The real story isn't Harvey's $15.5B valuation—it's that legal AI still can't reliably trust its own outputs. That's what "Learning When to Trust via Selective Context Preference Optimization" actually solves: LLMs hallucinating with confidence is the blocker for enterprise AI, not compute. Meanwhile, the video language model paper exposing how easily they fail at basic event tracking (who picked up what, when?) is a brutal reminder that scaling parameters doesn't equal reasoning—and this matters for every "AI camera" startup getting funded on surveillance promises. The real hardware trend isn't funding hype; it's that operational AI (cheap inference via llama.cpp, certified evaluation methods cutting testing costs 74x) is quietly outpacing the models themselves, which tells you where actual competitive advantage lives.