Σ 0:00▸ 0:00

Module 2 · Building Agents · scripted

The Day After: Predicting a Research Opportunity

by Jules White

90 min
🎤 Talk view← Course

Overview

The day before reading and writing, human knowledge was bounded by human memory. The day after, it was bounded only by the resources we could put into clay, scrolls, and paper — and that change gave rise to libraries, to the internet, and eventually to AI itself. This session asks you to take the parallel seriously and predict what it means for your field.

You do not start from a blank page. You already did two pieces of research for this course — one on what was predicted for an industry and what actually happened, one on what is happening to the cost of intelligence. Today you use that evidence to discipline your own predictions, then build rough prototypes to test whether the things you are predicting are already easy. You will work in teams.

Everything you make today joins what you already made: one master deck, presented as a whole.

The thought experiment

Imagine a world where all knowledge work — programming, data analysis, writing, synthesis, hypothesis generation, planning — is near zero cost and can be scaled up dramatically.
What are the freed resources, and what do they allow?
What is the opportunity?
What are the new bottlenecks in your domain?
What are the new skills we have to learn?
What remains scarce — and therefore valuable?

The session at a glance

  1. 1
    Mine the two research questions you already answered — the misses tell you how forecasts fail here.
  2. 2
    Form teams and run the thought experiment against your domains.
  3. 3
    Work through the seven parts below.
  4. 4
    Build rough, human-in-the-loop prototypes of one or two of your ideas — quick feedback, not experiments.
  5. 5
    Assemble one master deck: your earlier slides set the context for today's research direction.

The core mechanic: value migrates

Twice before, a human capacity became suddenly, dramatically cheap — and both times, value behaved the same way:

When a capacity becomes free

writing made memory cheapmemory stopped being the prize — and ORGANIZATION became the skill: searching, indexing, managing information at scale. Libraries, catalogs, and citation are scarcity-management inventions
computers made arithmetic freecalculation stopped being the skill — and SPECIFICATION became the skill: stating the problem, designing the program, building the workflows that harness the capacity

When a capacity becomes free, value does not disappear. It MIGRATES — to the complements, to whatever the free thing cannot do, to whatever remains scarce.

Your job today:find where it migrates in YOUR field.

Ten thousand assistants — write one sentence each

  1. 1
    Your lab just hired ten thousand tireless, brilliant assistants who work for electricity. What does your field do next Monday?
  2. 2
    What breaks by Friday?
    Who reads ten thousand literature reviews? Who checks them? Who decides which of ten thousand hypotheses gets the one wet-lab slot?
  3. 3
    What still takes a decade — and why?

Part 1 · What you already found

You answered two research questions before this session. Put both sets of findings on the table now — they are the evidence base for everything that follows.

The two questions you already answered

One — the prediction and the outcome. What was being discussed and predicted about the impact of AI on an industry of your choice in late 2023 and early 2024? Then: one or two scholarly papers or concrete data sets from 2026 (or late 2025) that support or refute that early reporting. What actually happened versus what was predicted?
Two — the cost curve. Cost versus benchmark performance of AI models over the last six to twelve months. What does the trend line look like? What does it say about the claim that agents carry large cost and unknown value? What did you predict comes next, and which research gaps have to be filled?

You made those predictions with a method. Now audit the method.

Reflect on these questions

  1. 1
    What was predicted that did not happen. Name the specific forecast that missed. Then the interesting part: WHY did it miss — was the capability wrong, the timeline wrong, the adoption path wrong, or the assumption about what the work actually consists of?
  2. 2
    What happened that nobody predicted. Name the thing your 2026 sources show that the 2023–24 discussion never raised at all. A blind spot is worth more than a wrong number: it tells you which kind of thing this discourse cannot see.
  3. 3
    What the cost curve is doing — and what it makes true on a schedule. Read your trend line forward. If capability per unit cost keeps moving the way you measured, what becomes economically obvious in a year that is merely interesting today? What stays expensive no matter what the curve does?

Why this comes first

The gap between what was predicted and what happened is helpful data about how forecasting fails in this domain — and you are about to forecast.
Every projection you make in Parts 2–6 should survive the failure mode you just identified. If your team's finding was "everyone overestimated speed and underestimated integration cost," then a prediction of yours that assumes fast frictionless adoption is already suspect.
The cost curve does the same job from the other side: it turns "someday" into a date, and it tells you which bottlenecks the curve will dissolve on its own — those are the ones not worth researching.

Part 2 · The three that survive

Discuss three things in your domain that become more valuable — or stay valuable — after intelligence is cheap. Create one to three slides presenting these things, and why.

Part 3 · The three where you'd spend it

Name three fundamental problems in your domain where you would apply this scalable cognitive capacity through agents — and why those three. Then step back and ask the more interesting question: what patterns do your three share? The same shape of problem? The same bottleneck being attacked?

Part 4 · The missing skills

What skills, techniques, methodologies, or processes that we don't have today — or don't have enough of, or don't understand well enough — become critical for your field to scale and apply this capacity? Name two or three.

Part 5 · The research direction

Using your insights from Parts 1–4, design a research direction and three initial research questions. Not a topic that uses agents — questions about how your field harnesses the capacity given what remains scarce, or questions that would help create the methods, techniques, and skills that are missing.

Part 6 · The measures

For your research direction: how would you measure success, and how would you measure cost, as the capacity is applied? Two or three specific, checkable measures of each — not "quality improves," but the number, the comparison, or the verdict you would actually collect.

Part 7 · Rough prototypes — is it already easy?

You have been predicting. Now spend a short block finding out how much of it a crude agent can already do. In Exercise: Give the Agent Tools you declared a tool set and played the computer yourself. Do exactly that again in the simulator or another LLM — but this time point it at one or two of the problems from Part 3.

What this is and is not

This is not experimentation. No baselines, no conditions, no results you would defend in a paper.
This is quick feedback to inform your own thinking. You are finding out, in twenty minutes, whether the thing you just predicted was five years away is in fact half-working on your laptop — or whether it falls apart in a way your prediction never considered.

Build one or two rough prototypes

  1. 1
    Pick one problem from Part 3 — ideally the one your team is least sure about.
  2. 2
    Declare the tool set: three to six tools, each with a name, a description, and parameters. The tools ARE your claim about what the work decomposes into.
  3. 3
    Run it and play the computer, answering each call as your world would.
    Use the Tool Playground · or a custom GPT · or plain prompting in the LLM of your choice · or the Agent Forge if you want the tools to carry real code.
  4. 4
    Watch for the three things worth reporting: where it got further than you expected · where it asked for something your tools could not say · where a human had to supply the judgment that made it work.

Capture — the prototypes

💾 Capture

Write down:

  1. 1Your tool set, verbatim, and the task you gave it.
  2. 2One surprise: something easier or harder than your team predicted, and what it does to your Part 5 research direction.

The deliverable — one master deck

One presentation per team, added at the end of you have already made. Your earlier research is not an positioning — it is the evidence that sets up the research direction, so it goes first.

What the master deck must contain

  1. 1
    Your two pre-work slide sets, first: the industry prediction-versus-outcome analysis, then the cost-versus-performance trend. Two to four slides, as you made them.
  2. 2
    Part 1 — one to two slides: what was predicted that didn't happen, what happened that nobody predicted, and what the cost curve makes true on a schedule.
  3. 3
    Part 2 — one to three slides: the three things that become or stay valuable in your domain, and why.
  4. 4
    Part 3 — one to two slides: the three problems where you would apply the capacity, and the pattern they share.
  5. 5
    Part 4 — one slide: the two or three missing skills, techniques, or methodologies your field will need.
  6. 6
    Part 5 — one to two slides: the research direction and three initial research questions.
  7. 7
    Part 6 — one slide: the measures — success and cost, specific and checkable.
  8. 8
    Part 7 — one to two slides: the rough prototypes. What you built, what surprised you, and what it changed.

The deck should read as one argument: here is what was predicted and what actually happened, here is what intelligence now costs, here is what that makes scarce in our field, here is where we would spend it, here is what we tried by hand — and therefore, here is the research.

Before we regroup

💾 Capture

Assemble your team's master deck.

Write down:

  1. 1The master deck.
  2. 2Your prototype capture: the tool set and the surprise.

Be ready to discuss:

  • The discussion is a hunt for convergence: across all the domains in the room, do the same scarce things and missing skills keep appearing? If they do — what does that say about the research agenda of the next decade?
  • Whose prototype most changed their own prediction? Did building the crude version make the team more or less ambitious?