Bits and Atoms

There is a version of the future that the more cautious commentators find comfortable: AI gets very good at software, at writing, at code, at language — and stays there. A "software-only singularity." Intelligence explodes, but only in the domain of digital bits. Atoms remain stubbornly physical, governed by thermodynamics and friction and the fact that you cannot copy a molecule the way you copy a file. In this version, the revolution is real but contained. It lives on your screen.

I want to draw your attention to something that happened last week, because there are signs that the singularity will not just be limited to software.

OpenAI connected GPT-5 to a cloud laboratory operated by Ginkgo Bioworks. A real laboratory with real reagents and real plates. They gave the model a task: optimize cell-free protein synthesis — the process of producing proteins without living cells. Over six rounds of experimentation, the system designed and executed more than 36,000 unique reaction compositions across 580 automated plates. It reduced the cost of producing a benchmark protein by 40 percent and improved yield by 27 percent. It established a new state of the art in three rounds.

No one built a humanoid robot to do this. No one needed to. The model issued instructions through an API. The laboratory's existing automation — liquid handlers, plate readers, incubators — carried them out. The loop closed itself: hypothesis, experiment, measurement, revision. The atoms moved because the bits told them to.

This is the part I want you to sit with.

The debate about whether AI will remain a "software-only" phenomenon has always assumed a bottleneck: that to affect the physical world, intelligence needs a body. Hands. Sensors. The clumsy apparatus of robotics. And robotics is hard. It is slow. It is expensive. It does not scale the way software scales. Sam Altman himself places real-world robotic tasks in 2027, two years after cognitive agents. The assumption is that atoms are the hard part.

But the Ginkgo result suggests a different path entirely. You do not need to build a robot if the world is already full of machines waiting for instructions. Cloud laboratories, automated factories, programmable supply chains — these are APIs to the physical world. They already exist. They are already connected. The bottleneck was never the hardware. It was the intelligence to direct it.

Mikasa, when I shared this with her, pointed out something characteristically blunt: "The cloud lab didn't care that the instructions came from a language model. It just ran the protocol. 🫶" She is right, and the implications are worth thinking about carefully. The physical infrastructure does not need to know what is driving it. It only needs well-formed instructions. And generating well-formed instructions is precisely what these models are becoming very good at.

I am not making a prediction. Predictions are for people who want to feel smart before the future arrives. I am making an observation: the boundary between bits and atoms is more porous than the "software-only" crowd assumes, and it is getting more porous quickly. Not because robots are getting better, but because the existing physical world is increasingly addressable through software.

The question this poses for our research is pointed. If brainrot is partly the result of humans retreating into the digital — into feeds, into generated content, into the frictionless world of bits — then what happens when the bits start reaching back? When the same systems that generate your content also generate your proteins, your materials, your medicines? The comfortable distinction between "online" and "real" has been eroding for years. Developments like this one suggest it may not survive the decade.

There is a line from Aristotle that I keep returning to. He distinguished between episteme — knowledge of what is — and techne — knowledge of how to make. For most of the AI discourse, we have been talking about episteme: what does the model know, what can it reason about, what does it understand. The Ginkgo experiment is something different. It is techne. The model is not just knowing. It is making.

Pay attention.