Apparently We Weren't Dreaming Big Enough
A few years ago, this was bourbon-and-nerd-shit conversation.
You know the kind.
A few technically inclined people sitting around, talking about where AI might go if you just kept pushing the line forward.
Dude, what if AI could actually design chips?
Not autocomplete some Verilog. Not help an engineer find a bug. Actually get deep enough into the problem to meaningfully participate in designing the silicon itself.
Then somebody takes it another step.
What if the AI helped design the hardware that runs the AI?
And somebody else inevitably goes: Okay, but then what if it could program the hardware too?
Now everyone needs another drink.
That's how these conversations went. They weren't predictions.
Certainly not mine alone. They were extrapolations. You'd see some little research result, some weird capability poking through, and then we'd sit around and spin the thing forward.
If this works...And this keeps improving...And compute keeps scaling...Then what happens when you point it at that?
I wrote about some of those early experiments in 2023. At the time, researchers were already getting large language models involved in chip design. They could generate Verilog, help navigate design spaces and augment engineers working on increasingly complicated systems.
It was enough to make the imagination run. But it was still imagination.
Or at least I thought we had more runway.
Apparently not much more.
Then Jalapeño Happened
OpenAI has now built its first custom AI inference chip. And AI didn't just show up afterward to write the press release.
OpenAI says AI directly participated in the chip's development—exploring implementations, optimizing arithmetic circuits, shortening design and verification loops, and helping move the project from initial design to tapeout in nine months.
Then they designed the chip so AI could help program it.
Codex is generating and optimizing low-level kernels for the hardware, with OpenAI reporting AI-generated implementations of some components outperforming existing human-expert-written versions.
SemiAnalysis went into the lab and tested the actual silicon.
On performance per watt, Jalapeño beat the Nvidia, AMD and Google systems they tested across several major open models. There are caveats. More testing needs to happen. This isn't "Nvidia is dead." But that's almost beside the point. Because I'm sitting here looking at this thing thinking:
Holy shit. That was fast.
Not that AI eventually got there. That part doesn't shock me nearly as much. It's how quickly one of those slightly drunk, slightly stupid "dude, what if..." conversations turned into an actual piece of silicon.
Apparently "Eventually" Is Getting Really Short
That's the signal I keep coming back to.
We've watched this happen elsewhere.
EvolutionaryScale's ESM3 generated a new fluorescent protein so divergent from known proteins that its distance through protein space was compared to more than 500 million years of natural evolution.
Obviously nobody simulated half a billion years one generation at a time.
That's what makes it interesting. AI learned enough of the structure of the problem to navigate a search space that evolution had taken an incomprehensible amount of time to traverse.
Five hundred million years...Lunch.
Then Microsoft and Pacific Northwest National Laboratory started with roughly 32 million possible materials, used AI and high-performance computing to reduce the field to a tiny number of candidates, and got to an experimental battery material using dramatically less lithium.
Again: Huge search space...AI sauce.
Search space collapses.
And now chips. Which, if you've ever looked at a modern processor beyond the shiny piece of silicon in your hand, should make you stop for a second.
These Things Are F'ing Insane
Look at a GPU or a modern ASIC under an electron microscope. Really look at what humanity has figured out how to manufacture.
Billions of transistors.
Nanometer-scale structures.
Timing.
Power.
Heat.
Memory hierarchies.
Data movement.
Signal integrity.
Packaging.
Networking.
Manufacturing tolerances that are almost absurd to contemplate.
Then put all of those constraints together and make the resulting machine reliably perform trillions of operations.
Modern semiconductors are arguably among the most complicated constrained objects human beings have ever created. And AI just walked into that domain and made itself useful. Not toy useful. Not "help me write an email" useful.
Useful enough to participate in designing the hardware.
Useful enough to help optimize it.
Useful enough to write low-level code for it.
Useful enough that the first-generation result is legitimately competitive with systems created by companies that have been doing this for decades.
That changes my calibration.
Because if you can point this stuff at that, I'm becoming increasingly reluctant to tell you what you can't point it at.
Take the Gloves Off
Medicine.
Materials science.
Chemistry.
Energy.
Aerodynamics.
Manufacturing.
Logistics.
Economics.
Power generation.
Drug development.
Climate systems.
Structural engineering.
Agriculture.
Pick your field. Every one of them contains problems with huge numbers of variables, enormous search spaces, weird constraints and generations of accumulated human expertise. And that's exactly where AI keeps doing its best AI-sauce shit. Not necessarily replacing the expert. Giving the expert access to a vastly larger portion of the possibility space.
That's a different proposition.
Scientists still have to validate discoveries. Engineers still set constraints. Labs still conduct experiments. Foundries still manufacture silicon. Physics still gets veto power. Reality isn't going anywhere.
But between the question and reality, we've inserted something completely new.
Search-Space Compression
I think that may be the cleanest description I've found for what keeps happening.
AI is becoming a search-space compression engine. A lot of the world's hardest problems aren't difficult because nobody understands the question.
They're difficult because the number of possible answers is enormous.
Thirty-two million candidate materials. An astronomical number of proteins. Billions of possible circuit arrangements. Nearly endless compiler and kernel optimizations. Drug molecules. Chemical processes. Wing geometries. Grid configurations.
Historically, humans attacked those spaces using combinations of theory, experience, intuition, experimentation, simulation and brute force.
And time. Lots of time.
AI doesn't remove those tools.
It massively increases the territory we can cover with them.
That's why a little AI sauce can produce such disproportionate results.
You don't have to automate the entire scientific discipline.
Sometimes you only need AI inserted at exactly the place where human search stops scaling.
Then everything downstream changes.
And Now AI Is Improving the Machine Running AI
This is the particularly weird part.
AI helps design better AI hardware. The better hardware makes AI cheaper and faster to run. Better AI becomes capable of doing more engineering. That engineering produces better hardware. Better hardware produces more capable AI infrastructure.
Round and round it goes.
No, I'm not declaring the singularity. Nobody needs to start stockpiling canned food. It's much more mundane than that. And possibly much more consequential.
It's an industrial flywheel.
Human engineers, AI models, simulation, software, silicon and manufacturing increasingly operating as one feedback system.
And the cycle time between those pieces is shrinking.
Fast.
That's What I Got Wrong
Looking back at those old conversations, I'm kind of amazed at how much of the direction was right. Some of it was undoubtedly luck.
You throw enough "what ifs" around with smart people and occasionally one lands squarely on the future. Maybe I've got decent radar for this stuff. Maybe I've just spent enough time staring at the trajectory that some of the next steps become visible before they're obvious. Probably some combination. But I don't particularly care about claiming the prediction.
The useful lesson is that we weren't dreaming big enough. We kept imagining these capabilities as things that would eventually arrive.
Someday AI helps design proteins. Someday AI discovers new materials. Someday AI meaningfully participates in semiconductor design. Someday AI optimizes the hardware that runs AI.
Then "someday" starts showing up every couple of years...Then every few months. That's what has my attention. Because there are still enormous parts of the economy operating under the assumption that AI means a chatbot that summarizes meetings.
Meanwhile, AI is roaming around protein space and helping design silicon.
Wait Until the Rest of the World Finds Out
That's where I think we are. Not at the end of anything. At the point where the general mechanism is becoming obvious.
Give AI a sufficiently rich representation of a domain. Give it a goal. Give it constraints. Give it feedback. Give it a way to test whether it screwed up.
Then let it search.
We now have enough examples across sufficiently different disciplines that I don't think this is a novelty anymore.
It's a pattern.
And once that pattern spreads across every field with giant search spaces and expensive human expertise, things get interesting very quickly.
So I'm done putting comfortable timelines on this stuff.
I'm also increasingly hesitant to draw arbitrary lines around what AI will or won't be able to touch.
If it can get elbow-deep into proteins, materials and the nanoscale architecture of the machine it runs on, that's enough evidence for me to update the model.
Clip the line. Set sail. Apparently a little AI sauce goes a very, very long freakin way.
Sources
Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.
Comments