AI will not be automating your job any time soon. Here’s why.

GPT-4 is not going to start automating your job any time soon. Neither is GPT-20, nor GPT-40, nor any other LLM.

At Sharestep we foresee a different long-term scenario. Here’s why.

What does not yet exist cannot now be known. The future is imagined by each man for himself and this process of the imagination is a vital part of the process of decision. But it does not make the future known. — G. L. S. Shackle

It often feels like human knowledge is a close-to-complete project — something like a built house that just needs the finishing touches of decor. But this is an illusion. Most things that can be known, are not known. And the unknown comes in two flavours: what can be arrived at by interpolation, and what cannot.

An example of interpolation — pick two words completely at random (e.g. “fox” and “baseball”), and find something that’s similar between them (e.g. “they both travel distances”). You just created new knowledge by interpolation. Granted, this knowledge is not likely to be very valuable, but most knowledge isn’t (until sometimes, by chance, it is).

New knowledge created by interpolation can be generated endlessly. GPT-4 and similar systems are very good at interpolation. For example, if you ask ChatGPT, “what are the similarities between a frog and Stamford Bridge stadium”, it’ll do a very good job.

Frogs undergo metamorphosis, changing from a tadpole to an adult frog, which signifies growth and development. Similarly, Stanford Bridge stadium has undergone several expansions and renovations since its opening in 1877. — GPT 4

However, most knowledge that is valuable cannot usually be arrived at by interpolation. There are two reasons for this, one related to the nature of the future and one to the present.

First, the future is random and mainly unforecastable. To test an example, write down a detailed list of 10 tasks that you will need to complete in exactly one week. Be adventurous and don’t pick things like “brush teeth” or “make coffee”. Put the note away and revisit it in a week, and check how well your forecast matched reality.

But what about the present? It turns out that the number of possible novel observations in the present is infinite, and cannot in general be interpolated.

Think, for example, about this task: “there is a garden behind 36 Southwater Road in the town of Saint Leonards in England. Two trees are planted there. Of what species are they?” This task cannot be solved by any conceivable intelligent being with interpolation alone. You must interact with the world to solve it.

Most human work, in the case that it is valuable, is focused on this kind of task: active, tunneled discovery in an infinite and novel observation space. Value is mainly derived from those things where we must force our environment to provide us with completely novel information — we must use agency to force new information from the environment.

In other words, most valuable human work is not like the example we tried earlier, where we picked two arbitrary words and interpolated a similarity between them. Most valuable human work lies in dealing with new unknowns that are latent everywhere — for which a process of targeted rediscovery and action is required.

It takes some careful thinking to realise that most things we do, when they are of value, are like this. Most valuable tasks are like the one I mentioned: going to Saint Leonards to discover the species of the two trees in the garden. When we are not solving tasks like these, and instead we are interpolating, it is much rarer to discover real value in what we’re doing.

An intelligent system that could match this capability would not just need to have advanced cognitive and physical agency. It would need to be capable of physical agency at almost every location on earth (and maybe elsewhere in the universe).

Reinforcement learning, twinned with robotics, has, theoretically, a future in that game, while LLMs don’t. But who in their right mind would sink capital into this capability, when we have roughly 8 billion agents doing the work for us already?

That is the question that interests me.