Parker Ferguson's talks on AI and what it changes, as one map of evidence, open questions and history that he keeps up to date.
The featured talk
Five things that stay scarce, and what work becomes
When something gets cheap, value moves to whatever stays scarce. The talk names five things that intelligence doesn't make cheap: verification, context, atoms, legitimacy and direction.
Each one follows a single subject from then, to now, to next, and ends on a question. About 11 to 22 minutes each.
How fast this is moving, how you'd know, and why so few people can see it.
Short talk · 15 slides · 16 min · updated Oct 2026AI gets good fastest at work someone can check. The same fact explains how it cheats.
Short talk · 14 slides · 15 min · updated Oct 2026How people at the frontier set up their work today: the tools, the files they own, and many agents at once.
Short talk · 20 slides · 22 min · updated Oct 2026An agent that acts for you and remembers everything. Who controls it, and who owns what it knows?
Short talk · 16 slides · 18 min · updated Oct 2026Anyone can build software in a weekend. What is still hard, and where is the moat?
Short talk · 10 slides · 11 min · updated Oct 2026What happens to jobs and wages when labor becomes something you can buy.
Short talk · 13 slides · 14 min · updated Oct 2026Intelligence runs on things that don't scale like software, and whoever owns them collects the return.
Short talk · 15 slides · 16 min · updated Oct 2026AI can now attempt open problems. Producing results got cheap. Checking them didn't.
Short talk · 15 slides · 16 min · updated Oct 2026A few labs, a few states, or something more open, and what history says about each.
Short talk · 14 slides · 15 min · updated Oct 2026If the jobs move, the opportunity moves with them. Each window closes when a model learns what it sells.
Short talk · 13 slides · 14 min · updated Oct 2026The scarcities are the board. States, capital, the military, the titans, the Church and crime are the players, and new institutions decide which of them gets constrained.
Short talk · 14 slides · 15 min · updated Oct 2026Each theme gathers the slides on one subject from across the library and plays them as a short talk. The build puts them together, so they grow as the library does.
Knowing what good looks like, and standing behind it.
Theme · 9 slides · 10 min · updated Oct 2026Who shares the gains, and who steers.
Theme · 9 slides · 10 min · updated Oct 2026How long an AI can work on its own, and what those hours buy.
Theme · 4 slides · 4 min · updated Oct 2026A test with a way to check the answer, which becomes the training signal for the next model.
Theme · 5 slides · 6 min · updated Oct 2026High reward should mean the task was solved, not that the grader was fooled.
Theme · 3 slides · 3 min · updated Oct 2026A model with a real computer and a loop: plan, act, observe, check, repeat.
Theme · 3 slides · 3 min · updated Oct 2026Hundreds or thousands of agents in parallel, from research swarms to a billion agents running all day.
Theme · 8 slides · 9 min · updated Oct 2026The funnel, verifiability, the LLM wiki and the move up to oversight.
Theme · 5 slides · 6 min · updated Oct 2026From Erdős problems to Navier–Stokes and the 722-paper MathDrop.
Theme · 3 slides · 3 min · updated Oct 2026Wet labs, biotech deals and citizen science pointed at open problems.
Theme · 3 slides · 3 min · updated Oct 2026Results announced from models the public can't use, and the gap between inside and outside the lab.
Theme · 3 slides · 3 min · updated Oct 2026Real work consumers hand to agents today, and how to set them up.
Theme · 3 slides · 3 min · updated Oct 2026One-shot apps, SaaS clones and the internal-tool arbitrage window.
Theme · 3 slides · 3 min · updated Oct 2026Pick two of three: untrusted input, private data, external action.
Theme · 6 slides · 7 min · updated Oct 2026Who holds the memory, the data and the instructions your agents run on.
Theme · 4 slides · 4 min · updated Oct 2026Chips, grids and the falling price of intelligence.
Theme · 3 slides · 3 min · updated Oct 2026Tokens turn money into work, so labor and capital start to look the same.
Theme · 5 slides · 6 min · updated Oct 2026Dividends, history's broken bargains, policy races and unrest.
Theme · 5 slides · 6 min · updated Oct 2026Timelines that keep collapsing, and the scenarios people argue about.
Theme · 7 slides · 8 min · updated Oct 2026Review falls behind, so our job moves up to oversight and understanding.
Theme · 7 slides · 8 min · updated Oct 2026Every question the library leaves open, from your own work out to society.
Theme · 73 slides · 80 min · updated Oct 2026Revised "When thinking gets cheap": a new subtitle ("Five things that stay scarce, and what work becomes"), a Part 3 that builds from jobs to the five scarcities and the players before alignment, a short version that covers the scarcities, and a Part 4 that ends on one question for the room. Two slides on the size of the change became one.
when-thinking-gets-cheapIf you can buy work by the token, how big is that?Five things that stay scarceWe already built an optimizer we can't switch offand 1 more
Corrected sources and wording on several slides: the task-length chart cites the METR releases that hold each model's result, the price of the same capability falls about 13 times a year everywhere, labor's share is labeled as business income, and the jobs slide cites the Stanford study it quotes.
How long an AI can work on its own keeps doublingPrice per unit of intelligence collapses every yearIf you can buy work by the token, how big is that?Big forecasts, small but sharp effectsand 1 more
Revised Part 4 of "When thinking gets cheap": the two endings now come before what each of us can do, the closing screen asks one question in the same words the talk uses to hand over, and six of its questions now cover the transition and alignment. "How do we work now?" adds an error log, and "What's worth pushing for?" became a deep dive under "What institutions does AI need?".
when-thinking-gets-cheapWhat could work become?If the jobs move, where does the opportunity go?How do we work now?and 3 more
Corrected how three places describe OpenAI's o3 cheating on a METR test: 30.4% is the share of its runs in which it cheated in any way, for example by patching the scoring program to always pass, not the share in which it patched the scoring program.
Alignment isn't solved, and the labs say soIf every public test of AI breaks once models study for it, how do we know how good the models are?How would we know an AI is pursuing what we mean, not only what we managed to score?
Library created. The Troglodyte Society #3 talk, its threads, and its questions and lenses now live as one shared pool of content that any number of talks can walk.
Watch plays a talk on the map, where you can also see the whole library as one network. Read opens the same slides as plain pages that print and save as a PDF. Every slide says what kind of claim it makes and when it was true.