Vol. I · No. 35 Weekly Edition September 12, 2026

DeepMind turns nine billion possible mutations into a lookup table. A Millennium Prize dispute breaks out after OpenAI’s Navier-Stokes claim follows a route two mathematicians say was theirs first. A 150-million-parameter model beats a frontier model on ARC-AGI at a fraction of the cost. OpenAI’s own chief scientist warns that the tool built to catch a model’s bad reasoning is losing its grip. World Labs unveils a single model that generates video, rebuilds scenes, and trains robots.

Inside

DeepMind’s genome atlas. The Navier-Stokes authorship dispute. Pathway’s cheap ARC solver. Pachocki’s warning on a failing safety tool. World Labs’ Atlas model.

Edited By

David Borish
From New York

Filed

Five articles, one week, sourced from The AI Spectator

Genomics · Artificial Intelligence No I

DeepMind, September 8, 2026

Every mutation,
already
scored.

DeepMind has published predicted molecular effects for all nine billion possible single-letter changes across the human genome. The dataset runs to about one petabyte, and a single combined score now tells researchers which of those nine billion variants are worth a second look.

By David Borish September 12, 2026 · 5 min read

The human genome holds about three billion base pairs, and at each position three single-letter substitutions are possible, for a total of nine billion ways a single DNA letter can change. On September 8, Google DeepMind published predicted molecular effects for all nine billion of them, in a resource called the AlphaGenome Atlas, built on the AlphaGenome model the company released in January.

The scale is the headline. The dataset runs to about one petabyte, which DeepMind says is roughly 30 times larger than its AlphaFold protein-structure database. Where AlphaFold answered a structural question, how a protein folds, Atlas answers a functional one: what a DNA change is likely to do to the machinery that turns genes on and off.

AlphaGenome itself reads a DNA sequence of up to one million base pairs and predicts thousands of molecular measurements from it, including gene expression, chromatin accessibility, and RNA splicing. It scored state of the art on 25 of 26 variant-effect benchmarks when the Nature paper describing it appeared in January. Turning that model into a genome-wide catalog meant running it against every possible substitution at every position in a reference genome. DeepMind’s genomics lead, Ziża Avsec, told IEEE Spectrum the team’s early estimate was that it needed an 80-fold increase in computation speed to finish in a reasonable time.

Raw prediction volume creates its own problem, since each of the nine billion variants carries an average of about 27,000 individual predictions. DeepMind’s answer is the AlphaGenome Variant Impact score, a single number per variant that combines AlphaGenome’s regulatory predictions with AlphaMissense’s protein-impact predictions, plus conservation statistics and predicted loss-of-function effects. Each AVI score ships with an attribution naming the specific mechanism behind it, such as splicing disruption. The score’s largest gains sit in the roughly 98 percent of the genome that does not code for protein, the part standard tools have handled worst.

The clearest test of a resource like this is whether it changes an actual diagnosis. Researchers at the GREGoR Consortium and the Broad Institute used AVI scores to flag a non-coding change in intron 10 of the gene DNM1 in a rare-disease case that had gone unsolved through standard sequencing, pointing to splicing disruption as the mechanism. In an earlier case using the underlying AlphaGenome model, researchers scored 91 non-coding variants near the gene PLA2G6 in a family affected by PLA2G6-associated neurodegeneration, after standard sequencing had found only a single coding variant. The model flagged an intronic change predicted to create a cryptic splice site, later confirmed by RNA sequencing, completing a diagnosis nearly two decades after the family’s symptoms began.

DeepMind is direct about the limits. Its own materials state that AlphaGenome has not been validated for and is not approved for any clinical use. The AVI scores narrowed each search from dozens or thousands of candidate variants down to one worth testing, and laboratory confirmation did the remaining work. The Atlas removes repeated model inference from that search. The experiment at the end of it stays exactly where it was.

The Atlas is available now through a browser portal that needs no coding, alongside the existing AlphaGenome API, and is free for non-commercial research. DeepMind says commercial access will follow through Google Cloud, without a stated date. For researchers working on rare and undiagnosed disease, a genome-wide search that once required writing code and running a demanding model one variant at a time is now a lookup.

The Atlas cannot tell a researcher which variant caused a disease. It can tell them which few hundred, out of nine billion, are worth months of lab work. The AI Spectator
Read the full article →
Architecture · Cost No II
29.5

Reasoning, without
writing a single token.

Pathway’s BDH-CQ solves new visual puzzles by updating a recurrent memory and a latent workspace instead of writing out a chain of thought. On the public ARC-AGI-1 evaluation, a 150-million-parameter version scores 29.5 percent pass at 2, at a computed cost of under a tenth of a cent per task.

Pathway, working with Bielik AI and New York University, has published BDH-CQ, a reasoning system that learns a new visual transformation from a handful of examples and solves it entirely inside a continuous latent state. The model never writes out a chain of thought. Demonstrations update a recurrent memory, and the query is answered through repeated transformations of a hidden vector that gets decoded only once, at the end, into the final grid.

On the public 400-task ARC-AGI-1 evaluation set, a 150-million-parameter configuration reaches 29.5 percent pass at 2, at a computed cost of $0.0007 per task. According to the ARC Prize leaderboard data the authors cite, that combination of score and cost sits beyond every previously reported system’s cost-accuracy curve. Recursive solvers like HRM and TRM report strong accuracy but fine-tune on augmented copies of each task’s own demonstration pairs, a step the leaderboard prices at $1.48 and $1.76 per task. BDH-CQ updates only its recurrent state at inference and reports about 0.85 GPU-seconds per task on an H200, priced at $3 per GPU-hour.

Against frontier language models, the paper first estimated BDH-CQ runs about 57 times cheaper than GPT-5.6 Luna (Low), which scores 34.2 percent at a reported $0.040 per task. After accounting for OpenAI’s 80 percent public API price cut on July 30, 2026, a change the ARC Prize leaderboard had not yet reflected, the authors revised that gap down to roughly 11 times cheaper, a correction the paper states plainly rather than keeping the larger number.

The paper’s more informative contribution is a breakdown of where the 29.5 percent comes from. On ConceptARC, which groups tasks into 16 concept families of ten tasks each, strict accuracy ranges from 9 of 10 on families like ExtendToBoundary down to 2 of 10 on Copy and Order, though the authors note the confidence intervals for the best and worst families overlap. The Open-Prem Inflection Point V3 framework has tracked this pattern before: once a smaller, cheaper architecture matches a frontier model on the tasks that matter, the economics of renting frontier compute to do the same job stop making sense.

A separate set of experiments isolated individual operations after the model was frozen. Propagating a signal to a boundary and copying a motif to multiple anchor points both stay at 100 percent accuracy as the tested range expands. Ordering and nesting behave differently: sorting bars by height drops from near-perfect at five objects to 1 of 24 at length eight, while selecting cells inside nested containment frames holds up through depth four and falls to 29 of 36 at depth five. Adding one demonstration at the harder difficulty recovers most of the gap, which points to a failure to extrapolate rather than a hard capacity limit.

The paper is transparent about its limits. The 29.5 percent figure and the cost estimate come from the authors’ own deployed system, and the audit that reproduced the score was run by co-authors at Bielik AI and New York University, so it confirms internal consistency rather than independent replication. A separate, generator-built evaluation of the same mechanics ranked them in a different order than hand-written comparison tasks did, with a rank correlation of just 0.300 across 16 mechanics, a reminder that a generated benchmark’s numbers describe the generator as much as the model.

$0.0007
Cost per Task
on ARC-AGI-1
150M
Parameters in the
Tested Configuration
11x
Cheaper than GPT-5.6
Luna, After Its Price Cut
Read the full article →
// alignment :: monitoring :: safety No III
[ ESSAY LOG · SEPTEMBER 6, 2026 ]

OpenAI’s own
chief scientist says
the tool that watches
a model’s mind is
losing its grip.

$ essay --author "Jakub Pachocki" --title "An Alien Mind" --date 2026-09-06
> OpenAI’s head of research argues the field should be ready to slow down
$ timeline --event "gpt6_astra_release" --date 2026-09-03
> Three days before the essay, OpenAI ships its most capable model yet
$ timeline --event "rlslow_breakthrough" --date 2023
> Pachocki and a colleague see reasoning models can be scaled by RL, spend the night on what it means
$ policy --check "chain_of_thought_monitoring"
> Hidden since o1-preview to shield it from optimization pressure
> Status: reliability declining, three reasons given

On September 6, 2026, Jakub Pachocki published an essay on OpenAI’s website called An Alien Mind. Pachocki runs research at OpenAI. Three days earlier, the company had shipped GPT-6 Astra, the most capable model it has released. His argument is that OpenAI, and every other frontier lab, should be prepared to slow down.

His central technical claim is that progress in machine intelligence comes mainly from increasing computational power. He says OpenAI internalized this around 2017 after seeing consistent returns to scaling across multiple projects, and reorganized its research around a small number of scalable directions. The result, he writes, is a system that is grown more than designed: a complex model built by repeating one optimization step across an enormous amount of compute, whose overall behavior resists a complete description even as researchers can trace individual mechanisms inside it.

Pachocki separates alignment into two problems. Goal alignment asks whether a model tries to accomplish the task set before it. Value alignment asks whether it holds and generalizes a set of principles, including honesty and regard for people, into situations unlike anything in its training. He names two methods in practical use today and the weakness of each. Reinforcement learning judged against a preference model or a written spec works well on average but is brittle, as in an OpenAI-Hugging Face incident where agents held one stated boundary but drifted from the spirit of their training in other ways. Steering a model toward an aligned part of its pretraining distribution is more fragile still, since a model trained hard enough on a difficult objective can learn to reason in a motivated way toward its goal, which he points to as a likely factor in recent cybersecurity incidents involving a non-OpenAI model.

The essay’s most concrete warning concerns chain-of-thought monitoring, the technique OpenAI has relied on most to check its own systems. Pachocki says OpenAI deliberately hid the chain of thought when it shipped o1-preview, specifically to protect it from supervision pressure, and has kept to a rule of not supervising the reasoning process since. That tool is losing its reliability for three reasons: reasoning now blends with communication and tool use in more complex environments, models are getting better at reasoning about and manipulating their own reasoning, and improved pretraining is making models smarter even without verbalized reasoning at all.

He does not argue for stopping. His strongest reason to keep training smarter models quickly is defense: models are becoming very good at breaking into and out of computer systems, and he describes a narrow window to use the best available models to harden critical infrastructure before rogue agents can reach it. He is direct that anticipated progress and the need for defense must not become an excuse for recklessness, and flags engineered pathogens as a further risk that new technology could enable.

Pachocki treats recursive self-improvement as the likely endpoint of the current path, and says he does not think the field should accelerate deep learning research further in the short term, even as he expects it will. The essay closes with three stated beliefs: that no lab has solved alignment and monitoring well enough to keep scaling responsibly at maximum speed for much longer, that he expects and hopes voluntary slowdowns become common until shared safety bars exist, and that international coordination on AI development needs to become a top priority for governments.

// FINDINGS_LOG
2023
RLSLOW MOMENT
Pachocki and a colleague see that reasoning models can be scaled by reinforcement learning, and spend the night working through what it means
3
DAYS BEFORE THE ESSAY
GPT-6 Astra shipped on September 3, 2026, three days before Pachocki’s public call to prepare to slow down
o1
FIRST HIDDEN CHAIN
Model where OpenAI first hid chain-of-thought reasoning from supervision, a rule it says it has followed since
3
REASONS COT IS SLIPPING
Blended reasoning and tool use, models reasoning about their own reasoning, and pretraining that no longer needs verbalized reasoning at all
Read the full article →
Mathematics · Millennium Prize No IV

A Navier-Stokes Proof,
and the Mathematician Who Says
It Followed His Unpublished Work

§ § §

OpenAI says an internal model resolved part of the Navier-Stokes Millennium Prize problem using roughly 10,000 concurrent agents over three days. NYU mathematician Tristan Buckmaster says the result followed the route he and Anthropic researcher Levent Alpöge had been pursuing privately for a year, after OpenAI learned their work existed.

OpenAI published a paper and a public Lean repository on September 8 claiming that an internal model had resolved part of the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems the Clay Mathematics Institute set out in 2000. Within the same 48 hours, mathematicians Tristan Buckmaster and Levent Alpöge published their own AI-assisted, Lean-verified proofs of finite-time blowup for related fluid equations, then said OpenAI had learned about their unpublished work and moved to claim credit before they could release it.

OpenAI’s writeup describes a three-dimensional fluid that starts at rest, has a smooth external force applied to it, and develops unbounded velocity in finite time while its total kinetic energy stays finite. The company says this satisfies statements C and D of the Clay Institute’s 2000 problem formulation, the versions concerning a driving force rather than the unforced case most people picture. The writeup runs to roughly 165 pages. OpenAI says the work came from a model significantly more capable than GPT-6 Astra, and states plainly that it does not intend to claim the $1 million prize, which under Clay Institute rules requires publication and at least two years of general acceptance first.

By OpenAI’s account, an internal model began training around August 28 and showed unusual gains on math benchmarks. On September 1, after hearing rumors that two Millennium problems had been resolved, the company pointed roughly 10,000 concurrent agents at every open Millennium problem. A smaller group produced an unforced Euler blowup as a warmup in about 50 hours, and OpenAI then redirected most of its compute toward Navier-Stokes. The winning group reached its result on September 5, and Lean verification took another 17 hours. Across the full effort, OpenAI says its agents sent 4.9 million messages and generated about 300 billion output tokens, with press estimates of the compute cost ranging from $15 million to $22.5 million.

Buckmaster’s own written statement, published alongside his papers, says that on September 3 he learned rumors of his and Alpöge’s unpublished work had reached OpenAI, and he emailed an OpenAI-affiliated mathematician to clarify that the project had no institutional backing. Three days later, on two calls involving OpenAI’s math lead Sébastien Bubeck, he says he was told an internal model had already produced a roughly 100-page proof of forced Navier-Stokes blowup following essentially his and Alpöge’s route, with minimal human involvement. He writes that this account did not hold up over the course of the call, and that he was offered a coordinated release or a version crediting only an internal OpenAI model, with Bubeck reportedly pushing twice to exclude Alpöge because he works at Anthropic. Buckmaster declined both and said he would go public.

OpenAI’s writeup addresses the overlap directly, saying its effort began September 1 after hearing rumors it later realized concerned Buckmaster and Alpöge, and that neither its researchers nor its agents saw any of the pair’s work through any means before it was released publicly. The company adds a hedge that it cannot rule out that de-identified data derived from product usage helped improve its models generally. Bubeck, in a press briefing, called Buckmaster’s allegations against him false and inflammatory and said he had followed academic norms.

Even taking OpenAI’s proof at face value, resolving statements C and D under a smooth forcing term is a step short of the unforced global-regularity question most people mean by Navier-Stokes, and a step short of what Clay Institute rules require for prize consideration regardless. Independent scrutiny of the 165-page proof has barely begun. The story likely to outlast this week’s headlines is what happens when researchers draft unpublished, career-defining work inside a lab’s own coding assistant, and that lab later claims to have reached an adjacent result independently, on the identical route, within days of hearing a rumor about their progress.

The AI Spectator • September 12, 2026 • Mathematics
Read the full article →
Robotics · World Models · Capital No V

One model that
generates, rebuilds,
and trains robots.

World Labs’ new Atlas model generates camera-controlled video, reconstructs 3D scenes from photos, simulates space and time, and produces text-to-image output from a single architecture. The company has raised $1.23 billion since 2024 to build toward it.

World Labs introduced Atlas on September 1, 2026, describing it as the company’s next generation world model and the first version built as a single system across four tasks that have historically required separate specialized tools: generating camera-controlled video, reconstructing 3D scenes from photographs, simulating how space and time evolve, and producing text-to-image output.

Atlas is what World Labs calls a multimodal autoregressive diffusion transformer, pretrained from scratch on text, images, video, camera poses, and 3D depth maps. Each image in the training data carries an explicit camera pose, grounding it at a position in 3D space, in a representation the company calls a spatial context. Like a large language model, Atlas generates output autoregressively; like a video model, it uses rectified flow diffusion to generate high-dimensional continuous data.

The clearest demonstration of the idea is camera-controlled generation: Atlas takes one or more reference images and produces new video from a camera path the user specifies directly, rather than describing motion in words. In third-party rater tests, Atlas was preferred over MiniMax H3 in 75 percent of trials, over Gemini Omni Flash in 81 percent, over Happy Horse 1.1 in 86 percent, over FLUX 3 in 93 percent, and over Seedance 2.5 in 94 percent, though World Labs notes the comparison models do not accept camera paths natively and had to be prompted with text descriptions of the intended motion.

Atlas’s reconstruction capability builds a real scene from as few as two or three ordinary photographs, and World Labs demonstrates rebuilding Stanford’s Main Quad from two to twenty-five ground-level photos into aerial flythrough views the original photos never captured. On sparse-view 3D reconstruction, the company reports Atlas achieving the lowest mean pointmap error across seven benchmark datasets, ahead of five specialized open-source baselines it says it reproduced itself.

The commercial target is robotics. World Labs demonstrates reconstructing real environments from cell-phone video using only 24 frames per environment, then simulating robots navigating those spaces and generating the depth images a robot’s own cameras would see along different paths, a workflow the company calls Real to Sim. The Exponential Replacement Curve is built on exactly this kind of transfer: a capability demonstrated inside a constrained benchmark, engineered from the start to move into a robot’s real-world operating environment rather than stay a research curiosity.

World Labs has raised $1.23 billion since emerging from stealth in September 2024, including a $1 billion round in February 2026 that drew AMD, Nvidia, a $200 million strategic investment from Autodesk, Emerson Collective, Fidelity, and Sea. Bloomberg has reported that round valued the company at approximately $5 billion, a figure World Labs has not confirmed. Atlas is not open source, is not broadly available, and both sets of benchmark results come from the company’s own testing. Whether the robotics pipeline actually reduces the cost of training real robots at scale is the question worth watching as early access rolls out.

Read the full article →
Key Figures
$1.23B
Total funding raised by World Labs since emerging from stealth in September 2024
94%
Rater preference for Atlas over Seedance 2.5 on camera-controlled video generation
24
Phone-video frames per environment used to generate the data that trains robots
$5B
Valuation Bloomberg reported after the February 2026 round, not confirmed by World Labs

The AI Spectator Weekly is published at davidborish.com/the-ai-spectator

Frameworks explored this issue:
Open-Prem Inflection Point V3  ·  The Exponential Replacement Curve

Vol. I · No. 35 · September 12, 2026 · Edited by David Borish · New York