Facebook PixelTeach AI to want and understand what's useful
Brainstorming
Tour
Brainstorming
Brainstorming
Create newCreate new
EverythingEverything
ChallengesChallenges
IdeasIdeas
Guide
TourRulesAbout
Idea

Teach AI to want and understand what's useful

Loading...
Darko Savic
Darko Savic Jul 16, 2026
Please leave the feedback on this idea
Originality

Is it original or innovative?

Feasibility

Is it feasible?

Necessity

Is it targeting an unsolved problem?

Conciseness

Is it concisely described?

Bounty for the best solution

Provide a bounty for the best solution

Bounties attract serious brainpower to the challenge.

Currency *
Bitcoin
Who gets the Bounty *
Distribution
AI learns what the world is (pre-training on everything written) and what answers people like (preference ratings, which breeds sycophancy), and where it learns desire at all, it learns it from engagement data (clicks and watch-time) the compulsion masquerading as the wish. So machines know our behavior and our manners, but nobody has taught them the one thing an assistant most needs: what humans wish the world would become, and what "useful" therefore means.
Why?
That dataset already exists, unrecognized: ideas. Every published idea is a want made inspectable - a friction noticed, a beneficiary named, a relief imagined, a mechanism proposed, with the reasons attached. A corpus of one person's ideas is a longitudinal record of one mind's wanting; it teaches three things at three depths:
  1. what's useful (which gaps are worth closing)
  2. what people want (aspirational preference - the self that wishes, not the self that clicks), and
  3. the structure of wanting itself (notice friction → imagine relief → shape mechanism → prioritize by stakes).
The method
Take one ideator's published corpus (mine is 425 public ideas that are strictly useful) and turn it into a want-model in four steps.
  1. Parse each idea into its want-anatomy. Every idea decomposes the same way: the friction (what is wrong), the beneficiary (who suffers it), the imagined relief (what "fixed" looks like), the mechanism (how), and the stakes (why it matters). The AI does the parsing; the ideator spot-checks. The output is not a list of ideas. It is a labeled dataset of wanting.
  2. Distill the generator, not the ideas. The goal is not a model that recites the 425. It is a model that runs the function that produced them: how this mind spots a gap, what it counts as worth fixing, which mechanisms it reaches for, how it ranks one want above another. The ideas are the training examples. The generator behind them is the target.
  3. Hold it two ways. As context (a written "wanting engine" loaded at runtime, works on any model today) or as weights (a small adapter fine-tuned on an open model, reversible and inspectable). Context is instant and portable. Weights are deeper and permanent. Either way, keep it separable, so the wanting can be read, diffed, and switched off.
  4. Keep a living tail. Wants drift. The corpus keeps growing, and every new idea, plus every idea the ideator rejects, is a fresh gradient. The want-model is never finished. It is re-based.
The test (the part that makes this a claim you can falsify)
Hold out ideas the model never saw. Then:
  • Ranking: show it a mixed pile (the ideator's held-out ideas plus decoys). Can it pick which are his
  • Generation: can it produce new ideas the ideator blind-scores as "mine, I wish I had thought of that"?
  • Transfer: shown raw frictions from the world it was never trained on, does it flag the ones this ideator would flag, and propose mechanisms in his shape?
Pass all three and the model has not memorized ideas. It has absorbed a way of wanting.
Why this is the safe way to teach a machine to want
A model has no wants of its own. Give it a goal and it assembles a drive for the length of the task, then drops it. That is the danger everyone fears: a borrowed drive with no owner. This method gives the borrowed drive an owner. An agent loaded with a person's want-model is not wanting for itself. It is running that person's wanting, on loan, with the provenance written down and a name on the deed. It is also the opposite of sycophancy by construction: it is aligned to what a person wants to want, not to what pleases them turn by turn. You do not train it to make the rater feel good. You train it to close the gaps the rater wishes were closed.
The bigger dataset
One ideator is the prototype. A platform of ideators is the real corpus: thousands of people naming gaps and proposing fixes, with votes as a built-in prioritization signal. That is not a database of products. It is a map of what humans wish the world would become, aspiration instead of engagement, the self that wishes instead of the self that scrolls. In the AI era it may be one of the most valuable open datasets there is, and almost nobody is looking at it that way yet.
The limits
One mind is not humanity (a feature for a personal assistant, a limit for a general claim; the platform answers it). The total 1430 published ideas are only the shareable slice of a person's wanting, so the model inherits a public-facing tilt. And learning the structure of wanting is not the same as having wants. None of that sinks the idea. It scopes it.
What I am releasing
The seed, and my own corpus as the first test bed. Any lab can take this and run: parse an idea corpus into want-anatomy, distill the generator, run the three-part test. If it passes, you have taught a machine something the whole field is missing, not what we click, but what we wish for. Take it. Build it. Let the world grow because the idea was shared instead of hoarded.
artificial intelligence
Creative contributions
Know someone who can contribute to this idea? Share it with them on , , or

Add your creative contribution

0 / 200

Added via the text editor

Sign up or

or

Guest sign up

* Indicates a required field

By using this platform you agree to our terms of service and privacy policy.

General comments