Skip to content

What changes in your working day?

Work.

Follow what people can delegate, build and do differently.

This robot fixes the experiment when the light goes out

MIT’s optical robot places components, measures the beam and adjusts again. Its laser demonstration explains why recovery matters more than a perfect first move.

Robotics 30 Sep 2026 3 min read

AI tutors passed the one-hour test. What about the next month?

A randomized GRE study found immediate learning gains comparable to human tutoring. It did not test retention, classrooms or solo practice.

AI 30 Sep 2026 4 min read

A robot can ace the lab and still fail the library

Three years of tests in parks, an underpass and a library show why task scores, interaction, safety and economics must be judged together.

Robotics 30 Sep 2026 4 min read

The cheapest memory for an AI agent may be forgetting

CliffCompaction drops old tool output instead of summarizing it. Coding benchmarks improved while model-usage costs fell—but lost history can still matter.

AI 30 Sep 2026 4 min read

This robot learned from demonstrations humans kept failing

GLIDE writes code that filters difficult teleoperation before it becomes training data. Three lab tasks improved sharply, but the guardrails remain part of the deployed capability.

Robotics 22 Sep 2026 4 min read

AI learns from the searches it already paid for

Dream-RSI replays earlier attempts to improve the code directing new searches. The foundation model stays fixed—and fewer agent calls do not prove lower total cost.

AI 16 Sep 2026 3 min read

Faster AI search starts before you ask

Google’s Retrieve-for-Train research moves expensive search planning into training. The aim is a useful collection of results, delivered faster.

AI 16 Sep 2026 2 min read

Digit 5’s most important move may be stopping

Agility’s next humanoid is designed to work close to people. Safe interruptions and predictable behavior will matter as much as lifting ability.

Robotics 16 Sep 2026 2 min read

AI can propose new materials. The lab decides what works.

MatBrain narrowed 30,000 proposed structures to 38 candidates in 48 hours. Human researchers chose what to test. The useful question is how much work that shortlist saves.

AI 14 Sep 2026 3 min read

One video can teach a robot. The whole shift is another matter.

Skild and Generalist are making robots faster to teach. Their results reveal why learning a task and delivering reliable work are different engineering problems.

Robotics 14 Sep 2026 5 min read

What Work Should You Hand Over to GPT-6 Astra?

Astra promises more capable agents. OpenAI's own deployment simulations show why the job, its permissions and the proof of completion need separate attention.

AI 5 Sep 2026 4 min read

Claude Went Into the Lab. The Bubbles Needed a Human.

A protein assay reveals the gap between connecting laboratory instruments and producing reliable scientific measurements.

AI 5 Sep 2026 3 min read

The Vastkind Briefing

Stay ahead
of the story.

A personal take on the week. By Dorian.

Occasional editions. Free to subscribe.
Unsubscribe anytime.