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:
- Seed-loss check: fall back to Lloyd when e-means ends worse than its seeds — fixes the revival counterexample and preserves k-means++'s O(log k) guarantee on every input.
- Pseudo-count floor: a one-step bound predicts τ = 0.1927; failure rate drops from 49% to 0% across 20,000 runs at that threshold.
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)→
More from Fun
- eigenrobot turns 'millennials won't buy brown furniture' news into a bleak AI joke — eigenrobot · 2026-11-20
- A WebGPU pinball table regenerates its whole art pipeline from one JSON swap — AI_Andrew · 2026-11-20
- Anthropic declares on IPO day that Claude could one day cure all allergies — Polymarket · 2026-11-20
- Classic movie line "machines cannot think for themselves" becomes AI-era irony meme — arpitingle · 2026-11-20
- Redditor makes 45 models draw the same car, agents self-correct via screenshots — qu1dow · 2026-11-20
- Living like a GTA character without buying GTA 6: gym, relationship scores, staying low-key — zeroxjackson · 2026-11-20