DeepMind's AlphaGenome Atlas predicts all 9 billion DNA swaps
Two releases and one uncomfortable result about readers: a petabyte-scale map of every possible single-letter mutation, a $1.35 billion bet on running models in milliwatts, and new evidence that people can't tell AI fiction from human fiction — and rate it highest when they think a person wrote it.
Google DeepMind published the AlphaGenome Atlas, a precomputed predictive map of every one of the roughly nine billion possible single-letter DNA changes in the human genome. Each variant carries estimated effects on molecular processes across hundreds of cell types and tissues: the dataset runs to one petabyte, more than 30 times the size of the AlphaFold protein database. The point is the 98 percent of the genome that doesn't code for proteins — the switches and dials that decide when and where a gene turns on, where most disease-linked variants sit and where their effects have been hardest to read.
Because thousands of prediction values per variant are useless at the bench, DeepMind also shipped the AlphaGenome Variant Impact (AVI) score, a small neural net that folds AlphaGenome and AlphaMissense predictions together with two evolutionary-conservation measures into a single number. It runs on 18 input features where the standard tool CADD needs more than 150, and it beat existing methods on clinically classified variants, especially outside coding regions. In one GREGoR rare-disease case, AVI pushed a previously ambiguous DNM1 variant in a child with severe epilepsy to the top of the candidate list and predicted the mechanism — a bad splice site that lengthens the protein by 13 building blocks, but only in a brain-specific gene version, which is why blood-based work had found nothing. A lab experiment confirmed it. DeepMind is explicit that this is research tooling, not a diagnosis.
Analog Devices agreed to buy Alif Semiconductor for $1.35 billion in cash, with up to $200 million more contingent on performance. Alif builds AI-native microcontrollers and fusion processors that put a low-power neural processing unit on the same die as connectivity and power management, so a sensor can classify what it sees without waking a bigger chip or opening a radio link. ADI is framing the deal as "Physical Intelligence" — inference that runs locally inside power, latency and reliability budgets — and says Alif's silicon is already shipping in production with consumer and industrial design wins.
The price looks strange for a company whose parts sell for a few dollars, until you notice what's actually being bought: deterministic inference. A factory safety system or a medical device can't accept an answer that usually arrives in 40 milliseconds; it needs one that always arrives within a bounded time, which no network round trip can promise. That requirement, more than privacy or bandwidth, is what forces models onto the device — and it's why a fifty-year-old analog company with industrial customers is the natural buyer rather than a cloud provider. It's also ADI's second move in the same direction, after agreeing to buy power-delivery specialist Empower Semiconductor. What's undisclosed is the part investors would want: Alif's revenue isn't public, so the multiple is unknowable.
A study in Judgment and Decision Making finds readers rate AI-generated short stories higher than human ones — and highest of all when they believe a person wrote them. Across 2,587 adults, participants given AI stories they were told were human-authored gave the top scores for quality and absorption; when asked to spot which story was which, they managed 39.93 percent in one experiment (below chance) and 51.97 percent in another (no better than chance). AI literacy helped: each one-point rise in self-reported AI expertise lifted the odds of guessing right by 14 percent, and 33 percent on the validated Artificial Intelligence Literacy Scale. Literary expertise did nothing.
Senior author Deena Weisberg of Villanova reads it as a bias running the wrong way — people assume creative writing needs lived experience, so they underestimate what the models do, and the AI versions state their themes outright rather than making readers infer them. That last detail is the part worth sitting with: the preference may be less about quality than about effort, and it lands on the same week the industry argued about whether AI labs are moving too fast to be governed at all.
What to watch: whether the Atlas follows AlphaFold's path from free research resource to drug-discovery infrastructure — and whether anyone ships a disclosure regime that survives contact with the fact that readers can't detect the difference.
Would you rather read a story you know was written by a person, or one you simply enjoyed more? Tell us in the comments.
