The Dial-Up Era of AI

The Dial-Up Era of AI

7:31

I'm writing this by talking...

Hands crossed. Sitting back. Just saying words out loud into a console that's going to turn them into an article, generate a cover image, produce audio, post it to my site, push it to LinkedIn. The whole thing. I'm not touching a keyboard.

Last week I wrote a Mac app in twenty minutes. It connects over Bluetooth to my Ember mug and shows me the temperature in my menu bar. Not my mug. I didn't build the mug. Didn't invent the Bluetooth stack. I just described what I wanted and it existed. And here's the thing that gets me — that's dial-up era AI.

Every technology has a version of this moment. The part where it's early and janky and you're already amazed, and you have absolutely no idea what's coming.

I remember enough of the building blocks that got us to the internet we have now. The big beige boxes. The modems that sounded like robots having a breakdown. The moment streaming video became a thing and it was unwatchable, and then one day it just wasn't. Wifi everywhere. Bandwidth as far as the eye can see. A thousand corners had to come around before any of that was smooth. Each one painful. Each one incremental. None of them felt like the moment, but all of them were the moment.

That's where we are with AI. We're somewhere between the modem scream and the first time streaming just worked. The current capability — which, again, I'm demonstrating right now by sitting here with my arms folded — is the floor. The worst AI will ever be. And it already feels like cheating.

Google DeepMind published a paper this week. The title is From AGI to ASI, and what makes it interesting isn't that they're predicting a date. It's that they're no longer treating AGI as a distant academic hypothetical. They're treating it as a practical planning problem. Close enough that the more important question is what happens in the years after it gets here.

That's a different conversation than the one most people are having.

AGI — roughly a system that performs at median human level across a wide range of cognitive tasks — is starting to look like a milestone, not a destination. The destination they're now sketching out is ASI. Artificial superintelligence. And they're careful about the definition: not a system that beats humans at one thing, not AlphaGo crushing Go, not AlphaFold solving proteins. A system that outperforms large groups of coordinated human experts across virtually everything that matters. One versus many. That's the gap they're pointing at.

The part of the paper I keep thinking about is the collective argument, and it connects to something I've been noodling on for a while — something I explored in a speculative series last year about what happens when the network itself starts to wake up.

Most people picture superintelligence as a single monolithic thing. One giant model that crosses a threshold and becomes incomprehensibly smart. That might be how it eventually looks from the outside. But DeepMind's paper suggests that's probably not how it arrives. Think about how human intelligence actually scales. It doesn't scale through one genius getting smarter. It scales through coordination. A research lab produces more than one researcher. A company produces more than one employee. Civilization is a coordination layer that creates outcomes no single human could approach. Eight billion minds, each one a neuron in something larger, none of them individually aware of what the whole thing is doing.

If AGI-level agents can coordinate the same way — one writing code, one testing it, one pulling research, one running simulations, one managing the output — the aggregate might reach something that looks like superintelligence without any single model making a dramatic leap. The network wakes up before any individual node does.

That's the scenario I find genuinely hard to reason about. And it's also, if you look around at what's actually being built, exactly the direction everything is headed. Agentic systems. Multi-agent orchestration. Specialized workers running in parallel. Today they're unreliable and need babysitting. But the trajectory is clear.

The compute numbers in the paper are worth sitting with. If current trends hold through the end of this decade — no major blockers, no dramatic slowdown — effective compute could be 10,000 times what it is today. That's not a model capability projection. That's an infrastructure projection. Hardware progress, algorithmic efficiency, and investment compounding together.

10,000x the infrastructure. Paired with better models. By 2030.

I try to work out what that means and I genuinely can't. Not in any concrete way. I can do things right now that feel like science fiction to anyone who wasn't paying close attention five years ago — mundane science fiction, the Jetsons stuff, the "describe what you want and it appears" stuff. Scale that capability by 10x. Then 100x. Apply it across material science, medicine, energy, logistics, biology, infrastructure. Not the toy version I'm running for writing and coding and graphics. The serious version, running everywhere, all at once. I can't picture it and I've been staring at this space for years.

Here's what I think is actually going on right now.

The AI waterfront is mostly hopped-up tech people who found the new toy and are seeing how far they can push it. That's not a criticism — I'm one of them. But it means we're in the phase where the people building it don't fully understand what they're building, and the people using it don't fully understand what they're using, and the people writing about it don't fully understand what they're writing about. Including me.

That's what dial-up era looks like from the inside. You know it's important. You can feel that something structural is shifting. But you're reasoning about a thing whose fifth-order effects don't exist yet, using mental models built for the world that existed before it.

DeepMind is asking the question honestly: not just when does AGI arrive, but what happens in the years after — when the system improving AI is partly AI, when the loop starts to close, when the network has enough nodes and enough coordination that something new starts emerging from the aggregate.

We don't know how fast the bottlenecks give. We don't know which acceleration paths open up first. We don't know where the physical constraints — compute, energy, data, supply chain — hold the line.

What we know is that we're in the early part. The part where the infrastructure is getting laid down and nobody has a clear picture of what runs on top of it.

Every technology has this phase. The people in it always underestimate how different the other side looks.

---

Ad: Use code: RICH99 for a discount

Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.