The Prediction Machine

First in a series exploring thoughts and directions Goldfish could go.

When I was a kid, we had a family computer running Windows 95. On it was an application where you could talk to a "doctor" who would answer your questions. It was a primitive chatbot, probably matching a handful of keywords, but I spent hours trying to make it answer everything, convinced I was talking to a real being. I wanted so badly for it to think. It matched Weizenbaum's ELIZA, jut a lot more basic.

Thirty years later, the whole industry is chasing the same feeling: a machine you can just talk to. I think that's the wrong destination. The interesting question isn't how to make the conversation better. It's what happens when there's no conversation at all.

Interfaces leak

An interface is anything that lets a person interact with a system, physical, digital, or so implicit you don't notice it's there. The designer's job is to communicate what the system can do on your behalf. Call an interface lossy when meaning leaks out along the way.

Not every interface leaks. A light switch is lossless, the intent is one bit, and the channel carries one bit. A spoon is close to lossless too. Interfaces are lossy in proportion to how much of your mind they have to carry. Design's oldest trick is to keep that small: constrain what the user can want until every intent has a button. It works beautifully, right up until the intent space is unbounded.

A computer's is. A general-purpose machine has to carry arbitrary intent, and that's where the leaking starts. How much does it leak and what can we do to fix it?

Language is the bottleneck

Language is the most flexible modality we have, and the one we lean on for nearly everything, with each other, and increasingly with computers. It's also brutally hard. Taking a thought and forcing it into well-formed sentences loses something every single time. It's harder still in a language that isn't your native one, but it's hard regardless.

That's exactly why LLMs feel like magic: they make expression trivial. You hand over a loose idea, a rough direction, and the model predicts what you probably meant. The prediction is so convincing that we often accept it without question. But the real story is that LLMs inject a lot into our vague directions. Sometimes that's what you want, outlining an article, bouncing ideas, grinding through something boring. Sometimes you're just accepting someone else's guess about your own mind. And the worst version of this is when the averaged output influences your own thinking. Think about TikToks prediction algorithm, which is optimized for you spending as much time as possible on the application.

And is chat even the right paradigm? Chat is itself a lossy interface: it forces you to serialize thought into text before the machine can do anything with it. On the other hand, the force of describing a thought almost always qualifies whether the thought is worth anything. What if there could be another way?

We build in our own likeness

Look at how we've handled that unbounded intent space so far. The command line put all the translation cost on you: learn the machine's language, or nothing happens. Then Xerox PARC flipped it with the desktop metaphor, model the computer after our world instead. Files, folders, windows, a trash can: borrowed intuitions from the physical office, so the machine could meet us halfway. Chat is the next rung on the same ladder. Don't model the computer after our objects, model it after us. Avatars, voices, chatbots, my Windows 95 doctor. Each rung borrows more familiarity to close the intent gap.

But the top rung assumes other humans are a good interface. They're not. Think about every conflict you've had: almost none of it was about what anyone actually thought. It was about what got said, and what got lost in the saying. Most interpersonal conflict is translation loss.

We celebrate the people who beat this problem, writers, storytellers, journalists, precisely because low-loss language is so rare. Alan Moore, who wrote Watchmen, describes writers as wizards: a spell is just a string of words composed to make people see things. The people we once called shamans or priests were masters of telling stories. I.e., low-loss communication is possible.

So humans are a lossy model too. The ladder of likeness doesn't top out at "human", it tops out at a workaround. Every rung, from programming, command line, chatbot to mice and keyboards, is a different way of coping with the same problem: thought has to be serialized and translated before the machine can act on it.

The most intuitive interface is thinking

What's the lowest-loss way to interact with anything? Just thinking it. A thought doesn't even need to be fully formed; the moment it's slightly tangible, you have it. No serialization, no translation, no leak.

The honest version of that vision is a brain-computer interface. But you don't need an implant to get most of the way there. You need context. We can see this with Wispr flow, ChatGPT voice et al, where people dump as much context as possible, to get the best output. We also see that LLMs are excellent at extracting signal from what looks like noise. It's the exact same thing we see at Goldfish. I think we can infer that information from more subtle cues.

The prediction machine

You don't need to read minds to remove the conversation. Most daily communication is trivial and predictable, Slack messages to colleagues, LinkedIn replies, routine outreach. No negotiations, no hard problems. Given enough context about what someone is doing, their intent is obvious before they type a word. Prediction gets you most of the way to thought.

And prediction compounds. Something that can finish your Slack messages can eventually draft your emails; can eventually check your calendar and book meetings; reach out to people and produce artifacts. Follow the curve far enough and you get something always present, always absorbing context, anticipating what you're trying to do without you having to express it at all. Not because it's in your head, but because it's seen enough of your world to not need to be. And that takes care of 99% of the heavy lifting.

Us humans can spend time on thinking about the rest and creating what needs to be created.