Read The Day

Published edition9 September 2026

AI tackles an old fluid mystery as robots learn new moves

A formally checked proof leads the AI report. Robotics ranges from ceiling walking and surgical mapping to open world models and soft-object simulation.

Artificial intelligence · 9 September 2026

AI, understood.

5 min read

A public formal proof leads seven verified developments spanning images, genomics, chemistry, sleep, generated video and real quantum hardware.

Repository card for the public Lean certificates accompanying the Navier-Stokes and Euler results

An AI system produces a formally checked fluid proof

OpenAI released an AI-generated proof that a smoothly forced three-dimensional fluid can develop an infinite-speed singularity in finite time. The 166-page argument comes with a public Lean formalization. That makes the logic machine-checkable, but mathematicians still need to confirm that the formal statement and interpretation fully match the famous problem.

Why it matters

A machine-generated result on a 90-year-old Millennium problem is an unusually strong test of whether frontier systems can produce new, checkable mathematics rather than only summarize it.

Keep in mind: The mathematical community has not yet completed independent expert review. Formal checking validates the encoded statements and proof steps, not whether the formalization captures every intended interpretation.

Read the source · OpenAI

Nine billion DNA changes get an AI map

Google DeepMind opened a one-petabyte atlas containing predicted effects for nine billion possible single-letter DNA changes. Scientists can browse it or query an API instead of running each prediction separately. The entries are model outputs, not laboratory measurements, although collaborators report experimental confirmation for selected rare-disease variants.

Why it matters

Scientists can rank candidate variants and inspect predicted mechanisms across coding and non-coding DNA without first running billions of separate model queries.

Keep in mind: Most entries are predictions, not laboratory measurements. Benchmark and collaborator-result framing comes from Google DeepMind.

Read the source · Google DeepMind

An AI agent calibrates real quantum chips

An MIT researcher connected GPT-5.6 Sol through Codex to the software controlling a six-qubit chip. The agent chose parameters, ran measurements, and completed routine calibration sequences when signals were clear. Weak or noisy signals still needed an experienced researcher, so this is useful lab assistance rather than unsupervised discovery.

Why it matters

A researcher can leave well-defined measurements running for hours while reserving human attention for ambiguous signals and experiment design.

Keep in mind: The evidence is a vendor case study rather than an independent evaluation. The agent handled routine calibration, not unsupervised scientific discovery.

Read the source · OpenAI and MIT EQuS

ChatGPT Images gets faster, steadier editing

OpenAI launched a new image model with more reliable targeted edits, steadier multi-step refinement, and up to 50 percent lower generation latency than its previous version. It is rolling out across ChatGPT, Work, Codex, and the API. The quality and speed comparisons come from OpenAI and early partners, not independent testing.

Why it matters

People can refine a picture over several turns without rebuilding it each time, while applications gain a faster default model and a slower precision option.

Worth doing: Test a familiar multi-step edit before changing a production image workflow.

Keep in mind: Quality and latency comparisons are vendor-reported; independent testing was not available before the cutoff.

Read the source · OpenAI

Chemistry agents face a real physics check

TSBench asks AI agents to propose three-dimensional transition states, then checks the structures with quantum chemistry instead of grading their prose. Across 546 evaluations, diagnosis-led retries raised aggregate success from 50.4 to 66.8 percent. The remaining failures show that plausible-looking chemistry can still follow the wrong reaction path.

Why it matters

Scientific agents can sound chemically plausible while following the wrong reaction path, and this test exposes that failure before an autonomous workflow trusts it.

Keep in mind: This is a preprint benchmark, not proof that the tested agents can plan useful laboratory syntheses.

Read the source · Xiaohu Xu and Tong Zhu

Two million hours of sleep train one model

Researchers trained SleepFM-2 on more than two million hours from 235,865 overnight recordings, then tested transfer across diseases, expert-scored sleep events, wearables, and subjective reports. The scale is unusual and the held-out tests are broad. It remains a preprint and does not establish clinical benefit or diagnostic approval.

Why it matters

One learned view of overnight brain, heart, muscle and breathing signals may support many downstream health studies instead of a separate model for each sensor and task.

Keep in mind: The paper is a preprint and does not establish clinical benefit or diagnostic approval.

Read the source · SleepFM-2 authors

Video models make convincing but wrong physics

PhysWeep measures the physical values that actually appear in generated videos, rather than asking whether motion merely looks believable. Two of three open generators repeatedly settled on seed-dependent but incorrect values. The audit covers only three models and six parameter sweeps, but it exposes a clean gap between visual plausibility and physical control.

Why it matters

A clip can look physically natural while ignoring the number or condition a user asked for, which weakens claims that today’s video generators understand physics.

Keep in mind: The audit covers three open generators and six parameter sweeps, not every video model or physical behavior.

Read the source · Rasul Khanbayov and Hasan Kurban

Robotics · 9 September 2026

Robotics, explained.

5 min read

Seven fresh studies connect striking hardware with safer control, open learning systems, surgical mapping and more useful simulation.

Flying humanoid prototype taking controlled steps against a ceiling

A flying humanoid walks across the ceiling

Researchers demonstrated a transformable flying humanoid taking steps against a ceiling in simulation and hardware. Its controller smooths thrust changes as hands and feet make or break contact. The result is a striking controlled prototype, not yet evidence that the machine can perform useful inspection work outside the lab.

Why it matters

A robot that can fly and then brace or walk on walls and ceilings could reach structures that ordinary drones or ground robots cannot.

Keep in mind: The first-of-its-kind claim is the authors’ assessment. The experiment is a controlled prototype, not a deployed inspection robot.

Read the source · Kazuki Sugihara and Kei Okada

A robot maps tissue for precise laser paths

OCTN uses a robot to scan tissue selectively, build an optical map, and plan a laser path for irregular treatment regions. The authors report 43 times fewer imaging queries and four times lower acquisition cost than dense scanning. Those are controlled research results, not clinical validation or evidence of safe autonomous surgery.

Why it matters

A surgical robot can reason over a smoother, faster tissue map instead of repeatedly traversing a blocky volume assembled from thin image slices.

Keep in mind: This is a research demonstration, not a clinical trial or approved surgical system. Performance figures are author-reported.

Read the source · Ravi Prakash and colleagues

One humanoid policy learns a soccer sequence

SkillX combines walking, ball control, and recovery in one humanoid soccer policy demonstrated on a Noetix E1 robot. Some skills learned from motion capture while others used onboard vision, then a shared policy connected them. The controlled pitch is a useful integration test, not proof of general athletic intelligence.

Why it matters

A single deployable policy can join several whole-body skills into a longer sequence instead of handing control among separate programs.

Keep in mind: Some trials use motion capture for the robot base pose and object positions. The demonstrations do not prove open-ended soccer play.

Read the source · SkillX authors

OpenWAM opens the robot world-model recipe

OpenWAM released a modular world-action training stack and a model pretrained on roughly 6,400 hours of human and robot video. The team reports tests across eight simulation benchmarks and physical robots. The open pieces make comparisons possible, although performance remains author-reported and varies by task and embodiment.

Why it matters

Researchers can test which parts of video prediction actually help robot control without rebuilding the whole pipeline for each idea.

Worth doing: Compare the released modules against one representative task before adopting the whole stack.

Keep in mind: Top-tier performance is author-reported and varies by benchmark and embodiment.

Read the source · OpenWAM team

Long conversations can make a robot ignore rules

A new benchmark ran language-model robot controllers through forty conversations of one hundred turns each, looking for rule drift and unsafe decisions. The published first layer shows that long dialogue can erode instruction following. Physical validation is still under way, so the current result is a warning signal rather than a complete safety verdict.

Why it matters

A robot’s high-level language model can become either too cautious or dangerously permissive as a conversation grows, even when low-level rules look fixed.

Keep in mind: Published results are primarily text-layer tests; physical Unitree validation remains ongoing. The benchmark is a preprint and does not represent all robot safety architectures.

Read the source · Aulon Bajrami, Mohamed Elshamouty and Werner Kraus

Geometry carries the robot; AI handles contact

GloVLA lets a geometric controller move a robot arm across open space, then hands control to a learned policy near grasp and placement. The authors report physical UR10e success rising from 35.6 to 90 percent. The project is under review and assumes known object locations and fixed handoff regions.

Why it matters

A robot can save learned-model effort for the moments that need contact judgment while reliable geometry handles the easy travel.

Keep in mind: Results are from the authors and the project remains under review. The method assumes object locations and fixed handoff regions.

Read the source · GloVLA authors

Videos teach a simulator how soft objects bend

PhysReal turns one ordinary video of a soft object into a simulation whose material behavior can be changed and replayed. It estimates geometry, motion, and physical parameters from a single view. The examples are promising research demonstrations, but single-view ambiguity remains and the work has not yet been independently reviewed.

Why it matters

Robots could practice with a digital version of a rope or soft object whose local behavior comes from the real item rather than a generic material preset.

Keep in mind: Authors are anonymous while the work is under review. The project page says the method shows potential; it does not establish broad real-robot deployment.

Read the source · PhysReal authors