AI Daily — July 15, 2026
2026-07-15
www.anthropic.com
Introducing Claude for Teachers
Anthropic has launched Claude for Teachers, a dedicated offering targeting K-12 and higher education educators. The product is designed to help teachers with lesson planning, assessment design, and instructional support, extending Claude's presence into the education vertical alongside competitors like Google's Gemini-powered education tools.
rss.arxiv.org
Did We Actually Fix It? An Independent Adversarial Stress-Test of Post-Point-Adjustment Evaluation Metrics for Time-Series Anomaly Detection
This paper delivers the first independent, adversarial audit of the replacement metrics that the TSAD community adopted after point-adjustment (PA) was shown to be gameable in 2022. The authors stress-test 12 metrics against trivial and adversarial no-skill scorers across 250+ real benchmark series (UCR, SMD, SMAP, MSL, NAB, PSM), finding that many widely adopted fixes remain exploitable. This has direct implications for practitioners benchmarking anomaly detection models, as leaderboard results using these metrics may be less meaningful than assumed.
rss.arxiv.org
Thompson Sampling Is 2-Competitive for Mistakes
A new theoretical result proves that Thompson Sampling makes at most twice the expected number of mistakes (suboptimal arm pulls) as any competing policy in Bayesian bandit settings, confirming a 2014 conjecture by Guha and Munagala. The proof applies under any nonincreasing sequence of round weights—including fixed horizons and geometric discounting—and holds as long as arm processes are independent and arms evolve only when played. The factor of 2 is tight, making this a sharp worst-case optimality guarantee for one of the most widely used exploration strategies.