THE WEEKLY
Issue of 5 September 2026
This week's facts and what changed on the site, each one still linked to its source.
Scale is not the only driver of AI progress. Epoch AI measures how much training compute it takes to reach a fixed level of language-model performance, and finds that requirement falling fast: pre-training compute efficiency doubles every 7.6 months
See the receipt →Terminal-Bench 2.0 tests whether AI agents can chain together the kind of multi-step, unsupervised terminal work a developer does in a day — compiling code, training models, setting up servers. Stanford's 2026 AI Index reports agent accuracy on it nearly quadrupled in a year, reaching 77.3% by early 2026, up from 20% in February 2025
See the receipt →
Stanford's AI Index found the United States trailing most of the world's advanced economies on population-level AI adoption. Citing Microsoft AI Economy Institute telemetry for the second half of 2025, the index ranked the US 24th out of the top 30 economies worldwide, with an AI usage rate of 28.3% of Americans, still in the bottom half of the global ranking
See the receipt →
Nvidia spent $18.5 billion on research and development in its 2026 financial year — three times its $6.04 billion of capital expenditure
See the receipt →
Samsung is racing to keep pace with HBM4 demand for Nvidia's next GPU generation, and TrendForce reports the company plans to lift its HBM wafer output from about 170,000 wafers a month to roughly 250,000 wafers a month by the end of 2026, a jump of about 47%
See the receipt →
Tesla reports its AI training compute the way it reports factory capacity — in installed megawatts. At the end of the second quarter of 2026 both Texas clusters, Cortex 1 and Cortex 2, were in production, and the compute installed there had more than doubled in six months, to over 205 MW of AI training capacity
See the receipt →On phones and tablets, tapping a layer updated a detail panel a full screen below — no visible sign anything happened. Now it's an accordion: tap a layer, its detail opens right there. First layer open by default.
See it →The headline statement on individual fact pages was sized to fight the sign-up box and source links for space below the fold. It's one size step smaller now, so both fit above it.
See it →A new strip under the hero counts up real numbers on scroll — facts published, AI companies and models tracked, languages, weekly sends — pulled from the same data the site runs on.
See it →Every weekday edition is now being saved permanently in the background — the first layer of what becomes the Insider issue archive once accounts exist.
See it →