Raj Reddy : The Future of AI : Doomers vs. Abundance

ELI5/TLDR

Raj Reddy — one of the founding fathers of AI, Turing Award winner, and the guy who helped build Carnegie Mellon’s AI lab in the 1960s — thinks both sides of the current AI panic are wrong. The doomers who say AI will wipe us out are overreacting. The abundance crowd who say we’ll all be rich and retired in five years are also overreacting. His quiet take: the technology will arrive, but the lawyers, unions, and regulators will slow it down by decades. Expect big changes in 20-50 years, not five.

The Full Story

The room the arguments are happening in

Reddy is not a neutral observer. He started in AI in 1963 at Stanford, the same year the U.S. government first funded the field, and he’s watched every hype cycle since. So when he looks at the current moment — trillions flowing into data centers, Hinton warning about extinction, Altman promising a post-work society — he sees two camps of people he respects, both extrapolating too hard.

On one side, the doomers. On the other, the abundance crowd predicting everyone will do a day’s work in an hour and create 10x the wealth. Reddy shrugs at both.

“I’m not taking the existential threat very seriously.”

The existential-risk argument gets the asteroid treatment: possible in principle, low probability, not the thing worth organizing your life around.

What actually changed: compute, by a factor of 10 billion

The backstory for the current moment is boring and physical. In 1967, Reddy’s lab got a PDP-10. One megaflop — a million floating-point operations per second. Today’s Nvidia Rubin R100 runs at 50 petaflops. Elon Musk is building a 100-exaflop data center in Texas. Exaflop is 10^18. That is a 10-billion-fold increase in 60 years.

“Grove giveth and Gates takes it away.”

For decades, Intel’s faster chips got consumed by Microsoft’s bloatier software. The insight nobody saw coming is that once you get to around 10^16 operations per second, you can suddenly train foundation models on essentially the entire internet, and the result feels like general intelligence. ChatGPT was the public’s first glimpse of this threshold being crossed.

Are we already at AGI? Depends who’s asking

There are two definitions of Artificial General Intelligence — the AI holy grail of being able to do anything a human mind can. The grand one says AGI should be able to invent writing from scratch, the way the Egyptians did, or invent the concept of zero. By that standard, we’re nowhere.

The practical one says AGI is a system that can answer any question at 90% of a world-class expert’s accuracy. By that standard, Reddy says we’re already there. His evidence: someone ran a foundation model through all 38 Advanced Placement exams. It passed everything. It aced the technical subjects — math, physics, chemistry, biology — and scraped through the humanities with Bs and Cs. That, he argues, is a polymath.

But the jump from “AGI exists” to “college is obsolete” has a hole in it.

“If you don’t know what the right question to ask, the prompt, right prompt, you won’t get the right answer.”

The AI makes expertise cheap. It doesn’t make taste, judgment, or the ability to ask a useful question cheap. Those still need a human who knows what they’re looking for.

The real bottleneck isn’t the tech

This is Reddy’s central move. He accepts almost every abundance prediction on the technical side. Yes, foundation models plus humanoid robots (projected at around $20,000 a unit, working four shifts non-stop) can in principle deliver healthcare at 10% of the current cost. Yes, AI tutors can deliver Benjamin Bloom’s famous “two-sigma” effect — one-on-one tutoring that turns average students into A students — to everyone.

The problem isn’t capability. It’s diffusion. Society runs on rules, licenses, unions, regulations, and laws, and those are what decide how fast technology actually lands.

He tells a story about his colleague Alex Waibel working in Karlsruhe, Germany, where the labor unions are strong enough that a security guard came over to tell him he wasn’t allowed to work on Sundays. Then Reddy delivers the punchline:

“An extension of that is not only human beings cannot work, but robots cannot work also. So, they passed a regulation law saying robots cannot work on Sundays.”

Multiply that by every hospital board, every medical licensing body, every teachers’ union, every insurance regulator, every country’s employment law. The technology gets built fast. The permission to use it gets built slowly. That’s why Reddy’s timeline is 20-50 years, not 2-5.

What happens to work, if it’s optional

If one person can do ten people’s jobs, you get one of two endings: nine people unemployed, or everyone doing a tenth as much work. The policy vocabulary for the first is UBI — Universal Basic Income, a government check to everyone. The vocabulary for the second is Universal Basic Services — the state provides electricity, water, healthcare, and education for free, funded through taxes, the same way most roads are free today.

Reddy’s bet is that work eventually becomes optional, the way it already is for certain aristocracies. He brings this back to his childhood village, where half the population didn’t work and someone would offer to carry your bag if they saw you walking down the street. He calls it the Maharaja model.

But he carefully separates two kinds of work. The neurosurgeon who is the only person capable of a particular surgery will still work long hours — both because she’s needed and because she wants to. The violinist will still practice. Skill-based and meaning-based work survives; routine cognitive and physical labor gets automated.

How education actually needs to change

Reddy has been around AI tutors since the 1970s, when Herb Simon and Allen Newell were building them at CMU. He thinks the pieces are finally in place. Three ingredients are needed:

One, one-on-one education. The two-sigma finding has been known since 1984 but was impossible to scale because it required half the population to be teachers. AI tutors fix that.

Two, learning to learn — teaching students how to ask good questions and how to recognize what they don’t know. The skill that matters is prompting, not memorizing.

Three, just-in-time learning. Reddy took four semesters of calculus and has never used most of it. His view: teach the basics, then let people learn specific topics when they actually need them. The exception is skill-based learning — piano, cooking, basketball, surgery — where you can’t just look it up. Those still need practice, social interaction, and real teachers.

He closes with a thought experiment borrowed from a 1980 Simon study in China. Two classes learned algebra — one through lectures, one through working out examples in silence with the teacher available only when asked. Three years later, the example-driven class remembered 50% more. Reddy thinks that’s the future classroom. The teacher hands out today’s problems and says “figure it out.” The AI tutor is the escape hatch when you get stuck.

Key Takeaways

Compute has grown 10 billion times since 1967 (1 megaflop → 10-100 exaflops). The current AI boom is what happens when you cross roughly 10^16 operations per second with enough training data.
Two definitions of AGI: human-level creative intelligence (we’re nowhere) vs. a polymath that can answer any question at 90% of expert accuracy (we’re already there, per AP exam benchmarks).
Foundation models passed all 38 AP exams — aced technical subjects, scraped Bs and Cs in humanities.
Moravec’s paradox: easy things for humans (walking, speaking) are hard for computers; hard things for humans (theorem proving) are easy for computers. Named after Reddy’s colleague Hans Moravec.
The bottleneck is regulation, not technology. Reddy’s 20-50 year timeline comes from diffusion lag — unions, medical boards, licensing authorities — not from engineering limits.
Germany passed a law banning robots from working on Sundays. Preview of what’s coming to healthcare automation everywhere.
Humanoid robots projected at ~$20,000 per unit, working four shifts non-stop. Paired with AI, could deliver healthcare at 10% of current cost — eventually.
Bloom’s two-sigma problem: one-on-one tutoring moves average students to A-grade performance. Known since 1984, unaffordable until now. AI tutors make it scalable.
UBI vs Universal Basic Services: two competing policy frames for a post-work economy. UBS says the state provides electricity, water, healthcare, education for free, funded through taxes — the road model extended.
Prompting is the new literacy. AI doesn’t devalue education; it shifts the core skill from knowing answers to asking good questions.
Just-in-time learning replaces just-in-case learning for informational subjects. Skill-based learning (music, sports, cooking, surgery) still needs practice and human teachers.
Herb Simon’s 1980 study at the Chinese Academy of Sciences: students who learned algebra from worked examples (with teacher on standby) remembered 50% more after three years than students who learned via lectures.
Programmers won’t disappear but will shift to writing evaluation functions, validation, and the RADS attributes: Reliability, Availability, Dependability, Security.
AI will not “coexist” with humans as a peer intelligence. Reddy’s view: it stays a tool, an assistant. Not a rival species.

Claude’s Take

This is the clearest AI-future talk I’ve seen in a while, and it’s worth listening to for one reason: Reddy has zero financial stake in the outcome. He’s 88, he’s watched every AI hype cycle since there were AI cycles, and he’s not trying to sell you a subscription or a cautionary book. The deadpan wisdom of someone who helped build the thing and has been around long enough to see which predictions aged well.

The strongest part of the talk is the diffusion argument. The tech-optimist framing of the last three years has been almost exclusively about capability — what can the model do, how fast is it improving, when does it match a human. Reddy pivots to the unsexy question: what does it take to actually change how medicine or education is delivered? And the answer is decades, because hospital boards and licensing authorities and insurance companies and teachers’ unions are all real things that move slowly. The German robot-Sundays story is the talk’s best joke and its best argument.

The weaker part is the implicit optimism. Reddy assumes we’ll land on something like Universal Basic Services, and that the diffusion lag will be “enough time for society to figure it out.” Both of those are optimistic extrapolations. The Luddite comparison he makes in passing is worth pushing on — the Industrial Revolution did eventually raise living standards, but the first two generations of displaced workers had brutal lives, and “the diffusion will be slow enough” isn’t much comfort to someone whose job vanishes in year seven of a fifty-year transition. He handwaves the transition costs.

Also: the claim that foundation models are already AGI because they pass AP exams is doing some work. Passing an AP exam is a test of knowledge retrieval and pattern matching in well-defined domains. It’s not the same as doing original research, running a startup, or writing a novel that matters. Reddy knows this — he makes the distinction between his two AGI definitions — but he’s generous with the label in a way that a stricter thinker wouldn’t be. Score of 7 because the framework is good but some of the specifics are loose.