Fun post, but I find the industry's obsession with compute to be rather vapid, and this is a good example:
> One million Claudes. To be able to search every book in history, solve math problems, write novels, read every comment, watch every reel, iterate over and over on a piece of code until it’s perfect – spend a human year in 10 minutes. 50,000 people working for you, all aligned with you, all answering as one.
We are already near the limits of what we can do if we throw compute at Claude without improving the underlying models, and it is not clear how we can get big improvements on the underlying models at this point. Surely geohot knows this, so I am surprised he thinks that "one million Claudes" would be able to e.g. write a better novel than one hundred Claudes, or even one Claude.
> We are already near the limits of what we can do
Hard disagree. If I had a million Claudes worth of compute I'd be livestreaming my entire reality feed to a local server 24/7 and having it organize my observations and thoughts, synthesize new ideas, implement prototypes and discard infeasible ones while I sleep. If you're in the business of knowledge creation, a million Claudes isn't enough. Text is an easy modality, I want foundation models that operate on text, images, audio, video, streaming point clouds, ...
Using a single Claude agent, ask it to generate "new" ideas and it will generate an immense list. Ask it to rank those ideas by novelty and it will comply.
The results will be lackluster. Additional agents will not improve the result.
I call this the Laurie Anderson fallacy, from a line in one of her songs:
> "Heaven is exactly like where you are right now, but much, much better."
If a million Claudes of compute were accessible, people would not be doing the same things they are now, but more so. They'd be doing very different things we likely can't imagine - in the same way that Alan Turing imagined machines learning from experience, but didn't imagine downstream products like Sora, ad tech, or social media, or their cultural and economic effects.
> If a million Claudes of compute were accessible, people would not be doing the same things they are now, but more so. They'd be doing very different things we likely can't imagine
You may as well make the same argument about commanding a billion Nintendo 64s. We cannot simply scale up Claude instances like you say, its ability to produce interesting output is bounded by the underlying model.
There's nothing (except their capacity and your token budget I guess) stopping anyone from having a million simultaneous conversations with Claude right now.
Maybe a million is a stretch but thousands is completely doable right now. That's thousands of Claudes. Knock yourself out.
But that's the point. No human has the capacity to handle a million simultaneous conversations, any million-sized workflow would have to be AI-managed itself, and it's not even clear what the goals would be.
If it ever becomes possible to say "build me a unicorn" you're going to get millions of people trying to do the same thing, and you no longer have the same economy.
Because the features that generate unicorns stop being unusual.
Startup slop instead of art slop.
Which is the real problem with AI. Work gets cheap, original value stays expensive no matter how much compute you throw at it.
Because if it gets cheap too, it gets commoditised and stops being valuable.
And globally that applies to everything.
AI will either have to be tightly rationed, or it will murder the competitive economy.
Agentic development allows one Claude, multiplexed a few times, to vastly improve its output and tackle much bigger problems than just prompting the one instance. If you had a million Claudes in layered networks like we do with matmuls to form Claude, you'd be really cooking with gas.
The output undoubtedly improves when looping LLM output back into the model at inference-time, but there is a limit to this and it is still bounded by the acumen of the underlying model. You cannot just recurse these models with tooling and compute to e.g. solve new physics.
Don’t get hung up on the Claude part. We already know an algorithm that produces AGI: evolution. We don’t have the compute to run this algorithm because it requires simulating the whole Earth. But with enough compute, it becomes possible.
The "Claude part" is important here, though. If you believe we can produce AGI by simulating physics then the existing paradigm is far too slow. A zettaflop does not even get close, you can have a moon-sized computer and it probably will not be able to produce AGI using the current transformer-based, extremely slow, classical computing architecture. We need to improve the underlying computation paradigm.
This is why the focus on raw compute is a poor use of engineering time. We have plenty of it, we are just headed in the wrong direction.
> One million Claudes. To be able to search every book in history, solve math problems, write novels, read every comment, watch every reel, iterate over and over on a piece of code until it’s perfect – spend a human year in 10 minutes. 50,000 people working for you, all aligned with you, all answering as one.
We are already near the limits of what we can do if we throw compute at Claude without improving the underlying models, and it is not clear how we can get big improvements on the underlying models at this point. Surely geohot knows this, so I am surprised he thinks that "one million Claudes" would be able to e.g. write a better novel than one hundred Claudes, or even one Claude.