Sections
00:00 What we're covering today
00:35 1. Anthropic says Claude now leads 26% of its AI research
01:44 2. Reuters reports an earlier Huawei chip target as supply runs short
02:44 3. Australia asks how to keep data-center costs off community bills
03:42 4. A new radiology model tackles 146 abdominal scan findings
04:42 5. Cornell's classroom pilots can use less AI, too
05:38 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Friday, September eighteenth. Anthropic says Claude led a quarter of its A I research in August, with people still supervising. We'll look at what the company counted, and what its numbers can't tell us.

Then, Huawei has a faster chip target but can't make enough equipment to meet demand. Australia is asking who should pay for new data centers. We'll also hear what a scan-reading model has shown in tests, and why Cornell's classroom trials may call for less A I.

Anthropic says Claude now leads 26% of its A I research. Anthropic released a new way to measure research automation on September 17. By its count, Claude led 26% of its A I research and development work in August, up from less than 1% in February. For Anthropic, leading means doing most of a task from a broad instruction while a person supervises. The company says none of the work it measured was fully autonomous. More than 90% involved at least substantial collaboration with A I. Anthropic measured its own work. It did not independently test how quickly A I improves itself. These are company measurements, not an independent audit. Anthropic uses its own models in evaluating its systems. It acknowledges that a model judging another model can share the same errors. Different labs also lack a common measurement method. Anthropic plans to let outside evaluators see its internal systems and data. Their findings could make the index more useful if Anthropic measures future work the same way. Those evaluators still need to check the ratings and the research behind them.

Reuters reports an earlier Huawei chip target as supply runs short. Huawei announced new A I computing systems at its Shanghai conference on September 17. Reuters reports that Huawei now expects its Ascend 960DT chip in the first quarter of 2027, three quarters earlier than planned. Huawei also introduced Peerium, a design meant to connect as many as 1 million processors. The company says it's deploying an Atlas 950 cluster with 256,000 cards, while its Atlas 960 system is still in testing. But Huawei's rotating chairman, Eric Xu, told Reuters the company can't make enough equipment to meet demand in China. That shortfall limits sales overseas. The size of a planned system tells customers little about whether they can get it, power it, or run their workloads efficiently. Customer deployments could show whether Huawei can deliver these systems, and independent workload tests could show how well they run. Delivery volumes matter while supply is this tight.

Australia asks how to keep data center costs off community bills. Australia announced a public consultation on A I training and large data centers on September 17. It opened at 6 a.m. AEST on September 18 and closes at 5 p.m. AEDT on October 9. The paper considers renewable electricity obligations, water use, community engagement, and requirements intended to prevent net costs to consumers and communities. It also asks which facilities should face the requirements. Australia has put these requirements out for consultation. They are not final obligations. The government says they should work alongside state and territory approvals. Communities cannot know how much protection they will get until authorities set the thresholds and enforcement rules. The final standards need to say which costs developers pay and how authorities check that they comply. Any renewable certificate requirement also needs to account for when a facility uses electricity.

A new radiology model tackles 146 abdominal scan findings. A study published in Science on September 17 introduces RADAR, an A I model for reading abdominal CT scans. In the researchers' evaluation, it had a mean AUC of 0.913 across 146 findings. The best competing vision-language model scored 0.776. AUC measures how well a model separates positive and negative cases; it isn't the percentage of patients it diagnoses correctly. The study also reports an AUC of 0.895 across cohorts from eight outside centers. Those results come from the researchers' evaluation. They don't show that routine use improves patient outcomes or that RADAR can replace a radiologist. Others can inspect the published code, but that isn't independent clinical validation. Independent evaluations and studies in working hospitals could show where RADAR helps. They need to count missed findings and false alarms, and test whether clinicians make better decisions.

Cornell's classroom pilots can use less A I, too. Cornell announced A I Integration Pilots and an expansion of its A I Undergraduate Fellows program on September 17. A $2 million Dake family gift supports the initiatives, alongside support from the provost's office. Faculty teams can propose more A I or less A I, depending on what students need to learn. Changes may begin in spring 2027, fall 2027, or spring 2028. Students will help research and evaluate classroom use. Cornell has announced the classroom trials, but it has not shown that students learn more. Their assignments might look better even if they can no longer explain the work without help. Teams plan to share results in spring or fall 2028. Tests of what students remember and can solve on their own could show whether they learned more or just finished assignments faster.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.