RLVR logic: with human data exhausted, security and math become the next verifiable training grounds
felpix_ · x · 2026-10-28
- The thread traces the training playbook: after 2024's reasoning wave, humanity's data was largely consumed by early 2025; RLVR proved itself with o1, and by mid-2025 models conquered the IMO and arguably 90% of SWE/PM tasks.
- The hardest, most time-consuming AND verifiable domains left: security work and unresolved math problems. Public models improved rapidly there through late 2025/early 2026, threatening incumbents in both fields.
- Expected payoffs of going further: (1) stronger models, (2) contributions to human knowledge, (3) fundraising hype — which the author argues 'unironically explains' the OpenAI vs. mathematicians conflict.
More from AGI Musings
- Nathan Young Asks Why AI Cyber-Risk Demos Are Obviously Unacceptable — NathanpmYoung · 2026-10-28
- Bill Gates predicts many people will reject AI and choose to 'live like the Amish' — Polymarket · 2026-10-28
- Neuroscientist Anil Seth checks in from Asilomar 2026, the storied AI safety venue — prof_kamilov · 2026-10-28
- Report breakdown: Claude sent zero emails in "one-person biotech," founder sent 12 after edits — faraz0x · 2026-10-28
- Silicon Valley influencer lists 39 Agent-era startup ideas: stop building software — FinanceYF5 · 2026-10-28
- When AI agents can do everything, picking what's worth doing is the hard part — bilawalsidhu · 2026-10-28