MIT Technology Review’s The Download on AI capability claims and nuclear power

MIT Technology Review’s The Download on AI capability claims and nuclear power

MIT Technology Review published this edition of The Download on February 5, 2026. Grace Huckins says an influential METR graph suggests certain AI capabilities are developing exponentially, and that newer model releases have outperformed the trend. She cautions that the graph’s dramatic reception oversimplifies the situation: “the truth is more complicated.” The newsletter points readers to a full story, but does not explain the complications there. Casey Crownhart takes up audience questions about advanced nuclear power after an online discussion of next-generation nuclear power, hyperscale AI data centers, and the grid. She says some questions went unanswered or were not covered in the depth she wanted, and that she is keeping some in mind for future reporting. The newsletter does not include the answers, so what those questions reveal about nuclear power remains unknown. In a separate item, Eileen Guo reports research finding images containing personally identifiable information in a small audited portion of DataComp CommonPool. The researchers estimate that the full dataset contains many more such images. The newsletter’s bottom line is that anything put online “can be and probably has been scraped.” The estimate is based on an audit of only a small subset, so its scale remains an estimate rather than a direct count of the full dataset.

Publication: 077c8c6d-7e40-c0eb-9ff8-5f2cd2a5cd99

attribution — pass: This is a single newsletter edition from MIT Technology Review. Its blurbs are written in the authors' own words. The METR item is signed by Grace Huckins, the nuclear item by Casey Crownhart, and the CommonPool item by Eileen Guo. The reading attributes each item to the right writer.

This is today's edition of The Download, our weekday newsletter

summary — uncertain: The lead item is hard to follow for a non-specialist. The reading calls it 'an influential METR graph' but never says that METR is an AI research nonprofit (Model Evaluation & Threat Research). It also never says what the graph measures: the length of tasks, timed by how long a human would take, that a model can complete independently. It drops the source's concrete example, Claude Opus 4.5 at about five hours. No definitions are supplied. Smallest repair: briefly identify METR and add the five-hour task example.

Opus 4.5 appeared to be capable of independently completing a task that would have taken a human about five hours

date — pass: The reading says February 5, 2026, which is the UTC calendar day of originalPublished (2026-02-05T13:10Z). The evidence kind is publisher-page, so no archive disclosure is needed. The modification timestamp precedes publication. Some Deep Dive sidebar entries may be later additions to the page, but the reading neither uses nor relies on them. There is no hindsight framing.

support — pass: Every claim matches the source text. Huckins notes the trend and that recent models beat it, then cautions that the truth is more complicated than the dramatic responses. Crownhart says questions went unanswered or were not covered in the depth she wanted. Researchers audited 0.1% of CommonPool and estimated hundreds of millions of images with personally identifiable information. The reading keeps that figure framed as an estimate and adds no unsupported benefits or costs.

But the truth is more complicated than those dramatic responses would suggest.

media — pass: The reading selects no media, and no verified media candidates were supplied, so nothing needs checking.