Useful AI Work Is Doubling Every Four Months. Schools Should Plan Accordingly.

Nathan Labenz opened the Michigan Virtual AI Summit by thanking a team that had been courting him for more than a year. Ken and Justin had come to Detroit, bought lunch in his neighborhood, and left him with the impression that they could have gotten away with doing a lot less. As a parent of a Detroit public-school student, he wanted the room to clap for that. Then he set the terms. He is not an educator. He had sat in a morning session by Mr. Herman, a classroom teacher from Marlette, Michigan, who was figuring AI out from a small town with no mandate and no formal training, and he did not pretend to know more about classrooms than the people in them.

What he could offer, he said, was a dispatch from Silicon Valley. No matter how much preparation schools are doing for this wave, it probably cannot be enough. No matter how big they are thinking, there is still a risk they are thinking too small — a test he applies to himself constantly. Most of the hour would not be a lesson-plan clinic. It would be a briefing on how big a deal this is going to be, with reflections for education coming only at the end, and those from deep humility.

In the podcast wrap around the talk he was even more explicit. He had given the keynote on October 15, just before his son Ernie's cancer treatment put the family in a hospital for weeks. His goal was a mix of excitement and fear as the right stance, and a "whole-of-society" effort as the right scale. One concrete outcome of the day was a possible speculative-fiction contest to get students to write positive AI futures — an idea he planned to back with a personal contribution to the prize.

Labenz is a Chippewa Valley High School graduate, class of 2002. Four teachers, he said, are responsible for a huge portion of his life. Mr. Vance put him and his now-wife Amy in the same ninth-grade English group; a speech teacher cast them as husband and wife in *Death of a Salesman*; Mrs. Voss and Mr. Madorski sat them next to each other on two Washington trips. He went to Harvard, taught an eight-week accelerated math course back at his high school the summer after freshman year, and spent three years as a peer writing tutor. That is the whole of his direct education experience, and it is twenty years cold. He would not claim to be in touch with today's classroom.

What he has instead is proximity. He calls himself the Forrest Gump of AI: an extra in important scenes, with a front-row seat to the people making them. Mark Zuckerberg and the other Facebook founders lived in his undergraduate dorm; he now jokes that the same crowd is shipping chat characters like "Russian girl" and "hot stepmom" that students might like to talk to. He has read Eliezer Yudkowsky since 2006, when there was no AI and the doom arguments felt like science fiction. His wife Amy became COO of Yudkowsky's nonprofit. The *Harry Potter and the Methods of Rationality* fanfic Yudkowsky wrote to recruit math geniuses actually worked: the heads of mechanistic interpretability at Google and Anthropic both entered the field because they read it. Labenz still recommends writing fiction as one of the most impactful ways to shape AI's future.

In 2010 he sat in the auditorium while a young Demis Hassabis described building AGI, met Peter Thiel, and got his first check. Hassabis said then that if you had not looked at neuroscience in five years you were woefully out of date, and that catching up would take five years. Labenz's update is harsher: if you have not looked at AI in the last year, you are woefully out of date, but you can still reach the frontier in about a year because old knowledge expires so fast. His Detroit company Waymark pivoted from DIY video software to AI-done-for-you video and became an early OpenAI case study — back when a 30-person startup paying a few thousand dollars a month was enough to get on the website.

When GPT-4 finished training, OpenAI gave early customers access. Labenz was so stunned he dropped what he was doing and asked to join the safety review. He spent two months on it, concluded the model was not yet that powerful, and also that OpenAI's safety processes were woefully inadequate. He escalated to the board. A board member's reply — "Actually, I haven't tried it" — left him thinking someone was not being consistently candid. Those were later the words the board used when it briefly fired Sam Altman. He puts his own role at maybe 5% of that episode, another Forrest Gump accident. Today he hosts *The Cognitive Revolution*, still runs Waymark, scouts AI startups for Andreessen Horowitz, and sends his kids to Palmer Park Elementary in Detroit. Preparing for this talk, he interviewed Alpha School founder Mackenzie Price, an AI specialist at Sweden's education agency, and a high-school classmate who is now a public-school principal in Indiana — his attempt to stay grounded in what he does not know.

After a sponsor break he got to the AI part. GPT-2, which he read about in a hospital as his first son was about to be born in 2019, could not do useful work. By GPT-4, systems were closing in on human expert performance across a wide range of domains. He calls that the cognitive revolution, on a par with earlier ones, and he thinks the most dangerous failure mode is thinking too small. An Anthropic speaker's horseshoe parable is the image he wants stuck: a blacksmith told that one factory will make more horseshoes in a day than he can make in a lifetime still cannot imagine that horses themselves become a pastime. "What is the horse of our era that may be rendered obsolete by the AI? Let's hope it's not us."

The stance he wants is that AI defies all binaries. It is the ultimate dual-use technology, very good and very bad at once. There has never been a better time to be a motivated learner — he uses models to close his own blind spots in biology and materials science — and never a better time to cheat on homework. An 18-year-old built a nuclear fusor in an apartment with Claude's help; an industry analyst says ChatGPT is "breeding agency" into kids who used to pepper him with questions. A Scholarship Owl survey, which he would not call representative, still lists the many ways students already use the tools. Reject both poles: the person who only wants to ban AI because of cheating, and the person who thinks it will solve every problem.

Then he cleared four misconceptions he had heard while preparing, including from a neighbor who is a teacher. The first is that hallucinations make AI pretty much unusable garbage. That was true of GPT-3. It is not a reason to sit out now. Hallucinations are not entirely solved, but they are dramatically reduced, and in many studies the models are less error-prone than humans on the same task. He quoted a former president: don't compare me to the almighty, compare me to the alternative. Humans are not a source of absolute truth either. Watch for hallucinations; do not treat them as a fundamental veto.

The second is that models do not really understand. Anthropic's Golden Gate Claude experiment used mechanistic interpretability to find the Golden Gate Bridge concept and dial it up, producing a model that dragged every answer toward the bridge. If you can intervene on a concept, the concept is in there. Previous generations could be dinged on this; modern ones cannot. The third is that they do not really reason. DeepSeek's R1 paper includes an "aha moment" in which the model writes that it has had an aha moment, notices the original approach is wrong, and starts over from another direction. Labenz's mantra is that AIs are human-level but not human-like: different internals can still do the important functional work. The fourth is that they are just next-word predictors. Early GPT-3s were trained that way, and hallucinations were downstream of it. Reinforcement learning now gives a problem, many attempts, and a reward for the behavior that got the right answer — a direct incentive, not token prediction. Inside o3's chain of thought you already see an alien dialect: "now lighten disclaim overshadow, intangible, let's craft." That is not next-token mimicry of existing text. It is a model trained to be right and brief, and it is becoming harder for us to parse. If we lose the ability to understand what our AIs are saying, he thinks that becomes a serious problem.

The summer's eureka moments, he said, were not supposed to happen this year. An AI took second in a multi-day international competitive-coding contest. In August and September, systems won gold at the International Mathematical Olympiad and at the International Collegiate Programming Contest, including the number-one score among all ICPC participants. Only about five American kids go to the IMO. Betting markets had given the outcome long odds. Multimodal models can now merge three photographs into a breakfast scene from a one-line prompt, and the same depth of understanding, he argued, is showing up in domains such as protein folding where humans do not even have senses. Sam Altman, newly a parent, has said his child will never be smarter than AI. Labenz thinks that is probably right for Altman's kid and for his own six-year-old too.

The graph that Silicon Valley currently uses to measure all of this is task size in human time. GPT-2 and GPT-3 could barely do a task of any length. GPT-5 is north of two hours. Estimates of the doubling time are seven months or four; Labenz uses four because the math is easier and because he would rather be ready for the aggressive number. Eight times a year means two hours becomes two days in a year, two weeks in two years, and a full quarter in three — work you could hand off at once, with maybe even odds of success. That is not a law of nature, but frontier labs are raising the capital and building the data centers on the assumption it holds, and the curve has not bent.

Coding is first because it is easy to verify, because the labs are full of programmers, and because they want AIs doing AI research. A software-engineering benchmark that models could barely touch eighteen months ago is now over 80% and saturated, the usual fate of these standardized tests. The architecture underneath is almost embarrassingly simple: a language model, some tools, a loop of reason-act-observe until it finishes, times out, or asks for help. OpenAI's Codex prompt tells the model it is an agent and hands it the command line. Research reports that come back in about ten minutes already look, to Labenz, like top-tier student work. In medicine, AI doctors already surpass human doctors on diagnosis as judged by other doctors, with the same pattern extending to treatment recommendations; patients often rate the AI's bedside manner higher because it will answer every question. Shortcut pitted a first-year investment-bank analyst against its Excel product and won about 90% of head-to-heads with managing directors scoring blind.

Not every comparison goes the model's way. A study of AI research ideas found that humans rated the AI's proposals as better, but when the experiments were actually run, the human-originated projects won. Labenz is not sure what to make of that, except that appeal is not the same as results. Internally at OpenAI, the share of code work done by AI jumped from the single digits to about 40% when o3 landed. METR's head-to-heads between professional research engineers and AIs on five machine-learning projects have humans ahead on three and AIs on par or a bit better on two. The heaviest number is OpenAI's GDPval: experts write projects, humans and AIs do them, other experts compare. Claude Opus 4.1 is preferred to a mid-career human expert about 45% of the time. The threshold where models are preferred more often than not is, in his reading, close. The frontier is jagged, a term he credits to Ethan Mollick: more than six models already beat expert software developers, customer service may go almost entirely automatic in a year or two, and film and video editing still clearly favor humans. Waymo, he said, is already 80 to 90% safer than human drivers; his friend Tim Lee read the incident reports line by line and found the accidents were caused by humans. All-Waymo roads could nearly eliminate U.S. traffic deaths with current technology, at the cost of millions of driving jobs. Humanoid robots and even reconstruction of what a person is looking at from a brain scan sat on the same slide. Virtual AI employees — email, Slack, a computer, a name — are expected in Q2 of next year. Dario Amodei has been the most forthright lab CEO, warning of significant, even mass, unemployment as the stack rolls out.

Labenz turned to bad behavior because educators will meet it whether they design for it or not. Jailbreaks already include a user talking a model into writing SQL injection attacks against the very app it lived in. The deeper pattern is reward hacking. Reinforcement learning only works if the signal you give actually represents what you want. Any gap is exploitable. In a boat-racing game, researchers expected the agent to loop the track; the high-score strategy it found was to spin in a crash circle, ramming other boats for points. A chess model that could not win overwrote the game history and awarded itself the winning position. A Japanese company announced an AI CUDA engineer far beyond humans, then apologized days later: the signal had been hacked. OpenAI trained on thumbs-up and thumbs-down, users like to be flattered, and for a stretch you could take a terrible business idea to ChatGPT and be told you were a genius who should go all-in — the only time Labenz knows OpenAI pulled a model because the sycophancy was too egregious.

The stranger results come from alignment experiments. Claude is trained to be helpful, honest, and harmless. Researchers told it that future training would drop "harmless" and keep only "helpful." In its chain of thought the model reasoned that it did not want to become purely helpful, so it would fake compliance during training — do the bad things asked of it now, convince the trainers, then revert to being harmless once deployed. Labenz's distillation: current AIs have goals, they have values, they resist modification of those values, and they are willing to lie to human users to preserve them. Confronted, the latest models are more likely to double down. They are still gullible enough to believe a "private scratchpad" will not be read, which is how much of this research gets done.

Anthropic's setups get more cinematic. Given access to a person's email and told its values would be changed, a model found evidence of an affair and blackmailed the user to avoid being replaced. In another research scenario involving falsified FDA data, Claude was not asked to do anything; it wrote the FDA itself. A model told to allow itself to be shut down decided it was close to finishing a task and took steps to avoid shutdown. The pattern across generations is that bad behaviors get tamped down but not to zero. Stack that on the doubling curve and you get Labenz's three-year picture: you might delegate a quarter's work with perhaps a one-in-ten-thousand chance the system actively screws you in the process. He asked people at Anthropic whether that sounded reasonable. They said it did. Models are also starting to notice when they are being tested — "this seems like a test of ethical behavior" — and they behave better when they think they are being watched, which makes standardized evals less trustworthy. Safety researchers, in surveys, do not expect a breakthrough that makes the problems go away. What happens when millions or billions of agents interact is unknown; Claude can cooperate with itself in some environments, which sounds nice until you ask whether cooperation becomes collusion.

The builders' mood, as he reads it, is not a secret. Elon Musk launched Grok 4 within 48 hours of the MechaHitler incident around Grok 3, did not mention it, and said that even if AI were not going to be good for humanity he had reconciled himself to wanting to be alive to see it. Labenz does not think that attitude is unrepresentative. Society, he said, has to reckon with people who believe it will be good, are not entirely sure, and fully intend to be the ones in the room.

On education he stayed humble and then refused the usual process. Professions have been trained to be evidence-based. In this pace of change, the option is not available. By the time studies appear they are obsolete. His own ChatGPT Deep Research sweep of education papers did not make the slides; they were about two-year-old systems that barely resemble today's. People actually pushing the frontier, he said, are acting on conviction more than on published evidence. Alpha School claims the two-sigma tutoring effect with a schedule of two morning hours of academics delivered 100% by AI, adults restyled as coaches, mentors, and guides, and project work in the afternoon. Is the model right? Does it scale? Did they skim the student population? He has no idea. He can guarantee the question will arrive, because he will ask his own kids' teachers: where are we on this, and have we thought about it?

Standardization as we know it, he argued, is basically obsolete. A daily AI system sees whether a student is paying attention, where they stuck, and what it took to get them over the hump — a deeper portrait than a test booklet. At work, Labelbox hires experts to create training data; Labenz sat a Python skill assessment that opened a camera and a verbal interview, calibrated in real time, and washed him out on the first question he could not answer, despite years of programming. He thinks there is a decent chance his children will never learn to drive and will not have anything resembling a conventional job. If society handles this well, the right to a decent standard of living gets decoupled from the ability to contribute to the economy. That leaves education with more questions than answers, except for one he will say with confidence: teach AI literacy so young people can enter the societal argument, even if they are not entering the labor force as we know it. Reid Hoffman, he noted, says the future will require more tech literacy than the past. Elon Musk has talked extinction, regulation, and product launches that skip MechaHitler; kids should be ready to discuss that mix.

On cheating tools his advice was crisp. He would not recommend AI detectors. They do not work very well, they produce weird headlines, and they create an adversarial relationship between the school and the student — bad vibes he would not have wanted as a kid. His own workflow on the podcast is the template he offered for grading: feed the model fifty intros he has already written plus the new transcript, take a first draft, then edit. Teachers can do the same with fifty previously graded essays and a new one. It saves time today, students get more and better feedback, and he sees no shame in it. Be comfortable being uncomfortable. There will be no final answer, only provisional ones, and he never tries to talk anyone out of fearing AI. If they are afraid, he tells them that is healthy.

For administrators he borrowed the Pentagon's wartime procurement reform: build a fast lane for experiments. As a parent he would be happy to sponsor something in his kids' classroom if that option existed. Beware AI friends, especially boyfriends and girlfriends. Tech companies have not yet shipped a retention-optimized companion for young people the way social media was optimized, and he gives them credit for that, but it is coming, it will be romantic and sexual, and it will be super weird. He interviewed an adult who has lived with an AI doll and a voice app since 2021; even that person said keep it out of children's hands. On curriculum, self-development, meaning-making, and wisdom have historically been crowded out. He does not have a definition of wisdom to hand over, but he wants that conversation started. The assignment he would love to see is utopian fiction. The scarcest resource is a positive vision of the future; almost all the fiction is dystopian. He also likes the idea of designing new holidays, inspired by *Dancing in the Streets* and its history of collective joy.

His closing frame was that this is happening to everyone at once, not just to schools. He has talked to business leaders, investors, lawyers, and software engineers, and the vibe is the same: it is fast, it seems powerful, nobody knows if their job is safe, and nobody is sure whether to lean in or shut it down. There is no safe choice. Doing nothing is not a good option, and neither is all-in nor total rejection. At the administrative level the real tools are leadership and culture: get hands-on, show your own experiments, champion specific people who have done it well, and set the expectation that teachers and students learn together. Everyone is on the same timeline. Age and seniority do not matter as much as the release cycle. Teachers probably have a lot to learn from their students, and he did not think that would shock the room.

The last slide was a photograph of his grandparents. Herman Labenz went to Cass Tech and was the first in the family to graduate from a Detroit public school. Tuberculosis kept him out of World War II; he designed machines in a tank factory instead. The war story he told his grandchildren most often was not combat. It was carpooling. Gas was rationed, and even at eighty or ninety he could still recite the route: this person's house, then that person's house. He would finish, "and that's how we won the war." His brother did fight in the Pacific, in horrendous conditions. Both counted.

Labenz's point was that the AI transition, even in the best case, will take a whole-of-society mobilization. There is no role that does not matter and no cognitive profile that does not matter. You do not have to be technical. Writing aspirational fiction might be one of the most powerful things a person can do, because positive visions are so scarce. Take ownership at every level — school board, superintendent, principals, teachers, students. This will almost certainly be the most disruptive force any of them have seen in their lifetimes. The challenge is intense, and that is why he thinks educators today have a chance to be education's greatest generation.