the tools around you are constantly changing.
Those are the things we teach when we run a Socratic seminar and make a student defend their position to their peers. When we give feedback on an essay draft and ask the student to revise — not because the words weren’t good enough, but because the thinking wasn’t finished yet. When we hold a student to a standard, even when they’re frustrated, because we believe they can meet it.
Clark ends his newsletter with a note of genuine uncertainty. He writes that he’s “reticent” to believe his own conclusion because “the implications are so large that I feel dwarfed by them.” That honesty is worth sitting with. This is one of the most informed people on the planet on this topic, and he’s saying he doesn’t know what comes next.
Teachers don’t always know what comes next either. But we know something that matters: the students in front of us need to become people who can navigate a world that’s harder to predict than it’s ever been. That’s not a new mission. It’s just more urgent.
The machine is learning to build itself. The best response I know of, as a teacher, is to make sure that the humans in my room are learning to build themselves, too.
Takeaway for Teachers
This week, find one moment in your class where you resist the urge to resolve the confusion too quickly. Let students sit with a hard question — about history, science, ethics, literature, doesn’t matter — longer than is comfortable. Not because struggle is good for its own sake, but because the capacity to stay with a hard problem without outsourcing it to a tool is exactly the kind of cognitive muscle that a self-improving AI can’t develop for them.
David Jacobson is a high school history teacher. He writes about AI, education, and the messy intersection of the two at shouldiuse.ai.
