e-means' two counterexamples fixed in 10 minutes with Astra 6.2; high schooler solves 6 conjectures

amuellerml · x · 2026-11-20

Rohan Arun reports using the new Astra 6.2 model to find improvements to e-means (mini-batch k-means via exponential moving averages) within 10 minutes, repairing two counterexamples published just the day before:

The combination, e-means+, matches 25 Lloyd iterations (0.9998× loss) on SIFT1M in 4 data passes instead of 26.

Also cited: high schooler @1337isnot1337, a top contributor to the Integer Mult repo, solved 6 conjectures with Astra 6.2 since yesterday, including sharp asymptotics for Erdős 607 line-size sets. A quoted reply quips that OpenAI trained a model great at math but useless specifically for "AI influencers."

Related event: AI Model Refutes Conjectures in Clustering Algorithm Within Minutes(2 posts)→

Original post →

More from Fun

Fun channel →