AI discovery methods · partial framing snapshot

# WHAT AI ACTUALLY CHANGES

*The frame for the 100 entries, and the reason 10 of them are negative*

---

## The claim being tested

"AI could discover X" is usually a statement about *possibility*, not about *mechanism*. The useful
version is narrower:

> **A discovery becomes feasible when its bottleneck is search, and search becomes cheap.**

That is a real and large class. It is also a *bounded* class, and most of the disappointment in this
field comes from mistaking the boundary.

---

## Where the bottleneck is search (AI helps, a lot)

| Bottleneck | Why AI relieves it | What it still cannot do |
|---|---|---|
| The literature is too large to read | Embedding retrieval, table extraction, cross-lingual search | Judge whether a paper is right |
| A design space is too large to enumerate | Generative models, Bayesian optimisation, active learning | Synthesise the candidate |
| One evaluation costs hours of compute | Surrogate regression | Extrapolate outside the training region |
| The inverse problem is ill-posed | Learned priors, regularised inversion | Recover information the instrument did not record |
| The pattern is invisible in high dimensions | Representation learning, clustering | Tell a real cluster from a sampling artefact |
| The experiment is slow and manual | Closed-loop automation | Discover something outside its search space |

In every row, the AI removes a *combinatorial* obstacle. None of them removes a *physical* one.

---

## Where the bottleneck is not search (AI does not help much)

- **Nothing has been measured.** No amount of learning recovers a quantity nobody recorded. The model
  will interpolate smoothly and be wrong with confidence.
- **The data are outside the training distribution.** Surrogates fail silently; the failure mode is a
  plausible number, not an error message.
- **Causality is required.** Observational data identifies associations unless an intervention, an
  instrument, or a strong assumption supplies the direction.
- **The claim needs a proof.** A model can suggest a theorem; it cannot certify one.
- **The claim needs a synthesis.** A generative model proposes a molecule; a chemist makes it.
- **The effect is below the noise floor.** A network that "denoises" below the floor is inventing
  structure, not recovering it.

These are the ten negative entries in `08_H_LIMITS`, and they are there because a document with
ninety optimistic entries and no negatives is a marketing artifact, not a research plan.

---

## The honest asymmetry

The two lists above are not symmetric in difficulty. The positive ones depend on **compute and data**,
both of which improve. The negative ones depend on **information theory and thermodynamics**, which do
not.

That is the whole argument in one line: *AI makes search cheap, and search is not the same thing as
knowledge.* A discovery that needs a better search is now cheap. A discovery that needs a measurement,
a synthesis, or a proof is exactly as expensive as it was, and the model's confidence about it is
unrelated to its difficulty.

---

## What the 100 entries are

Ten themes, ten or fifteen each:

- **Search** (A) — mining archives and literature for relationships nobody has joined.
- **Inverse design** (B) — spaces too large to enumerate.
- **Surrogates** (C) — what becomes possible when a calculation is 10⁶ times cheaper.
- **Automation** (D) — closed loops, and what they optimise that you did not intend.
- **Translation** (E) — the same mathematics in two fields.
- **Symbolic** (F) — recovering laws and invariants from data.
- **Measurement** (G) — instruments with learned priors.
- **Limits** (H) — ten things that will not happen soon, and why.

Every entry names the bottleneck, the capability, **and the step that remains physical**. An entry
without the third is not in this collection.

---

## The one test that separates the real from the promotional

Ask: *if the AI were perfect at its stated job, what would still be missing?*

- If the answer is "nothing" — the entry is wrong, because something is always missing.
- If the answer is "a measurement", the entry is an experiment proposal wearing a model's clothes,
  which is fine and should be labelled as such.
- If the answer is "a synthesis" or "a proof", the entry is a design target, which is also fine.
- If the answer cannot be stated — it is promotional.

Applied to all 100 entries, that test is what this folder is for.

# v9 BRIEF — 100 discoveries that AI makes *feasible*, and 20 that it does not

The question is not "what could AI discover?" That framing produces slogans. The question is:

> **Which specific discoveries have a bottleneck that is *search* rather than *measurement* or
> *understanding* — and is that bottleneck one current AI actually relieves?**

Every entry must answer three things plainly:

1. **What is the bottleneck today?** (search space size, literature volume, inverse problem, cost per
   evaluation, human attention)
2. **What does the AI do about it?** (a specific capability: surrogate regression, generative inverse
   design, embedding search, symbolic regression, active learning, automated experiment)
3. **What does the AI NOT do?** (the physical step that remains: synthesis, measurement, proof,
   validation, or the causal inference that no model supplies)

If (3) is missing or trivial, the entry is hype and must be rewritten.

---

## Format (200–260 words)

```
### A01 — <the discovery, stated as a result>
**What becomes possible.** <one or two sentences: the concrete finding or capability>
**Bottleneck today.** <why it has not been done: size, cost, attention, inverse problem>
**What the AI supplies.** <the specific capability, not "AI" generally>
**What still needs the bench.** <the physical, experimental or proof step that AI cannot do>
**First concrete test.** <a number, a small experiment, or a dataset that would show it works>
**Hype risk.** <the most likely way this entry is wrong or oversold>
```

## Sections

| Folder | Theme | Count |
|---|---|---|
| `01_A_SEARCH` | Mining archives, literature, and unpublished data for missed relationships | 15 |
| `02_B_INVERSE` | Inverse design over spaces too large to enumerate (materials, molecules, proteins, metamaterials) | 15 |
| `03_C_SURROGATE` | Replacing expensive simulation with learned approximations, and knowing when they fail | 15 |
| `04_D_AUTOMATION` | Self-driving laboratories, adaptive experiment, closed-loop characterisation | 15 |
| `05_E_TRANSLATION` | Finding isomorphic structure between fields (same equation, different subject) | 10 |
| `06_F_SYMBOLIC` | Equation discovery, invariant finding, and the recovery of laws from data | 10 |
| `07_G_MEASUREMENT` | Instruments that get better with learned priors (calibration, denoising, super-resolution) | 10 |
| `08_H_LIMITS` | **Ten things AI will not deliver soon**, stated as honestly as the positive entries | 10 |

Total: 100.

## Hard rules

1. **No "AI will revolutionise X."** Every claim is a specific result.
2. **The physical step must be named.** If the discovery needs a synthesis or a measurement, say so.
3. **No fabricated citations, datasets, results, or benchmark numbers.**
4. **Hype risk is mandatory** and must be a real failure mode, not false modesty.
5. **Section H is adversarial.** It exists because a document with 90 optimistic entries and no
   negatives is a marketing artifact.
6. **No prohibited subjects:** human demographics, social constructs, politics, economics, climate,
   ecology, marketing.
7. **Distinctness:** no two entries share a core mechanism or a core AI capability argument.