Anthropic's automated researcher beats humans at $4 an hour

Two research stories landed today on the same theme: machines doing work that used to be a human's job. In one corner, an Anthropic fellow shows that an LLM-driven research loop can already out-design its human handlers on alignment post-training. In the other, Meta is testing robots that pull the cables inside its data centers.

Anthropic's Automated Alignment Researcher (AAR) is cheaper, faster, and — in a six-hour window — better than the human researchers it replaced. A paper from Anthropic Fellow Chen Yueh-Han describes a closed-loop system that scans the literature, proposes a method, trains a model against it for half an hour, scores the result, and iterates. Across 10 categories of alignment failure, the system found fixes that improved each benchmark without degrading capabilities, the paper says. "The best AAR method beats what experienced humans propose, on average within six hours," the team writes. "Human guided research directions do not lead to stronger performance." The cost line is the eye-opener: roughly $4 per hour in API inference, against $150 per hour the lab pays its human researchers — a 37× gap. The paper is honest about the limits — the system only improves what the benchmarks measure, and the benchmarks still need humans to maintain them — but the trajectory is the point. Once automated research is on a cost curve, "we have time to do more research" stops being a budget line and starts being a throughput number.

Meta is testing ABB-built robots to swap server cables and reset machines in its data centers, Wired reports, citing unnamed sources familiar with the effort. The push is meant to keep labor costs down as Meta's AI buildout balloons: a faster way to physically maintain the racks an AI fleet depends on, and one that scales with the footprint rather than the headcount. Workers' concerns are the obvious subtext — Bluesky commentary on the story noted that the very "AI factory" jobs the buildout was supposed to create are exactly the ones now being automated away. Meta's footprint is large enough that even a small robotics rollout becomes a meaningful test of whether hyperscalers can keep growing without proportionally growing the people who keep the lights on.

What to watch: whether the AAR pattern holds outside alignment and into capabilities research — and whether Meta publishes numbers on what fraction of data-center tasks robots actually replace, instead of just assists.

Do you think labs should publish automated-research internals the way they publish model weights — open by default? Tell us in the comments.

AI 101 — How do AI video generators work?

An AI video generator is a model that starts with a grid of random noise and repeatedly refines it into a moving picture that matches your prompt — the same trick image generators use, stretched across time. Nothing is filmed, and nothing is stitched together from clips the model has seen. The pixels in the finished video were invented, frame by frame, all at once. Why it matters right now Kling 4.0, Kuaishou's flagship video model, went fully live on October 15 with up to 30 seconds of vid

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Read more

AI 101 — How do AI video generators work?

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone