The Frontier Has a Floor

Every week I read more AI news than any one person should. Most of it is noise. The signal shows up when you line the dots up across a week and notice they point the same direction.

This week they did. A neural network learned to see through skulls. A transistor got shrunk to the width of a DNA strand. A humanoid stopped assuming the floor was flat. And the country that builds the machines that build the world's chips ran short on electricity. Different fields, one shape underneath: the digital story keeps colliding with the physical world, and the physical world keeps having the last word.

---

The brain, read from outside the head

Aleph Neuro is reading neural activity with sound waves. No implant. No surgery.

The physics is elegant. Focused ultrasound passes through the skull; blood flow shifts as neurons fire, and those shifts change how the waves scatter back. Read the echo, map the activity. What used to mean drilling into the skull and threading electrodes now happens from outside the head.

Most neurotech attention goes to implants — chips in the cortex, headlines about typing with thoughts. Meanwhile the non-invasive side has been quietly closing the gap, without the risk of open-brain surgery. The same week, MidJourney — an AI image lab — announced a full-body scanner it pitches as "as powerful as MRI, as casual as a trip to the spa." It reads like a joke until you notice a preventive-imaging company already runs $500M+ in revenue on radiation-free scans with zero VC funding.

The model that learned to read patterns in pixels turns out to be the same kind of model that can read what's happening inside a body. The perception layer is leaving the screen and pointing at us.

The machines keep shrinking; the map keeps growing

IBM just built a transistor 7 angstroms wide. That's about the width of a single strand of DNA.

Sit with the number: 0.7 nanometers. Sub-1nm was the point where a lot of people quietly assumed Moore's Law finally ran out of room — you can't keep shrinking flat transistors forever before the physics stops cooperating. So IBM stopped shrinking flat and started stacking. Their nanostack architecture piles transistors vertically, the same trick that turned single-story cities into skyscrapers when the land ran out. Roughly another decade of density scaling, bought.

This matters more than a spec bump because every AI capability upstream — longer agent runs, cheaper tokens, bigger models — sits on this floor continuing to move. Which is why the export-control story got sharper this week: the US told ASML it's worried China may have gotten hold of a leading-edge tool. That's the exact leak the whole regime was built to prevent. The theory is simple — you can design a chip anywhere, but you can't fabricate a bleeding-edge one without a handful of machines only a few companies know how to build. Control the tools, control the frontier. The entire strategy rests on one assumption: the chokepoint holds.

The robots stopped trusting a remembered world

For years, humanoid robots assumed the floor was flat. Most lab demos ran on the same polished surface every time. The robot didn't see the ground. It memorized it.

That assumption is breaking. Perceptive Behavior Foundation Models let a humanoid take one learned human motion prior and adapt it to whatever terrain it actually sees — a slope, a stair, gravel — decided on the fly instead of scripted in advance. PNDbotics' Adam became the first full-size humanoid to scale a 1-meter box: not step over a curb, climb something taller than a kitchen counter. And NVIDIA's Isaac GR00T is trying to unify the fragmented dev stack — simulation, data, training in one place, setup in hours instead of days.

The pattern underneath all three: perception is catching up to motion. The robot stops trusting a remembered world and starts responding to the real one. That's the same move Aleph Neuro made with the brain and MidJourney made with the body — machines learning to read the physical world as it actually is, not as a model assumed it would be.

Capability is racing; the way we use it is standing still

Claude Opus 4.7 built 2 to 17 weeks of engineering work in 14 hours. Not an autocompleted function. A full software package — the kind of thing a human team would scope, plan, and grind through over weeks.

The new MirrorCode benchmark measures exactly this: the largest engineering tasks a model can finish coding autonomously for days at a time. METR-style time-horizon extrapolations now push predicted 50% task-completion into dozens of hours of continuous work; one projection landed at 61 hours.

Then you look at how people actually use these tools. 64% of Codex users run exactly one agent at a time. And of all US adults, half now use AI but only 18% call themselves confident — that's about 37% even among people who've tried it. Two curves are pulling apart. Capability is sprinting toward workweeks of autonomous output. Human habit is still sequential: ask one question, get one answer, move on. The productivity story people keep waiting for doesn't arrive evenly — it accrues to the few who change how they work, not the many who touch the tool and call it done.

The digital economy has a physical floor

The Netherlands makes the machines that make the world's chips. It's running out of electricity to grow.

ASML in Veldhoven builds the lithography systems every advanced fab on earth depends on — arguably more silicon expertise packed into one small country than anywhere else in Europe. The thing holding it back isn't talent or capital. It's the grid. Companies wanting new or bigger electricity connections sit on waiting lists that run for years.

We keep talking about AI as a software story: models, tokens, agents. But every token runs in a data center, and every data center needs a connection to the grid. AI compute has been doubling every seven months — triple the pace of Moore's Law — since xAI's Colossus launched. The popular ceiling was water, but US data centers use about 0.2% of the country's daily water; that's a local siting problem, not an industry ceiling. The real ceilings are power and silicon. You can't double compute every seven months without doubling the electricity to run it. Meanwhile $110 billion in real generative-AI sales booked over the past year — a run rate north of $175B — is demand pushing hard against a physical supply that can't move at software speed.

---

The theme across all of it: this was the week AI stopped being weightless. Sound waves through bone, transistors at DNA scale, robots reading gravel, agents that run for days, and a grid that can't keep up. The frontier keeps moving, but it's discovering its floor — and the floor is made of physics, electricity, and machines only a few places on earth know how to build.

Watch where the bottleneck sits next quarter. My bet: less about whether a model can, more about whether the world underneath it can supply the power and silicon to let it.

If you'd rather have the map than the noise, this is what I do here each week — pull the developments into one clear picture of where it's heading. Subscribe and I'll see you next edition.