The old company scaled by hiring people.

The next company scales through compute and refactoring.

Completely different organisms with different economics.

Cognixion's CHOICE study gives 30 people with ALS three ways to talk at home: eye-tracking tablet, AR-glasses BCI, or touchscreen.

Primary outcome: which one they choose to use.

The real benchmark for a neural interface: preference.

Cognixion: prnewswire.com/news-releases/…

Inside a mixture-of-experts model, sparse autoencoders show who actually does the work: separate experts for Java and front-end web code, for geometry and for arithmetic.

The black box is starting to have an org chart.

ByteDance Seed: arxiv.org/abs/2611.08089

NASA picked Astranis, Global Enterprise Solutions, and Viasat to develop commercial Ka-band relay concepts as it retires its TDRS fleet.

Even the network that talks to spacecraft is becoming a service you buy.

NASA: nasa.gov/news-release/n…

Brain decoding + spinal stimulation + a hybrid exoskeleton for chronic stroke.

In an early pilot, one 67-year-old patient walked without a walker after 22 days.

Not a cure. A glimpse of the nervous system getting external support.

Tianjin University: news.tju.edu.cn/info/1003/6196…

3D-printed tungsten tiles went inside the DIII-D tokamak, faced 74 high-confinement plasma discharges, and came out intact.

One of fusion's hardest problems is the wall around the star. Now we can print the wall.

NC State/ORNL/JEOL via ANS: ans.org/news/2026-11-0…

A mineral supplement as a senolytic: vanadyl sulfate selectively killed senescent cells. In aged male mice, 3 months in drinking water raised median lifespan to 952 days vs 886.5.

Mice, not humans. But aging keeps looking like engineering.

Nature Aging: nature.com/articles/s4358…

Tokens aren't equally expensive: each enters the KV cache and raises the cost of the next.

The SIMPLE law predicts inference energy across 13 LLMs without per-model fitting. Bigger reasoning budgets drive super-linear energy growth.

SIMPLE law paper: arxiv.org/abs/2611.03330

OpenFrontiers maps 12,151 open research problems.

Closed-book, the best of seven agents rediscovered at most 7 of 40 recently solved ones. With live web, success mostly meant finding the published answer.

The frontier is now a benchmark.

OpenFrontiers: arxiv.org/abs/2610.37739

Open-HumanScale: 50,000 hours of egocentric human video and a log-linear scaling law from human video hours to robot success.

Pretrained only on human footage, then robot post-training: 92.08% on RoboTwin 2.0 Clean2Clean.

Our hands are the dataset.

PKU: arxiv.org/abs/2611.06817

Agents don't need to read tool outputs as text. Coca-Cola compresses them into latent tokens: 70–89% fewer agent tokens, 24–56% faster runtime, and better task performance.

Machines are starting to talk to machines in their native tongue.

Tsinghua: arxiv.org/abs/2611.06500

A tool-rich autonomous agent plateaued at 0.50x FlashAttention-3 after 133 steps.

KernelSeek, a plain coding agent asking itself short questions at decision points, reached 1.13x.

The bottleneck wasn't tools. It was direction.

Snowflake: arxiv.org/abs/2611.03517

SwarmEvolve: hundreds of cheap agents reason in latent space and share KV caches, not text.

Median 3–8 min to hit the best scores rivals reach in 60. With an 8B model it beats some AlphaEvolve solutions.

Discovery stops talking out loud.

Brown: arxiv.org/abs/2611.08483

"Shall We Approve?": Caper runs a coding agent's shell command in a snapshotted sandbox first, shows an LLM what happened, then commits or rolls back.

Harmful outcomes: 6.7% on SABER.

Permission after consequence, not before hope.

SMU: arxiv.org/abs/2610.40753

When there are no jobs nature finds a way.

Video generation meets the agent harness: multi-shot production, interactive world models, and paths toward recursive self-improvement of the generator itself.

The clip is no longer the end of the loop. The loop is.

Video Rebirth/NUS: arxiv.org/abs/2610.40191

Hybrid Computer Use: agents pick among code, accessibility trees, and GUI—with live-state guards so stale actions die.

Highest Partial and Binary on OSWorld 2.0 among compared systems.

The desktop stops pretending to be a screenshot.

Microsoft Research: arxiv.org/abs/2610.41432

Axion 1.0: open agent foundation model in three stages—two light SFTs, then RL. No mid-training, no DPO, no on-policy distillation.

Recipe, data, prompts, and code all released. Strong agents don't have to be closed.

ByteDance/Duke: arxiv.org/abs/2611.03233

Avatar Intelligence: persistent personal agents that carry your goals and values—less routine supervision, you keep the consequential calls.

L0→L3 in delegated domains: human-for-the-loop to out-of-the-loop.

Your digital self, on the clock.

SJTU/NUS: arxiv.org/abs/2611.03092

Recursa: adapts to new domains from its own experience—no experts, no stronger teacher.

Parallel learners propose skills; the system validates and merges. Held-out all-pass: 30% → 55%.

The mind that rewrites its playbook.

Alibaba Research: arxiv.org/abs/2611.04035

CoworkOS: the model decides; the runtime remembers—docs, CAD, Blender, long jobs.

Five RSI post-training rounds: 33.19 → 42.95 across eight benchmarks.

The coworker that trains the next coworker.

Renmin/Fudan: arxiv.org/abs/2611.03178