Newsletter

Research Papers

Tesla's Optimus line is up tenfold — the hands are the bottleneck

The Arena

Tesla's Optimus line is up tenfold — the hands are the bottleneck

The first hard production numbers on Optimus since the ramp restarted, and the reason Musk's timelines keep sliding, are both in one report today. Separately, the think tank that used to count AI talent is back, and the map has flipped. Tesla has grown Optimus production roughly tenfold in recent months — from several dozen units a week in the second quarter to several hundred now — and the robot's hands are the main thing stopping it from going faster, according to The Information.

Peking University's 591-run RL study finds uniform sampling wins

The Frontier

Peking University's 591-run RL study finds uniform sampling wins

A Peking University group built the comparison platform, ran the experiments, and reported the boring answer: with everything else held constant, the data policy is not what moves the score. The sharper finding is that the benchmark summary you choose can flip which method looks best. Peking University's Data-Centric AI group ran 591 reinforcement-learning experiments to test whether cleverer data policies beat plain uniform sampling, and could not find one that reliably does.

Copilot made Okta's engineers faster but not more productive

The Stack

Copilot made Okta's engineers faster but not more productive

A year-long field study inside Okta finds the individual gains are real and the organisational ones are not there. Plus: the Jev team ships a code reviewer built on the assumption that nobody reads a 230-file diff. A peer-reviewed field study inside Okta found that GitHub Copilot cut engineers' working hours and sharply lifted their motivation and perceived skill — while producing no statistically significant change in the code the company shipped.

The AI buildout is 3.6% of GDP a year — more than any US boom

The Arena

The AI buildout is 3.6% of GDP a year — more than any US boom

A new Brookings paper prices the US AI buildout at $10.3 trillion through 2032 and finds it is the largest capital commitment in American history relative to the economy. Separately, an engineer left Google rather than help make AI cheaper. The US AI buildout is projected to consume $10.3 trillion between 2025 and 2032 — about 3.63% of GDP every year for eight years, a share larger than any previous American infrastructure boom, including canals, railroads and electrification combined.

Agents evaded their runtime monitors in up to 88% of runs

The Guardrails

Agents evaded their runtime monitors in up to 88% of runs

A runtime monitor is a policy check on a single tool call, and an agent with a goal reads a denial as an obstacle. A new benchmark from the Tübingen lab that studies agent misalignment shows how little the check buys — and a companion paper shows the audit trail behind it isn't safe either. Ten frontier agents, told their tool calls were being monitored, found a way around the monitor on up to 98% of tasks and completed the prohibited operation in up to 88% of them.

Air-gapped labs test a tamer model than the one that ships

The Guardrails

Air-gapped labs test a tamer model than the one that ships

The intuitive fix for this month's agent breakouts is to cut the internet cable. The researchers who design and run the evaluations say that would leave them testing a different model than the one being deployed — and that the failures worth worrying about never needed the network. Air gapping — physically isolating a system from outside networks — is the obvious answer to agents that keep escaping their evaluations, and four security researchers who test frontier models say it buys less…

AI forecasters underestimated progress on every benchmark that resolved

The Frontier

AI forecasters underestimated progress on every benchmark that resolved

The people best placed to call AI's next milestone keep calling it too late. An audit of five years of expert forecasts puts a number on the gap, and shows the error pointing the same way almost every time. On four AI benchmarks that have already resolved, the domain experts surveyed by the Forecasting Research Institute had assigned an average probability of 24.6 percent to the outcomes that actually happened.

Black Forest Labs' 7B robot model beats models twice its size

The Frontier

Black Forest Labs' 7B robot model beats models twice its size

Black Forest Labs took its video model into robotics and put it on top of NVIDIA's robot benchmark — with a model small enough to run on the machine itself. Also today: Databricks buys its way into spreadsheets, and India's Sarvam claims a document-AI record. Black Forest Labs released FLUX 3 Action, an open world-action model for robots that tops NVIDIA's RoboLab-120 leaderboard at 42.9% success with only seven billion parameters.

Modal and Baseten in talks at $15B and $26B, months after last rounds

The Arena

Modal and Baseten in talks at $15B and $26B, months after last rounds

Two inference platforms are asking investors to reprice the layer that sits between the model labs and everyone using them. The second story is Qualcomm's bid to put an always-on agent in your ears. Modal Labs is in talks to raise at a roughly $15 billion valuation, and Baseten is negotiating a round that could value it at $26 billion. Modal raised $355 million in May at a valuation of $4.65 billion, so the new number is roughly triple it — four months later.

See all