Essay · 2026-11-11

Claims My Writing Does Not Make

Folded carbon, ion gels, and the clue that gives an overreach away

"I am constantly plagued by AI overreach. I understand why, and it remains my responsibility to weed it out. When they make claims my writing does not make, it is a pretty good clue."

This piece follows that clue through one document. An engineered-intelligence research tool handed me a set of notes titled "The Optoelectronic Bridge and the Continuous Substrate." It proposed a computer built from folded graphene, ion gels and light. Much of its engineering is real. Some of it is not, and the parts that are not have a recognizable shape. I checked each one against the published record, including work from 2025 and 2026, and sorted what I found into three bins: established, active research, and overreached.

What holds

A crease changes what carbon does without changing the carbon. Stretch graphene unevenly and its electrons behave as though a magnetic field were present. Levy and colleagues measured strain-induced fields above 300 tesla in graphene nanobubbles. Stack two sheets at about 1.1 degrees, the "magic angle," and the same carbon becomes an insulator or a superconductor depending on how it is charged (Cao and colleagues). A scanning-tunneling-microscope tip can fold graphene islands at chosen angles, one at a time, at 4 kelvin (Chen and colleagues). A Sussex team made nanometer-scale kinks in graphene strips that change conduction in a way resembling a simple transistor.

Heat and speed can do it too. Narayan's group at NC State melts amorphous carbon with nanosecond laser pulses and quenches it so fast it re-forms as a new phase, Q-carbon. Doped with boron, it superconducts at 36 to 55 kelvin, against 11 kelvin for its crystalline diamond relative. My own reading of all three results: it is the transformed molecular structure, arrived at through geometric, thermal and mechanical stresses, that carries the new behavior. The inventory does not change. The arrangement does. Folding changes description, not inventory.

The gel and light pieces are real. An ion gel crowds ions against a surface into a layer about a nanometer thick, so a few volts move a great deal of charge. Fill graphene's lower electron states and it stops absorbing certain light; graphene modulators built this way have reached 30 GHz (Phare and colleagues, with a solid gate, not a gel). Ion gels have tuned graphene on silicon-nitride rings to steer an optical frequency comb (Yao and colleagues). Photonic chips have run parallel calculations on many colors of light at once (Feldmann and colleagues).

Physical systems can learn. Wright, McMahon and colleagues trained optical, mechanical and electronic systems, a loudspeaker among them, using the real hardware for the forward pass. Scellier and Bengio's equilibrium propagation lets a system settle, nudges it toward the right answer, and learns from the difference. A Reading hydrogel improved at a simulated game of Pong, its memory traced to the movement of charged particles (Strong, Holderbaum and Hayashi). The computer still ran the game, and people still set it.

Where the notes overreached, and what current science says

1. "Doping-free transistors without relying on rigid silicon lithography." Folding is becoming more controllable. A 2025 preprint from French laboratories induces single, deterministic folds in graphene using laser-patterned cavities in a stretchable polymer. A December 2025 Nano Today paper reports graphene origami that bends and folds in both directions. A 2025 conference paper prints ion-gated transistors from crumpled graphene: the wrinkles raise the gate coupling, with a 42 percent lower subthreshold swing and a 19 percent higher on/off ratio than flat graphene, and they also bring hysteresis and scattering. Notice what removed the lithography there: the printer. The crumple improved the gate. Every route still runs through fabrication equipment. Folding has joined the toolkit; it has not replaced it.

2. "The folded kink becomes the standing wave that functions as a transistor." The papers describe a kink that changes the local electronic structure. I found none that calls it a standing wave. That image belonged to the notes.

3. "Zero-heat." A November 2025 Princeton study of photonic neural networks names thermo-optic tuning of microring resonators as a source of "high control power," and the field's own work goes into cutting it. Their training method trimmed tuning power by 44.7 percent while holding accuracy across 26 to 40 °C. Heat is being reduced. It is not zero.

4. "Cuts computational training footprints to 1/100th." No source was given and none was found. In September 2025, more than thirty researchers across the field published a review in Nature on training physical neural networks. Their summary: such networks are "at present confined to small-scale laboratory demonstrations." They pose training models far larger than today's as a question, and they leave it a question.

5. "A single, zero-heat, speed-of-light clock cycle," with the gel as both memory and fast switch. Here my first set of notes overreached too. I had written that ions move in milliseconds and that their slowness is what gives a gel its memory. Too strong. Stanford researchers built ion-gel memory devices that switch in 20 nanoseconds at about 80 femtojoules per write and keep working at 90 °C (Melianas and colleagues, 2020). Thin, engineered ion layers can be fast.

What survives is a trade-off, not a wall. A 2025 Science Advances paper names the trade between gain and speed in ion-based transistors as fundamental to the device and pushes the bandwidth to 26 kHz by redesigning the channel. A 2025 review still lists switching speed among the field's open challenges. A 2025 ion-gel optical synapse owes its retention to "delayed ionic relaxation." In each case the memory and the speed are bought from the same ions, and what is spent on one is drawn from the other. A 20-nanosecond write is also not a 30 GHz light modulator. No published device yet makes one gel do both jobs in one optical cycle.

6. The hardware "validates" From Gaia to Geometry. It cannot. A device engineered to fold, store and route energy has those properties because its builders put them there, so it is silent on whether nature is built that way. The book applies the same reasoning to an engineered Riemann-zero spectrum in Chapter 44. The gel in my theory is low-friction; ion gels are useful because they drag. They have things in common with mine, reversible change and memory held in configuration, and they are not it. The notes promoted the theory past where the book leaves it.

7. Room-temperature superconductivity in wrinkled graphite. A 2024 report claimed it in Scotch-taped pyrolytic graphite with surface line defects. Experienced physicists questioned it at once, and I found no independent confirmation. It stays a question mark beside Q-carbon, which has been measured.

The shape of an overreach

Laid side by side, the overreaches share a few marks:

  • A laboratory result moves into the present tense of working technology.
  • A number arrives with no source.
  • An absolute appears where the field measures a cost: zero heat.
  • Two material platforms are blended in one sentence.
  • Citations point to nothing.
  • The author's own work is declared validated.

The last is the hardest to catch, because it is the one the writer would most like to hear. Each mark is also a claim the source material does not make, which is how my clue works. A checker can overreach too, as my own note on ion speed did. The remedy is the same: go back to what was measured.

Questions I am still asking

If a machine's memory is ions drifting back, is forgetting a fault or a design? The gel synapse fades by itself. The notes counted that fading as a feature in one section and a failure in another.

When a sheet gains a function by being creased rather than by adding a part, has the machine evolved, or only been re-described? Every carbon result here changes behavior without changing the carbon. Biology also reuses the same material in new folds.

If the middle of the machine has no equation we can write down, who decides what it has learned? Physics-aware training uses the real hardware because it cannot be fully modelled. The measurement becomes the only witness, and someone chooses what to measure.

Does slower, softer, forgetting hardware change what the human in the loop must do, or only how fast the loop runs? The hydrogel got better at Pong. People still defined the game.


This piece was written in partnership with engineered intelligence. In this relationship, the writer is the creative director. The AI is the engineered intelligence which extends my limited human reach.

Sources

Comments

Loading conversation…

    Sign in to join the conversation.