Machines That Understand Humans

How and why a computer understands things, and how it might understand humans — including how this shows up in science fiction. "How would you create a computer that understands humans" is the wrong objective. The correct objective is analogous to ChatGPT: to predict the next word really well, you need to be really smart and understand the world. At the end of a crime drama the text reads, "and the killer is ______" — a truly intelligent system does this. ChatGPT proved the pattern (intelligence fell out of next-word prediction); the open question is what the analogous objective is for humans, and whether the data for it even exists. Representation is a means to an end: human representations will be formed through a computer's alien self-learning to accomplish an objective, and after, you could direct those representations to other general tasks. If you can represent it, you can simulate it, and if you can simulate it you can generate it.

The history of personal computing, and the core of how we use computers to help us. Recommender systems (ChatGPT is in some way a recommender system), and their holy grail: hyper-personalization through perfect serendipity. Advertisements, ChatGPT, personalized AI (i.e. AI therapists), AI memory, matchmaking (Boardy), human simulations. Much of our interaction with software is already personalized, and it is only becoming more so. It's still day 1.

Recommender systems — content-based to collaborative filtering, Wide & Deep, TIGER and deep neural networks, generative recommendations, scaling laws. The No Free Lunch theorem, the theoretical upper limit of embedding-based retrieval, federated learning, AI memory (comprehensive overview), generative agents, generative user models — and the core gaps.

Data is upstream of everything. An exhaustive account of what data on us exists — genetic, descriptive, action. Tabular vs. token. Data as fossil fuel: what exists, how much of it, and what we might need to create. Alignment and structured ontology vs. aligned meaning. It is interesting that global text has signals a computer can learn about humanity from — what other data exists in the world that we can create a solution for?

Probably the most important section, and it deserves rigor — because we don't actually know the answer. Verifiability. Shallow objectives and the failure case of the wrong objective: slop. Some candidates: matching, recommendations, personalized AI, next-action prediction, and simulation. Simulation feels like the abstract version of computers that understand humans — maybe it gets solved first, or maybe you need to precisely understand each individual human that composes the abstraction. Maybe each individual is difficult to predict, but on the order of populations, it averages out.

The failure of purely probabilistic systems, and what we can learn from the success and failure cases of non-LLM foundation models.

A deep examination of existing approaches. Model layer: Simile AI, Unbox AI, Humans&, Aaru. Data layer: not sure. Dev tools: the AI memory players, Shaped AI. App layer: many of the personalized AI players — and notably, much of large tech exists as the app, data, and model layer at once. Walled gardens.

Assuming you solve the data and objective problem — on the surface, much better product recommendations. But we should not view ChatGPT as better word prediction. What happens in that process is you have developed a machine that understands the world; here, you would have a machine that understands humans.

Data should be viewed as relatively unchangeable. The one thing you can reason about is the objective — and the historical objectives we have built systems around are quite trodden. What might a new objective look like?

The question with ChatGPT turned away from what you could do with better word prediction, toward what you could do with an approximation of intelligence — with high-fidelity representations of the world. Could you do the same with not only an approximation of understanding humans, but actually understanding humans better than any human?

I simply do not know, and it is impossible for any human to know. Self-understanding itself is one of the sharpest pain points of the human condition. Perfect search and relationship simulation. The perfect therapist, the perfect mentor. Complete loss of human agency. Matching you to the perfect person on earth, the perfect job — and telling you why. Bad actors, addictive programming, market incentives. Roughly: if aliens arrived and they understood us better than we understand ourselves, don't you think that deserves examination?