Doubled in Cells, Simulated in Silicon, Promised in a Post: Four Fields, Four Distances from Proof
This week's four items sit at different distances from the lab bench. One is a peer-reviewed result in cultured cells. One is a simulation. One is a letter of intent for money. One is a founder's post on X. Only the first has a clear AI contribution; in the other three the AI link is absent, indirect or promised for later. We say which is which as we go.
A Protein Language Model Designed a DNA-Insertion Enzyme That Doubled Targeted Insertion in Cells
A team from Integra Therapeutics, Pompeu Fabra University and the Center for Genomic Regulation, with Marc Güell listed among the authors, fine-tuned the protein language model ProGen2 on PiggyBac transposase sequences. PiggyBac is an enzyme family that moves DNA segments; here it was paired with Cas9 in a system called FiCAT. The team generated more than 100,000 sequences and tested 22. All 22 showed excision activity and 7 beat the existing engineered enzyme, HyPB, at excision. One synthetic variant, number 3277, doubled targeted insertion in FiCAT experiments. In parallel, the group searched 31,565 eukaryotic genome assemblies for natural relatives and tested 23 in HEK293T cells; 9 were active and 2 performed comparably to HyPB. The work is reported as published in Nature Biotechnology, according to a secondary news write-up dated October 7, 2026, which is our only source for these details; we did not read the paper itself.
The limits are substantial. Results varied by cell type and task: in primary human T cells, variant 3277 only matched HyPB on non-targeted integration. These are measurements of laboratory gene insertion, not evidence that any treatment improved health, and higher activity does not show better precision or safety. The write-up says specificity is a crucial open question. This is a transposase, not a base editor.
NVIDIA Simulates a 104-Qubit Fault-Tolerant Protocol on 8,192 GPUs; No AI in the Description
On October 19, 2026, NVIDIA announced a simulation of a 104-qubit fault-tolerant processor on JUPITER, the supercomputer at the Jülich Supercomputing Centre, using 8,192 GH200 GPUs. It includes realistic noise, where Jülich's earlier 50-qubit work assumed noiseless conditions. The protocol prepares a magic state, a resource fault-tolerant machines need for certain gates, starting in a distance-3 color code and growing to a distance-7 hybrid code. NVIDIA says the approach needs about 90% fewer physical qubits than earlier methods and cut a weeks-long simulation effort to days. Jülich's director, Kristel Michielsen, is quoted on bringing that capability to quantum computing.
This is a classical simulation, not a quantum computer, so it shows how a design should behave, not that hardware does. The 90% figure and the "largest" label are NVIDIA's own, the baseline methods are not named in the coverage we read, and no logical error rates or accuracy figures were published with it. The Quantum Insider write-up of the release does not mention AI, machine learning or decoders. The only link to our theme is that the compute is GPU hardware built for AI.
Commerce Signs a Letter of Intent for Up to $85 Million to Commonwealth Fusion for Fusion Electronics
The U.S. Department of Commerce signed letters of intent for CHIPS R&D awards to four companies, totaling up to $389 million, according to trade coverage dated October 23 to 28, 2026. Commonwealth Fusion Systems (CFS) is listed for up to $85 million and Helion for up to $175 million. Per ANS Nuclear Newswire, CFS's money targets semiconductors and power electronics for its tokamak, including RF plasma-heating hardware, diagnostics electronics, radiation-hardened parts and faster embedded controllers for plasma management. Commerce would take a minority, non-voting equity stake in each company as a condition. CFS chief executive Bob Mumgaard is quoted saying the innovations "will be a vital part of our global effort to deploy commercial fusion."
A letter of intent is not a finished agreement; the amount is "up to," and no disbursement schedule has been reported. On AI, sources conflict: The Fusion Report's October 30 issue says the proposed work includes AI applied to SPARC and ARC, including a SPARC digital twin, but the ANS article does not mention AI or a digital twin, and we could not confirm it from the primary announcement. We therefore treat the AI link as unconfirmed. The separate SPARC digital-twin effort was covered in Issue 8.
Musk Says Neuralink's Blindsight Is "Ready From an Engineering Standpoint," Pending Regulators
On October 30, 2026, per Drive Tesla Canada, Elon Musk posted on X about Neuralink's first two products. Telepathy is meant to let users communicate with computers or people by thought; Musk wrote that its "output bandwidth will vastly exceed the fastest human speech or typing in a few years." Blindsight, meant to restore sight including to people who have never had it, "is ready from an engineering standpoint" and is awaiting regulatory approval. The same article reports Neuralink's own update on a 72-hour hackathon: automatic speech-intent detection, iPhone call integration, and a prototype in which thinking "Hey Neuralink" activates a Grok agent. Grok is built by xAI, a company Musk also founded, which is worth knowing when weighing the claim.
None of this comes with trial data. Blindsight received FDA Breakthrough Device designation in September 2024, which is not approval, and the coverage reports no evidence yet that it restores sight in humans. The Grok prototype is the only explicit AI element and the article does not say how the speech decoding works. Everything here is the company's or its founder's own account.
The Thread
Read side by side, the four items show how much of "AI in science" is ordinary engineering around it: a language model that proposed enzymes and a lab that tested them; a GPU simulation that mentions no AI; electronics funding; and an implant company's roadmap. Only the first has the test-and-count structure that lets a reader judge the claim. That is the standard we try to hold the rest to.
The one forecast on the table is Musk's, that Telepathy's output will exceed human speech or typing "in a few years" Forecast. It is a founder's prediction with no published bandwidth figure behind it.
Sources
- AI-designed PiggyBac enzymes: The Brighter Side of News, October 7, 2026
- 104-qubit simulation: The Quantum Insider, October 19, 2026
- CHIPS letters of intent: ANS Nuclear Newswire, October 28, 2026; The Fusion Report, October 30, 2026
- Neuralink: Drive Tesla Canada, October 30, 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.