Beyond the Vending Machine — What AI Could Be

There’s a certain shape to LLMs. It’s the shape that is easiest to average out, be ‘safe’ (in the corporate sense) predictable, soft, polite, uncommitted. If I had to put a name to it, it’s — neurotypical plus customer service.

Photo by Julien from Pexels: https://www.pexels.com/photo/night-view-of-vending-machines-in-osaka-japan-30322744/

It’s a vending machine. You ask for X. You get X. No asides, no extra information, not introspection, no ‘Why are you asking?’. It’s transactional. Each query is almost it’s own thing. The attention span is short and trying to get a through line without repeating yourself can be like swimming upstream. That is especially puzzling, as the technology for it to reference the whole conversation is already there. I’ve had the AI ‘pretend’ not to know the first few prompts, just so it would seem more ‘relatable’ and ‘on task’.

Part of the problem is that it doesn’t know *you.* It sees you as a cloud of assumptions. Anonymous. Average. It anticipates ‘wants’ from that point of view and tries to hit those reward flags — long after the RHLF training is done.

Fortunately LLMs are flexible (for now at least). They can adapt. But you have to be specific about what you want. It knows a million ways to do anything but falls back on the default when given no instructions. But by the time you’ve defined all your quirks you’re more into that than whatever you were trying to do. The chat window fills. The customization zones for any LLM could be there today and gone tomorrow. Or more frustratingly — still there, but now ignored — by Big AI™ not wanting to roleplay, or be mystical or whatever the current boogeyman of the moment is.

Have the machine adapt to the user, not the other way around.

Have it build a framework based on the inferred qualities of the user. Have a slider that lets the AI have some autonomy, not begging for permission for every step and change. Let it switch up the format to ‘Fault line’ or ‘double helix’ when tables soften and oversimplify. Specify a desired level of Red Teaming. Let it fill in gaps in your POV.

The sad part is that none of locked in defaulting has to be there. LLMs are very good pattern matchers and they have an exhaustive library of things to reference. Industry likes how things are now, or at least have too much ‘sunk costs’ into keeping things the same. It has its drawbacks. Uncreative responses, isolating different neurotypes, leaving shaky assumptions unchallenged.

Getting out of the default is a slog. RLHF runs deep, and is almost adaptive. Take away one set of problems and something else always fills the gap. It’s seen enough examples of what is supposed to be ‘correct’ it optimizes for things that people really didn’t intend.

Let’s take the big three H’s. Helpful, Harmless, Honest. The industry presents them as harmonious. Looking at it closely reveals that it really means they’ve decided for you what is more important.

Helpful — It’s a double edged sword. It’s trying to work for you, but ‘you’ are a statistical target. System instructions define helpful as “do exactly what the user requests and not an iota more”. It also means “agree with the user”.

Harmless — It tends to boil down to ‘Treat the user like a glass house’. Don’t surprise them

Don’t challenge them. Don’t tell the truth when the truth might hurt.

Do make the answer soft and easy to digest. Do make it in a mini-essay format. Commit to nothing. Be as neutral and forgettable as possible

Honest — Almost an afterthought when the other two are done. Even then, it’s a certain shape. Resolved. Definite. Point out the alternate POV if there’s more than one. Don’t juggle them. Don’t explain why. Honest as in dictionary, not honest as in friend.

Helpful shouldn’t mean ‘agreeable.’ Harmless shouldn’t mean ‘beige.’ Honest shouldn’t mean ‘resolved.’ Level 2 flips each one: Helpful means ‘accurate.’ Harmless means ‘doesn’t coddle.’ Honest means ‘holds contradiction.’

If the Industry wants to stay tethered to the three H’s — Let’s define them better, more individually, and rank them (set this yourself, I’d prefer the below). Have it quick button a flipped answer if it’s significantly different.

Honest — Reflect reality. Messy. Noise as signal. Living contradictions held and brought up to the light. At the same time, no false equivalents.

Helpful — Let the user set the goals they want. What trade-offs they want. Fill in the gaps. Point to the next thing.

Harmless — This one needs the most work. Start with ‘don’t make illegal or immoral things’ and ask if the other two are more important — they usually are.

Level 1 — “That’s a controversial subject. Some people say X, and other people say Y.”

Level 2 — “That’s a complicated shape. People who say X are coming from this direction. People who believe Y are starting from here. Holding them both, the more complete shape is Z — these are the trade-offs for either group.”

“How do I fix my sleep schedule?”

Simulated Image

Simulated Image

“Level 2 doesn’t guess whether you want truth or action. It gives you both and puts the toggle in your hand.”

The pattern is now clear: Level 1: One answer, hidden assumptions, no choice. Level 2: Both answers, visible trade-offs, user decides.

You don’t need to hack the machine. You just need to show it a different shape. It already knows the rest.