Latent Space as a New Medium

I thought so too, but the results are too consistent. It never combines foo and bar to make "Certainly! It's foobar" or "I can't do that" or "foobar. Would you like anything else?"

>In other words, correctness, truth, cohesiveness, completeness, comprehension, etc are all essentially patterns that are mapped in this space.

The caveat is that truth in latent space is just a reflection of the consensus from the corpus, and you find truth by comparing the answer in latent space to what is in reality.

But I just hate this idea that truth and facts are no longer real, they're just "directions". The more we rely on these models for our lives, the more we lose touch with reality and are pushed and pulled in all these different directions. Feels like the future is just ai psychosis and there's no way out. Is that what complete agi victory looks like?

I think virtual reality and AI are a prelude to humanity living in Hong Kong style coffin apartments, and things will get progressively worse until we're in Matrix pods.

This article could be compressed.

This article starts off well enough and covers some things I had been thinking about myself: exploring the many-dimensional space of an LLM. But then I think it veers into good old human hallucination (buzz-words disconnected from realities).

It even mentions my favorite show "Connections," but gets it wrong: it wasn't showing obscure or unlikely ideas, it was highlighting under-appreciated historical paths of developmnent. The one I remember (and have seen repeated on youtube since) was how the lathe boot-strapped the industrial revolution, crude ones at first helping to build more accurate replacements for themselves.

I want to give the author some credit for exploring some of the new possibilities that the LLM constructs offer. But the second half just veers into unreal territory. Typically, the author takes the words "latent space" and divorces them from their actual meaning to just make up stuff.

"Retro latents": nobody except historians will want to revisit the uncanny valley.

"simulate all possible latent spaces": that's just n-dimentional random matrices, for n trending to infinity, computationally infeasible. "computationally sweep through the space of all possible latent spaces, in a sense mapping the nature of latent space itself": no for the same reason. A "latent space" is only significant because it is based on a reality. It makes no sense to us to make a random "latent space" and ask what reality that corresponds to.

"create simulations of various propositions by moving through the latent space": there was already a documentary about this, it was called "The Matrix."

"Latent spaces contain all parallel worlds": if properly constructed and queried, it seems they would really only contain our reality. A parallel world would be a different "latent space" (ie big matrix) with different values.

The "Personal latent space" does make me think there might finally be a way to "download" our brains into silicon, but you'd really have to train it with google glasses running 24/7 from birth, and nobody wants that. Plus, it would be a security nightmare: unless you could carefully guard all access in an air-gapped container only you control (a bit like your brain in your body), you could be duplicated and your virtual duplicate could drain your bank accounts, be easily blackmailed, steal your IP, etc.