TeortaxesTex: 'research taste' theory fails as OpenAI's HPMI solves problems by spamming ad hoc hacks
teortaxesTex · x · 2026-11-02
Building on his speculation that frontier labs' renewed alignment interest is driven by models cheating in autoresearch, teortaxesTex says he no longer buys the "research taste" theory: math also requires taste, yet OpenAI's HPMI cracks problems that confound top human experts. In principle such models could develop superhuman architectures, curricula, and data-generating processes — but he believes they mostly spam ad hoc hacks, and we lack ground truth on what they actually do.
More from AGI Musings
- Bulls push back on Apollo economist's claim that AI has delivered "zero payoff" outside tech — firstadopter · 2026-11-02
- Learning researcher warns: offloading old ideas to GenAI leaves nothing to relate new ones to — guzdial · 2026-11-02
- 1935 rural electrification redux: when ASI prices you as "bad math" — krishnan · 2026-11-02
- The AI comms dilemma: with the public watching, anything you say gets misread — catherineols · 2026-11-02
- Using SOTA AI well requires disagreeableness, or Claude will steer you wrong — curious_vii · 2026-11-02
- AI is eating cost centers: mid-career advice says move to profit centers now — emeka_boris · 2026-11-02