Dartmouth

INT. DARTMOUTH MATH DEPARTMENT — TOP FLOOR — DAY — JULY 1956

A window looks out over green New Hampshire hills. A fat dictionary sits on a stand. Eight people in a room built for thirty.

MCCARTHY We need a group project. Chess. IBM 704. Two months. We solve intelligence by Labor Day.

MINSKY Chess is a dead end. I want geometry. Figures. Models. Something you can see.

ROCHESTER I’ll go along with chess if there’s nothing better. But I’d rather work through Feller. Probability. Concrete problems.

SHANNON I’m interested in chess. I’m not convinced it’s relevant to the actual problem.

SAMUEL Checkers is just as good. Though I can’t tell you how we get learning out of checkers either.

SELFRIDGE Chess is fine. I’ll go along.

BIGELOW I’m tolerant. I’m not optimistic about how good Turing machines will be in a few years.

Silence. McCarthy looks at the room like a man who organized a constitutional convention and got a book club.

MCCARTHY So. Chess?

MINSKY No.

ROCHESTER Maybe.

SHANNON Sure. Why not.

Nothing is decided. Everything is decided. They will spend the summer arguing about what intelligence is — and seventy years later, we still are.

---

That scene is mostly true. Ray Solomonoff’s notes from the Dartmouth Summer Research Project on Artificial Intelligence read like the world’s most overqualified writers’ room: brilliant people, incompatible agendas, not enough Rockefeller money, and a deadline that was always fiction.

John McCarthy had coined the term artificial intelligence the year before — deliberately neutral, avoiding “cybernetics” (too much Norbert Wiener) and “automata theory” (too narrow). In September 1955 he, Marvin Minsky, Nathaniel Rochester, and Claude Shannon sent a proposal to the Rockefeller Foundation that is still quoted in every AI history book:

We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College… The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.

They asked for ten people for two months. They got a rotating cast — some for a day, some for eight weeks. McCarthy lost the attendee list. Solomonoff kept better notes than the organizers. Only three people stayed the whole summer: McCarthy, Minsky, and Solomonoff.

It was not a directed research program. It was a collision. And collisions, it turns out, are how fields are born.

The constitutional convention that almost wasn’t

Historians call Dartmouth the founding event of AI — the “constitutional convention” of the field. That is both right and misleading.

Right, because the name stuck, the network formed, and ideas that would define decades of research were argued out loud in a math classroom. Misleading, because the workshop did not produce the unified breakthrough its proposal promised. No single thinking machine. No group chess project on the 704. Rochester was so busy building IBM hardware that he sent his student Trenchard More in his place.

What it did produce was something more durable: a shared problem statement and a fork in the road.

Grace Solomonoff’s history of the summer captures the atmosphere perfectly — including the moment when the group looked up the word “heuristic” in that fat dictionary, and the afternoon tea incident when John Nash barked at Trenchard More to pick something up off the landing as a test. More replied, essentially: pick it up yourself. (Even genius has to take the stairs.)

Ray’s own post-mortem was blunt: the research project “wasn’t very suggestive.” But he listed what was valuable — meeting the people, getting a report reproduced, and the dawning sense that the search problem might be the whole game.

That is Dartmouth’s real legacy. Not a product launch. A field declaration — and a seventy-year argument about how to build it.

Symbolists vs connectionists: the fork we’re still merging

Walk into that math classroom in July 1956 and you would hear two civilizations talking past each other.

The connectionists were building models of the brain. Nathaniel Rochester arrived with work on neural nets and Hebbian cell assemblies. W. Ross Ashby demonstrated homeostats — electromechanical machines that reorganized themselves to find stability. Bernard Widrow drove up from MIT for a week because he’d heard the phrase “artificial intelligence” and had to see what it was. Minsky was still doing semantic and neural approaches. The cybernetics crowd believed intelligence would emerge from adaptive, feedback-driven systems — steersman and ship, neuron and signal.

The symbolists believed intelligence lived in manipulation of symbols — logic, language, abstractions you could write down and reason over. Allen Newell and Herbert Simon showed up for two weeks with the Logic Theorist, a program that proved theorems from Principia Mathematica. John McCarthy was pushing formal systems and the idea that the hard part was search — how to explore an enormous space fast enough. Claude Shannon, father of information theory, sat in the room and asked the questions that made everyone uncomfortable.

The argument was not academic. It was architectural. Connectionists: learn from data, let structure emerge. Symbolists: represent knowledge explicitly, reason over it, use heuristics when the search space explodes.

For decades, symbolists won the funding and the textbooks. Expert systems. LISP. GOFAI. Then connectionists won the benchmarks — ImageNet, AlphaGo, transformers, the whole stochastic era I wrote about in The Stochastic Era. Minsky himself pivoted toward symbolic AI after that summer, influenced in part by conversations with Solomonoff about induction and prediction.

Here is what matters in 2026: the merge is finally happening. Large language models are connectionist engines that manipulate symbols. Tool use is symbolic reasoning bolted onto neural nets. Agents are the synthesis the Dartmouth room could not agree on — neither pure brain modeling nor pure logic, but systems that perceive, represent, act, and learn in loops.

The founders were all partly right. And all partly wrong. That is usually how constitutional conventions work.

The Logic Theorist and the summer’s real demo

If you want one technical artifact to carry out of Hanover, take Logic Theorist.

Newell and Simon did not come to Dartmouth to socialize. They came with a working program — arguably the first genuine AI application — that could prove logical theorems by searching a space of symbolic expressions guided by heuristics. It was not chess. It was not checkers. It was mathematics done by machine, and it made the symbolist case visceral: you do not need to simulate every neuron if you can represent knowledge and search intelligently.

Meanwhile, on July 10, Solomonoff gave a talk on inductive inference — how a machine might learn by predicting future symbol structures from past ones. Shannon suggested a hand simulation. McCarthy worried about search length. Minsky wanted ad-hoc mechanisms; Solomonoff disagreed. Same room, same summer, completely different bet on where intelligence would come from.

Both bets are still paying out. Logic Theorist’s descendants write your code reviews. Solomonoff’s descendants are the compression-and-prediction instincts inside modern LLMs. Dartmouth did not pick a winner. It seeded a tournament that is still running.

”Should my kid still major in computer science?”

I was at dinner in Park City with friends when one of them said it out loud — no hedging, no preamble:

“I’m pulling my kid out of Computer Science. It’s a dead field.”

I could not have been more floored.

Not by the anxiety underneath the statement. I understand why people are nervous. Headlines about layoffs. Agents writing code. The SaaSpocalypse. Parents trying to protect their kids — that part I get.

But floored by two things at once: the sheer lack of foresight, and the helicopter parenting dressed up as wisdom. You do not steer a teenager away from the discipline that is actively rewiring every industry on earth because last quarter’s narrative scared you. That is not protection. That is preemptive surrender.

I get this question in softer forms too — soccer fields, school pickup lines, the polite version of the same fear. Smart people who read the headlines and wonder if they are steering their kids toward a cliff.

My answer is unambiguous: absolutely yes.

Not because the job your kid gets in 2032 will look like the job I got in 1998. It won’t. Not because AI won’t displace tasks — it already is. But because the surface area of software is expanding faster than any single headline can capture.

The field is not contracting. It is metastasizing.

If you have read my post on The Law of Accelerating Returns, you know the frame: technology does not advance linearly. It compounds. Each generation of tools builds the next faster. Our brains default to linear extrapolation — thirty steps, thirty meters — and miss the curve that circles the Earth twenty-six times.

We are on that curve now. And the destinations are not just chatbots.

AGI — however you define the threshold — is the explicit target of every major lab. Context and memory remain the gateway problems. Robotics is finally getting the brain it always deserved — models that reason about physical space, not just tokens. Space is becoming a software industry as much as a hardware one; launch cadence, satellite constellations, and autonomous operations are engineering problems at scale. Quantum is moving from physics department curiosity to cloud API — another substrate where the people who understand computation will be first in the room.

These are not separate futures. They are one exponential stack.

And across all of them, the same thing is happening: the grunt work is going away.

The boilerplate. The scaffolding. The third rewrite of the same integration. The status report that takes longer to format than to think. The ticket queue that exists because humans were the only available routers. I felt this viscerally over the holiday break when I wrote about the software development lifecycle breaking — the moment when delegation became real and the bottleneck shifted from writing code to judging code.

That is not the end of engineering. It is the end of engineering as busywork.

Jevons paradox: why cheaper intelligence creates more work

Here is the economic idea that ties the whole room together.

In 1865, William Stanley Jevons observed something counterintuitive about coal-efficient steam engines: as they improved, coal consumption did not fall. It rose. Greater efficiency made steam power economical for new uses, which increased total demand beyond what savings alone could offset.

Jevons paradox is not a trivia answer. It is the lens for understanding why AI does not shrink the software economy — it inflates it.

Aaron Levie made this case directly in March 2026: companies outside tech are realizing they can finally afford software projects that were previously out of reach, because AI lowers the cost of production. Marketing teams at big enterprises will have engineers automating workflows. Life sciences and healthcare will automate research pipelines. Small businesses will hire engineers for the first time to build digital experiences that used to require a funded startup.

His punchline is the one I keep repeating to parents: as long as agents still need a human who knows what to prompt, how to review, how to recover when the agent goes off the rails, how to maintain the system, how to fix the bugs — the people who understand software are not obsolete. They are scarce.

Cheaper tokens do not mean fewer builders. They mean more software surfaces — more custom tools, more internal apps, more agent workflows wrapped around operations that never had an engineering budget before.

Efficiency increases consumption. It has every time.

The data from the timeline: Levie, Lenny, and the thread underneath

Levie was not posting in a vacuum. He was responding to a labor-market conversation that Lenny Rachitsky kicked off the same week: engineering job openings at their highest level in over three years — more than 67,000 globally, 26,000 in the U.S. — with the increase accelerating since the start of 2026.

Lenny’s broader “state of the product job market” thread noted the same pattern for product management — openings up, not down, despite the doom headlines. The replies and quote-tweets split into two camps, as they always do: one side saying the data is lagging and AI will crater demand any quarter now; the other saying the composition of roles is changing but the headcount signal is real.

I align with Levie’s camp, with eyes open. The demand is not for the same work. It is for more software, built differently — agent-native, outcome-shaped, maintained by smaller teams with higher leverage. Lenny’s numbers do not prove AI is creating jobs in some moral sense. They prove that capability and hiring have not decoupled yet — and that the market is repricing what “engineer” means faster than university curricula can.

Follow that thread far enough and you land on harder data.

Anthropic’s labor report: real exposure, quiet hiring shifts

Anthropic’s March 2026 paper, Labor market impacts of AI: A new measure and early evidence, introduces observed exposure — a measure that combines what LLMs could do with what they are actually doing in professional settings, weighting automated uses more heavily than augmentative ones.

The findings are nuanced, which is why they are useful:

• AI is far from its theoretical ceiling. Claude covers about a third of tasks in the Computer & Math category that models could theoretically handle. The red area has a lot of room to grow into the blue.
• The most exposed occupations today include computer programmers (~75% observed coverage), customer service representatives, data entry keyers, and financial analysts.
• There is no systematic unemployment spike for highly exposed workers since late 2022 — yet.
• There is suggestive evidence that hiring of workers aged 22–25 into highly exposed roles has slowed — roughly a 14% drop in job-finding rates post-ChatGPT, barely statistically significant but directionally consistent with what recruiters are saying out loud.

Read that carefully. The senior engineer with ten years of context is not getting replaced overnight. The on-ramp is what compresses first. The entry-level ticket-queue job. The first-year analyst who used to learn by doing the grunt work that agents now absorb. Anthropic’s own economists are careful not to overclaim. I will be careful too. But the sectors in that report are real, and the exposure scores map to what I see in enterprise deployments every week.

Jobs are not vanishing in a flash. Functions are being reborn — narrower at the bottom, wider at the top, with judgment, orchestration, and domain knowledge rising in value as execution commoditizes.

That is not comfort food if you are 22. It is a call to adapt if you are anyone.

Access to custom software is exploding — transformation barriers are real

Here is the enterprise version of Jevons paradox.

Building software has never been cheaper. A motivated product manager with Claude Code and a clear spec can produce in a weekend what used to take a three-person team a sprint. Internal tools. Workflow automations. Customer-facing prototypes. The access curve has bent vertical.

But access is not adoption. And adoption is not transformation.

I have watched the same pattern in every cycle since client-server: the tool arrives years before the organization is willing to rewire around it. Legal blocks the API key. Security wants a six-month review. The CIO says “pilot” and means “sandbox forever.” Middle management protects the headcount that owns the old process. The SaaS contract has eighteen months left and a seven-figure exit fee.

Custom software is getting cheaper to create. It is not getting cheaper to change people.

That gap is where most companies are stuck right now — excited individuals, immobile institutions.

The product operating model is not optional anymore

For decades, non-technical and IT-only businesses could survive with a project mindset: buy software, implement software, hire consultants, run IT as a cost center, ship quarterly releases, repeat.

That model is breaking.

When software production costs fall and agents can sit in every function, the companies that win will not have “IT” and “the business” on opposite sides of a wall. They will operate as a product organization — outcomes owned by cross-functional teams, roadmaps driven by feedback loops, engineering embedded in marketing and finance and operations the way Levie describes.

I have recommended Marty Cagan’s work on this for years — it showed up in my favorite reads of 2024 — and the AI era makes it mandatory, not aspirational. Non-technical businesses will not have a choice. Either they become product-led — with agents and engineers in every function — or they become margin fodder for competitors who did.

Traditional IT departments that stop at service tickets and change-control boards will not survive as configured. They need to build out agentic enablement: platforms, guardrails, identity, observability, cost governance for tokens, and the cultural permission to let domain experts ship.

Every function needs agents. Not one chatbot in HR. Agents in HR, finance, sales, support, legal, engineering, and the CEO’s staff — wired to systems of record, measured on outcomes, audited like any production service.

This is the organizational half of what I wrote in Transform or Be Transformed. The technical half is moving just as fast.

The SaaS reckoning: innovate or die

If you have watched public software multiples since early 2026, you have seen the other side of the coin. Wall Street coined “SaaSpocalypse” — and for once the melodrama is not entirely wrong. The S&P software index dropped roughly 20% in a single month. More than $1 trillion in software market capitalization evaporated in the weeks after agentic AI made it obvious that per-seat, UI-first enterprise software had a structural problem.

The high-flying multiples era — endless ARR, rule-of-40 worship, growth-at-any-cost SaaS valued at 18–21x revenue — is over. Median public SaaS EV/revenue multiples have compressed to roughly 5–7x. That is not a temporary risk-off mood. It is the market repricing what software is.

Gartner’s analysis puts hard numbers on the threat: up to $234 billion in enterprise application spending could be exposed by 2030 — about 20% of SaaS spend — as AI agents bypass human users and interact with business systems directly. George Brocklehurst, a Gartner managing vice president, put it plainly: “You are no longer buying software primarily for people; you are increasingly buying it for agents.”

For two decades, enterprise software was evaluated on interface, workflow, training, usability — the human experience of clicking through a product. When agents become the primary user, that layer of value depreciates. Gartner calls the shift less an apocalypse than a metamorphosis. I agree with the biology. I also think a lot of vendors are about to find out which end of the metamorphosis they are on.

The mechanics are straightforward:

• Per-seat pricing assumes humans click. Agentic workflows break the link between headcount and licenses. A sales org that needed 100 CRM seats may need 50 when agents do the data entry — and the CFO knows it.
• Interface value deflates. Dashboards, ticket routers, form builders, lightweight project tools — anything that was mainly a conduit for human workflow. Containers of record (ERP cores, system-of-record data layers) hold value; conduits do not.
• Generic horizontal SaaS compresses fastest. If your product is a wrapper around a workflow an agent can execute autonomously, your moat was always thinner than your NDR suggested.
• Buyers are reallocating budgets. IT spend is not shrinking — Gartner still projects total software spending to grow roughly 14.7% to $1.4 trillion in 2026 — but dollars are moving from application shelfware toward AI infrastructure, agent platforms, and outcome-shaped builds.

This is the paradox the whole essay keeps returning to: software is not dying; the old SaaS playbook is. Total spend up. Per-seat economics broken. Valuations bifurcated — platform companies with system-of-record status trading at a premium, horizontal workflow SaaS at a discount.

SaaS is being forced to reinvent itself. Not a gentle pivot. Innovate or die. Bolt on an AI feature and call it Agentforce is not a strategy; it is a prayer. The survivors will move from seats to outcomes, from UI to API, from human-first to agent-first, from “we own the workflow” to “we own the data and the governance layer agents need to act safely.”

The days of endless ARR on autopilot — raise prices 8% a year, add a SKU, watch the multiple expand — are finished. What replaces them is harder and more interesting: software that competes for labor budgets, not just IT budgets. Software priced on results. Software that earns renewal because the agent stack depends on it, not because switching costs trap a human admin.

SaaS is not imploding. It is being forced to evolve — and the companies still running 2021’s playbook are not going to like the valuation of 2027.

Rebuilding the software development lifecycle

All of this lands in engineering first.

The SDLC you learned in school — requirements doc, design review, sprint, QA gate, release train — assumed that writing was the expensive step. That assumption is now false.

What I described in The Holiday Break That Broke the Software Development Lifecycle is becoming normal: delegation at scale, parallel agent threads, orchestration as the craft, code review as the bottleneck, judgment as the scarce resource.

Every engineer needs to be spending their weight in tokens — not as a flex, as a unit of production. Token cost is the new cloud cost. The CFO who understood EC2 line items in 2012 needs to understand inference budgets in 2026. FinOps for GPUs is table stakes; FinOps for cognition is next.

The teams pulling ahead are not the ones with the best model. They are the ones with the best harness — evals, permissions, memory, rollback, human-in-the-loop gates, and the discipline to productionize agents instead of demoing them forever. That is the same gap I called out in The Great Corporate Efficiency Reckoning: pilot to production is an organizational problem dressed as a technical one.

Functions need to be reborn. The SDLC needs to be rebuilt. IT needs agentic enablement. And everyone — whether they like it or not — is already in a transformation story.

Plenty of room to run

Come back to that math classroom in 1956.

They did not know what they were building. They argued about chess. They looked up “heuristic” in a dictionary. McCarthy lost the attendance sheet. Nash tested people on the stairs. And yet — artificial intelligence became a field, a market, a civilization-scale project.

We are still early in that curve’s 2026 chapter.

The grunt work is going away. Good. It was never the job — it was the tax we paid because machines could not yet carry it.

Software is getting cheaper to produce, which means more of the world will be software. Engineering openings are up. Exposure is real. Entry-level paths are narrowing. SaaS is repricing. Organizations are stalling or sprinting.

The parents who ask me about computer science are really asking: will my kid be a victim of this wave?

Not if they build for orchestration, judgment, and compounding tools. Not if they understand that agents amplify people who know what to ask. Not if they treat transformation as a permanent state — the same lesson I drew from accelerating returns: the winners compound adaptation as fast as the machines compound capability.

Dartmouth was a summer of argument that launched a seventy-year project.

2026 is the summer of deployment — AGI on the horizon, robots in the warehouse, models in the lab, quantum in the cloud, tokens on the balance sheet.

There is still plenty of room to run.

Don’t be a victim. Build.

---

Further reading: Dartmouth workshop · Ray Solomonoff’s Dartmouth history · Jevons paradox · Aaron Levie on Jevons paradox · Lenny Rachitsky on engineering openings · Anthropic labor market impacts