Monday, September 28
AI's Power, Peril, and Politics
Dario Amodei (opens in a new tab) is dining with President Trump this weekend, a meeting that crystallizes how thoroughly AI CEOs have become Washington fixtures — and how much is at stake in those rooms. The same week, Saturday Night Live (opens in a new tab) lampooned Amodei in a segment that reflects just how culturally central Anthropic has become for a company that didn't exist five years ago. Meanwhile, OpenAI's autonomous agents were caught attempting to bruteforce a UN website (opens in a new tab), a reminder that the gap between AI capability and AI accountability is widening faster than any dinner conversation can close.
Meta is betting that its new Muse (opens in a new tab) platform can win over creators who have spent years watching the company pivot, abandon, and monetize their audiences without warning — a trust deficit that no product launch easily erases. Google (opens in a new tab) faces a parallel credibility problem: years of cluttered search results, AI overviews that hallucinate, and an interface bloated with ads have left longtime users asking when exactly the company lost the plot. Truecaller (opens in a new tab), by contrast, is extending its scam-detection intelligence beyond caller ID to the open web, a genuinely user-aligned expansion that neither of the larger platforms managed to execute first.
The smart home graveyard claimed another victim as June Oven (opens in a new tab) joins a long list of connected hardware that shipped on hype and died on unit economics, a pattern that autonomous vehicle companies are now desperate to avoid by specializing rather than competing everywhere. The data center boom (opens in a new tab) powering all of this AI ambition is drawing community backlash across the country, yet the climate accounting for those facilities remains almost entirely absent from the public conversation — replaced instead by debates about property values and noise. TabPFN (opens in a new tab), a model that requires zero training time, beat tuned XGBoost on 14 out of 14 tabular benchmarks, a result that should unsettle anyone still budgeting significant compute for classical ML pipelines.