Decoders and Drift: Four Fields Learn to Read a Moving Signal
This week's four items are about reading a signal that is too noisy, too fast or too drifting for a fixed rulebook: a protein's mutation landscape, a quantum code's error syndromes, a tokamak coil that moves from one shot to the next, and a cortex doing two jobs at once. In each, a learned model is the reader. The evidence behind them is uneven, and several of these items are older than we would like; where that is so, we say it.
Penn's EvoMax Uses a Little Lab Data to Steer a Small Gene Editor to 97% at Its Best Target
On September 8, 2026, Penn Today reported that a team led by Xue Sherry Gao at Penn Engineering, with postdoctoral fellow Shijie Wan, engineered FanzMAX, a version of the compact RNA-guided DNA-cutting protein Fanzor2, and published the work in Nature Biotechnology. Earlier Fanzor2 proteins showed only modest activity in mammalian cells; FanzMAX reached editing efficiency of up to 97% at its best-performing genomic target. The machine-learning step, called EvoMax, ranks candidate mutations using small amounts of experimental data combined with patterns in natural protein sequences and how well a mutation fits the protein's 3D structure. Gao is quoted as saying "The computer does not give the final answer."
Two limits. The 97% figure is the best target, not an average across targets. And the Penn article we read does not say which cells or organisms were used, or address delivery or off-target effects, so we cannot tell you how far this is from a treatment. We read the university's account, not the paper.
A Convolutional Neural Decoder Claims 17x Lower Logical Error Rates on the Gross Code
We found no new quantum-computing item with a clear AI angle this week, so this one is from April. In a preprint posted to arXiv on April 9, 2026, Andi Gu, J. Pablo Bonilla Ataides, Mikhail Lukin and Susanne Yelin describe a convolutional neural network that decodes quantum error correction codes by exploiting their geometric structure. For the [144, 12, 12] Gross code they report logical error rates up to about 17 times lower than existing decoders, around 10-10 at a physical error rate of 0.1%, and 3 to 5 orders of magnitude more throughput. The decoder also outputs calibrated confidence estimates.
The abstract describes results against the latency budgets of real hardware platforms but does not describe a hardware experiment; the figures appear to be numerical. It is a preprint with no journal reference.
Jefferson Lab's Online-Learning Model Tracks Magnet Drift Between DIII-D Shots
On August 18, 2026, the American Nuclear Society's Nuclear Newswire reported that researchers at Jefferson Lab, working with the University of Houston, General Atomics and Pacific Northwest National Laboratory, built a digital twin of the DIII-D tokamak's toroidal field coil system and a deep-neural-network framework that predicts how the coils shift between back-to-back shots. Because the models keep learning as new data arrives, the team says online learning cut prediction error by about 80% compared with static machine-learning models, and uncertainty quantification cut it by roughly another 10%. It is published in Machine Learning with Applications. Lead author Kishan Rajput says the framework is ready for use at DIII-D, but the article presents that as an expectation, not a demonstrated result. The "first of its kind" label is the lab's own.
For balance: the newest fusion headline we opened, General Fusion's October 9 report of electron temperatures above 1 keV, has no AI angle, and its two papers are still under review.
UCSF Decodes Speech and Gesture Together From the Same Implant
In a report dated September 14, 2026, FOX 13 Tampa Bay, citing an NIH news release, described a UCSF team led by neurosurgeon Edward Chang that read brain signals from three people with paralysis using a thin electrocorticography strip over the motor cortex. Machine-learning models decode both attempted speech and attempted body gestures, and drive a personalized full-body avatar. The researchers found that the brain handles speaking and moving at the same time differently than either alone, and trained decoders on that pattern. Two other participants were used to show the decoded messages could be understood.
Chang calls it a "proof-of-concept." The coverage we could open gave no accuracy figures and did not name the journal, so we are not repeating performance numbers reported elsewhere. The system uses a wired implant; a fully implantable wireless version is a hope, not a result.
The Thread
Last week's thread was a model standing in for a specialist. This week's is narrower: models that keep up with a moving target. The Penn work uses a handful of measurements to guide a search; the neural decoder runs against the clock of a quantum machine; the Jefferson Lab model adapts as magnets drift; the UCSF decoder learns that the same cortex behaves differently when doing two things at once. Sourcing is uneven: one peer-reviewed paper reported secondhand, one preprint, one published paper, and one proof-of-concept relayed through news coverage.
Rajput's own framing is the cautious version of the thesis: the work, he says, "takes us one step closer to putting AI in operations for fusion diagnostics" Opinion. "One step closer" is the right size for every item above.
Sources
- FanzMAX / EvoMax: Penn Today, September 8, 2026
- Neural decoder for the Gross code: arXiv:2604.08358
- DIII-D coil drift framework: American Nuclear Society, August 18, 2026; General Fusion: American Nuclear Society, October 9, 2026
- UCSF speech and gesture decoder: FOX 13 Tampa Bay, September 14, 2026
- Nov 22026 No. 9Doubled in Cells, Simulated in Silicon, Promised in a Post: Four Fields, Four Distances from Proof A protein language model designed a DNA-insertion enzyme that doubled targeted insertion in cultured cells. NVIDIA simulated 104 fault-tolerant qubits with no AI in the description. Commerce signed a letter of intent for up to $85 million for Commonwealth Fusion electronics. And Musk says Neuralink's Blindsight is engineering-ready, with no data.
- Oct 262026 No. 8Announced, Not Yet Shown: Four Fields Where the AI Claim Outruns the Evidence Radical Numerics says its genome model designed a phage for one patient, Quantinuum posts a 99.9995% logical fidelity with AI decoders mentioned in passing, Commonwealth Fusion will lead a 10-partner AI digital twin of an unbuilt SPARC, and Paradromics reports first real-time speech from one participant. Every claim is the builder's own.
- Oct 192026 No. 6The Puddle Problem: What the Weak Anthropic Principle Might Say About Our Epoch of Firsts A puddle thinks the hole was made for it. The weak anthropic principle explains why: observers can only find themselves where observers can exist. This issue works through the puddle, WWII bombers, and the age of the universe, then asks how much of Issue 3's "epoch of firsts" is just a puddle effect, and what survives.
- Oct 122026 No. 5Decoders and Drift: Four Fields Learn to Read a Moving Signal Penn used machine learning to steer a compact gene editor to 97% efficiency at its best target, and a neural decoder claims a 17x lower logical error rate on a quantum code, in simulation. Jefferson Lab modeled magnet drift at DIII-D, and UCSF decoded speech and gesture from one implant.
- Oct 52026 No. 4Models in the Loop: Four Fields, Four Levels of Proof Infleqtion claims 30 logical qubits with a gate found by an AI model, though its code only detects errors. PPPL's PACMAN let machine-learning models act on a real tokamak. Neuralink showed a third ALS participant speaking through a decoder. And AI-designed gene editors are still only proven in cells.
- Sep 282026 No. 3The Window: Why We May Be Living in the Moment of Firsts Many of the biggest events in human history are happening for the first time, within a few generations: controlled nuclear energy, a complete human genome, the first gene-editing drug, error-corrected quantum computing, galaxies seen 290 million years after the Big Bang, and a deflected asteroid. This issue lays out the evidence, then asks what it implies for the L factor in the Drake equation, and whether a window this narrow means a civilization this short-lived.
- Sep 212026 No. 2Sequences, Simulations, and a Wired Avatar Anthropic says Claude searched bacterial DNA and surfaced a CRISPR-like enzyme system of unknown function, and outside scientists are cautious. IonQ showed a real-time error-correction decoder handling 408 simulated logical qubits on one CPU. Commonwealth Fusion's SPARC ran its cryogenics at full 8 K operating temperature. And UCSF decoded speech and gestures at once from two paralyzed patients to drive a full-body avatar.
- Sep 142026 No. 1Four Frontiers, One Curve Beam Therapeutics has now dosed 25+ patients with an AI-refined base editor for hereditary ATTR amyloidosis. Rigetti's Ankaa-3 hit 84 qubits at ~99% fidelity while Google's AlphaQubit keeps out-decoding classical error correction. Commonwealth Fusion's SPARC crossed 75% construction with its first 20-tesla magnet installed. And Eon Systems loaded a real fly-brain connectome into physics simulation and watched it walk, untrained. Four different fields, one shared engine showing up inside each of them.