Fable Made the AI Compute Boom Make Sense

A few days with Claude Fable changed how I thought about the demand for AI compute. Better models could make ordinary work cheap while making much bigger investigations worth attempting.

My first stretch with Fable lasted only a few days, ending when Anthropic suspended access in June following a US government directive. Anthropic later restored Fable access on July 1. What stayed with me was the impression from those first days, rather than anything I would call a full review.

The responses felt more insightful and patient than I was used to. Fable carried a thought well enough that I started wondering how much of my normal computer work I should still be doing myself. That was a subjective reaction to a short trial, but it changed the question I wanted to ask.

I also expect models with comparable everyday capabilities to become available locally over time, including through open models. I do not know when, or what compromises that will require in speed, context, or quality.

If that happens, what will all the giant data centers be for?

Cheap local intelligence could make larger ambitions practical

It is easy to imagine a laptop or phone handling most routine assistant work: writing emails, explaining documents, editing code, and helping with spreadsheets. I would welcome that, especially when it lets me work with my own files locally.

I suspect powerful clusters would then take on a different share of the work. Once ordinary tasks are inexpensive, we may become more willing to spend heavily on questions that need sustained investigation.

Many familiar AI uses start with a job someone already intended to do. A model drafts the memo, summarizes the reading, or writes the initial code. There is value in finishing that work faster.

The possibility that interests me more is being able to attempt a project that previously cost too much to consider. A small lab could investigate more hypotheses. A solo builder could discard more prototypes. A team could explore a difficult design question before committing to an expensive physical experiment.

That is one reason the compute boom makes more sense to me now. Better local models could raise our expectations of what computers should do, leaving us with more reasons to use the big machines.

Persistence can look like a breakthrough from the outside

There is an observation about magic that I love: sometimes the secret is the unreasonable amount of work behind it. A magician practices for years to make a coin disappear for two seconds. An effects team builds an elaborate rig for a shot that lasts less than a minute.

The audience sees the moment that works. Most of the effort remains invisible.

I wonder how often AI will bring that pattern to research and engineering. A system could try many approaches, revise them, test them, and carry forward the few that survive. The final result might look like a sudden insight even though it depended on a long trail of failed attempts.

Now imagine doing that across thousands of parallel runs, with time to revisit the promising ones. I find that easier to picture than a model producing a perfect answer in one inspired response.

There is an obvious constraint: more attempts help only if there is a way to distinguish progress from error. Compute cannot make a bad assumption true. Mathematical arguments need checking, software needs testing, and physical claims need evidence from the world.

That makes verification part of the workload. Some of the most valuable compute may go toward finding out why an attractive answer is wrong.

The useful unit of work may become an investigation

Fable made this feel less abstract because it was already useful for so much of what I wanted to do. I kept moving from whether it could help to deciding which problem deserved its attention.

Simon Willison's early account of using Fable offered another practical example: he described being impressed by the API design, tests, code, and documentation it produced. I recognized the feeling of getting through work that would otherwise have taken much longer.

For everyday tasks, that level of help might eventually be enough on a local machine. A serious investigation could benefit from a much larger arrangement.

I picture something like a research group assembled around a question. One model proposes an approach, another looks for weaknesses, another writes code to test it, and another searches the literature. Findings feed into a revised attempt. Tools, simulations, memory, and human review connect the pieces.

Making that arrangement useful is its own hard problem. Coordination can consume the savings. Several models can reinforce the same mistake. Some fields offer much weaker feedback than a test suite or a checkable proof.

Still, the possibility changes what I imagine buying with compute. A useful result might come from hours or days of organized work, with an answer emerging only after several approaches have failed.

The appeal is having more room to try

AI discussions often center on which jobs will change or disappear. Those questions matter, and wider access to capable tools will not make the disruption painless.

What draws me toward this particular future is the possibility of asking questions we currently leave alone. What would a small lab attempt if it could explore ideas at the scale of a much larger organization? What would a solo builder try if the first hundred failed prototypes were affordable?

That could matter in medicine, materials, mathematics, and engineering. It does not promise a cure for every disease or a solution to every hard problem. It could give people more chances to find something useful.

This is also not a calculation showing that every new data center will pay off. The financial costs, energy demands, and difficulty of turning compute into reliable discoveries remain real. I see the boom as a bet on new kinds of work becoming valuable, with plenty of room for that bet to be made badly.

Fable gave me a small, personal glimpse of why someone would make it. If local models become excellent, I expect to use them constantly. I can also imagine wanting far more compute for a problem that deserves a sustained investigation.

I would love to hear what you would try if the cost of exploring a difficult idea dropped enough to make the first attempt worthwhile.