Loading…
The Convergence
Issue No. 4 · October 5, 2026

Models in the Loop: Four Fields, Four Levels of Proof

Genome, Quantum, Fusion, Neural

Three of this week's four items come with the same asterisk: a company's own announcement, a demo video, or a result in cells rather than patients. One is a peer-reviewed tokamak experiment. The tie between them is narrower than a grand theory and more useful: in each field, a learned model is doing a job that used to be done by a person, a hand-tuned controller, or a lucky guess.

✦
GenomeLab · Cells Only

AI-Designed Gene Editors Keep Edging Toward the Clinic, Still in Dishes

No gene-editing story broke this week that we could verify, so this item is from July. As reported on July 24, 2026 by SingularityHub, two research efforts used AI on CRISPR tools. A Chinese team used Google DeepMind's AlphaFold 3 and a tool called ContactSeek to find the protein regions prone to off-target binding, then built an A-to-G base editor with only two changes that beat existing high-fidelity editors without losing editing efficiency. Separately, a team led by Jennifer Doudna asked AI to design entirely new nucleases; the synthetic proteins differed from natural ones by about 30%, edited genes in bacterial, plant and human cells, and in some cases outperformed their natural counterparts. We read the account of these studies, not the papers themselves.

Why this matters to Human Augmentation → Everything here was shown in cells. The researchers' own caveat is that the new tools still have to be tested in living animals to show they don't provoke an immune attack and can edit enough cells to treat disease. A better enzyme is a prerequisite for a therapy, not a therapy.
✦
QuantumCompany Claim · Unreviewed

Infleqtion Reports 30 Logical Qubits, With an AI-Found Gate and an Asterisk

On September 24, 2026, Infleqtion announced 30 entangled logical qubits encoded in 80 physical atoms on its Sqale neutral-atom machine, which it calls a first for a neutral-atom company on a commercial system. The company says the result was confirmed by a signal about 1000 times stronger than the underlying noise, and that it used a logical operation discovered with OpenAI's GPT 5.6 Sol that halves the physical gates needed for a key logical operation. Two qualifiers matter. A technical post by the company's CTO, as described by independent coverage at PostQuantum, acknowledges that the code used is distance-2: it detects errors but cannot correct them, and the experiment relies on discarding runs where an error was flagged. And no peer-reviewed paper exists yet; the company promised one in the coming weeks.

Why this matters to Compute → Counting logical qubits is only meaningful alongside what the code can do about errors. The genuinely new item is a language model proposing a better gate sequence, which is checkable independently of the company's marketing.
✦
FusionExperiment · Published

PACMAN Let Machine-Learning Models Run Parts of a Real Tokamak

On September 8, 2026, the American Nuclear Society's news service reported that researchers at the Princeton Plasma Physics Laboratory and Princeton University had built PACMAN, a software framework that chains machine-learning predictors and controllers together, and tested it at the DIII-D tokamak in San Diego. The full framework runs in about 20 milliseconds per cycle, and in one test a model predicted a tearing mode, a damaging instability, roughly 200 milliseconds before it would have occurred, in time to adjust the plasma instead of reacting afterward. The work is published in the journal Nuclear Fusion. The testing was a handful of scenarios at one machine; use on other tokamaks is still prospective.

Why this matters to AI → A plasma changes too fast for a person to steer, so a learned model is the only thing that can. The point of PACMAN is that the pieces are modular, so a new predictor can be added without rebuilding the control system.
✦
NeuralClinical · Company Demo

Neuralink Shows a Third ALS Participant Speaking Through a Decoder

In mid-September 2026 (outlets disagree on the exact day, between the 16th and 18th), Neuralink posted a video of Terry, who has bulbar-onset ALS, the form that hits speech first, using its implant in the company's VOICE trial. Machine-learning models decode the speech he attempts into words that are spoken in a synthetic voice; coverage says it is modeled on his own voice from before the disease. He is reported to be the third participant in the company's voice study. Neuralink says its devices are investigational and not approved by regulators, and the video came with no peer-reviewed data, error rates or latency figures, so what viewers can judge is a demonstration, not a measurement.

Why this matters to AI → Speech decoding is a model trained on one person's signals and one person's old recordings. Whether it generalizes across patients, and how often it is wrong, are the numbers that will decide whether this is a product.
✦

The Thread

The common element this week is a model standing in for a human specialist: one that proposes proteins, one that proposes a gate sequence, one that holds a plasma steady, one that turns brain signals into a voice. The sources are not equally strong. The tokamak result has been through a journal; the quantum and Neuralink items are announcements, and the genome item is a lab result in cells.

PACMAN's authors put the case for the whole pattern most plainly. Princeton's Hiro Farre Kaga is quoted as saying machine-learning models "are the only way we have to model the plasma in millisecond times" Opinion. That is a claim about one field, but it describes why the same tool keeps showing up elsewhere: where the data is too fast or too large for a person to read, a model is the only reader available.

✦

Sources

— End of Issue —
All Issues
  1. 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.
  2. 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.
  3. 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.
  4. 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.
  5. 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.
  6. 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.
  7. 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.
  8. 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.
← Back to the dashboard