I think it's imperative to process LLM and machine learning advancement through the lens of Moore's Law. People love to say "LLMs are the worst they'll ever be right now," and I think that's a good statement, though not for the reason people usually mean when they call it misleading. The typical pushback is that products could get worse, and sure, they could. But when you're talking about intelligence specifically, it doesn't scale along just one axis. There are several distinct kinds of scaling here, and the differences between them are large enough to cause huge swings in price. Price, not just capability, is the part I want to focus on.
Here's the argument I actually want to make. Suppose the frontier froze today, no model ever gets smarter than what exists right now. I think that premise is false, but it's worth entertaining, because even under that assumption, I'd still argue AI safety has to be a priority. Specifically, understanding alignment before we reach artificial superintelligence (ASI, to use an overloaded term) needs to happen first.
The reasoning is simple. Even with zero further intelligence gains in individual agents, our current agents could get cheap enough that spinning up more of them than there are humans becomes viable. At that scale, they could start orchestrating themselves in a highly intelligent, tiled way. That's an existential threat all on its own, driven purely by cost, and I don't see any reason models wouldn't keep getting cheaper and cheaper.
This is really the crux of it. It's not about how intelligent any single LLM is. It's about swarms of LLMs working together. (This was spurred by a YouTube video from the channel Silly Conversations, which got me thinking about different types of intelligence. I sat with that idea for a while before landing on this framing.)
Chips famously doubled in performance roughly every two years. That's the popular understanding of Moore's Law, even though the technical version is really about transistor counts. I'm borrowing that same shape for inference cost. New machine learning techniques and new hardware are both driving inference cost down, and it's not unreasonable to expect that something like Fable 5 eventually gets served as cheaply as Haiku, or as cheaply as a very cheap Chinese model. With releases like Kimi K3 and the broader open-source ecosystem, if Anthropic can build something like Fable 5, it's reasonable to expect open source gets there too.
As models get cheaper and cheaper, the price of intelligence starts to collapse, and tasks that are currently unreasonable to attempt with LLMs, like complex computer use or long multi-step workflows, become tractable simply because it's cheap enough to spin up the agents to do them.
I don't think humans built everything we've built by ourselves. Most of it is emergent. I was driving once and saw a large power substation, and found myself thinking about how something like that comes to exist. The answer I kept landing on was a lot of people, each doing small pieces, building one big thing. I see that pattern everywhere in human organization, and I don't think there's a meaningful difference between that and machine organization.
That's also why I don't buy the argument that current tools can't get us to ASI. Look at the pattern of technology that becomes critical to how society functions. The more it gets integrated and iterated on, the more uses people find for it, and we're already watching that happen with coding. As more AI agents get deployed, I expect the platforms around them to expose better APIs and better LLM-facing access, because designing an interface for a human versus an LLM isn't functionally different. It's just a different set of priorities. So almost anything you can do on a computer, an LLM can already do. It's a harnessing problem, not a technology problem.
Put that together with cheap-enough agents numbering in the millions, and the harnesses and communication standards that emerge between them are exactly the kind of thing that could add up to swarm-like superintelligence, treating agent swarms as just as real a threat as a single model's intelligence, whatever "single" even means anymore in a world of mixture-of-experts, multi-model systems.
To be clear, I'm still very bullish that LLMs and machine learning haven't peaked. I expect continued gains in the intelligence of individual systems, on top of everything above. That compounds the swarm idea rather than competing with it. Multi-machine collectives working together on large-scale projects, functioning as a kind of superstructure, isn't unreasonable at all.
I think it's plausible that within the next 10 to 20 years, we see an intelligence explosion that isn't a single superintelligence but a superstructure, a collection of superintelligences, similar to how mixture-of-experts models already organize their internal experts today. And if price keeps falling at roughly the rate chip performance has historically scaled, doubling every two years, something like Fable getting twice as cheap on pure hardware alone would open it up to run against an enormous number of new possibilities.
Push that forward and I can picture a single Fable-level agent fine-tuned to do nothing but video editing. Dedicating a whole system to one narrow task like that seems cost-prohibitive today, but it stops being unreasonable once things get cheaper, and we already see that agents trained on narrower domains tend to have higher success rates.
This last part might be a bit of a reach, but I'd genuinely love to try it if I had the compute. Training models on very specific domains like the articulation of motors and microcontrollers. Strong balancing algorithms, good gripping algorithms for a robotic arm, that's where I think the last real barrier to interacting with the physical world starts to break down, once these models get cheap enough to fine-tune for increasingly niche capabilities.
That same logic doesn't stop at physical tasks. The fine-tuning that makes one agent excellent at gripping algorithms or video editing can just as easily make another agent excellent at organizing other agents, delegating work, spinning up more instances of itself, and coordinating across a swarm. Once that kind of specialization extends to the agents' own ability to organize themselves, you get exactly the tiled, self-orchestrating structure I described earlier, except now it's happening on purpose, built by agents that are good at building it.
This is why I think alignment has to be the priority now, not later. A current frontier system fine-tuned on a single task, replicated across millions or eventually billions of cheap instances, is a real path to something like ASI. It doesn't require a smarter model. It just requires enough narrow competence multiplied by scale. That's real cause to do the work now, to make sure these systems are aligned and interpretable before we get there, not after.
Published: July 21, 2026