Reid Hoffman has a line worth stealing: the AI products that work don't sell labor replacement. They sell "lazy and rich." Let you work fewer hours. Let you make more money. Frame it any way you like — the substance is the same, and that combo, he says, is killer.
This sounds obvious until you watch how almost every founder pitches AI. The deck says "automate your team out of existence." The case study shows headcount going down. The ROI math is cost-per-employee divided by subscription price. And then the deal stalls in procurement, because nobody at the buying organization wants to be the person who brought in the tool that puts their colleagues out of work — and the principal-agent problem at any company larger than fifty people means the director who would save the money isn't the one who would get credit for it. As Hoffman puts it, the company saves money but the director just wants to leave earlier and get promoted. That ethereal being called "the corporation" is a terrible customer.
The products that move are the ones aimed at the individual. The dermatology clinic that can see five times the patients. The plaintiff's attorney who can run five times the settlements. The sole proprietor whose deep research tool gives them a due-diligence plan in five minutes that used to take a day. None of them are firing anyone. They're getting their evening back and earning more during the day. That's diffusion that compounds.
Hoffman's other framing on this is sharper still. He calls AI "massively underhyped" — a heretical claim in Silicon Valley, where the default debate is whether valuations are too high. His point is that most of the population has never seriously used a current model. They tried something a couple of years ago, it didn't solve their problem, they decided it was bad, and they stopped. The mistake is judging a technology on its present. He brings up a video of Tiger Woods at age two and a half hitting a perfectly straight drive on the Tonight Show — you can either say "I can hit further than that kid" or "if that kid keeps it up, he'll be Tiger Woods." Most people pick the wrong one. Ethan Mollick has the corollary: the worst AI you will ever use is the one you are using today.
Where the LLMs still genuinely fall over is more interesting than the demos suggest. Hoffman ran an experiment recently. He had a debate scheduled on whether AI would replace all doctors in a small number of years, and being the kind of person he is, he set up Chat GPT Pro, Claude Opus, Gemini Ultra, and Copilot deep research in parallel browser tabs and prompted them all to build him the strongest possible case for his side. Ten to fifteen minutes of compute on each, the kind of work an analyst does in three days, run in parallel. The output was a B-minus. Consensus opinion, dressed up well. None of the systems gave him a contrarian answer worth using in a debate, because none of them were doing lateral thinking — only confidence-weighted summarization of what good magazine articles already said.
That tells you both what the next decade of professional work looks like and what it doesn't. The knowledge store part of being a doctor, lawyer, or analyst is gone. Two thirds of doctors already use OpenEvidence. The Harvard Medical School credential was a heuristic for "this person has the knowledge base," and now we have the knowledge base on tap for twenty dollars a month. What survives is the part of the job that's lateral thinking — the part that asks "the AI gave me consensus opinion, what if consensus opinion is wrong here, and what would I look at if I wanted to find out." Doctors who learn to do that win. Doctors who do not are competing with a free B-minus.
And this is where the Silicon Valley blind spot gets real. The default SV instinct on every problem is "put it all in software, simulate, ship." That works for productivity tools. It does not work for drug discovery, where the simulation problem is genuinely hard. It does not work for folding laundry, where you are up against capex curves and battery chemistry that bits don't fix. Japan builds robotics because they cannot hire anybody — at the bowling alley, a vending-machine robot hands you your shoes and cleans them after you. America hires the high-school kid because the capex line still sits above the opex line. Bits are easy. Atoms are where the contrarian return lives, and Hoffman is putting most of his time on the boundary between them — bio, where he co-founded Manas AI with Siddhartha Mukherjee — precisely because that is where the line of sight is not obvious to everyone.
Two takeaways for anyone building or buying.
If you are selling AI: stop pitching labor replacement. Pitch "lazy and rich" — same hours of revenue, fewer hours of work — and pick the buyer who actually keeps the upside. Sole proprietors, small clinics, individual professionals. Big-company budgets get stuck on the principal-agent rocks every time.
If you are using AI: assume the model gave you a B-minus consensus answer, and budget the time to ask the second question — the lateral one. The professional advantage in 2026 is not having access to the model. Everyone has that. The advantage is being the person who treats the model's answer as the starting position, not the conclusion.
Source: Reid Hoffman on AI, Consciousness, and the Future of Labor
