Lecture, Statement, Other
Being a pain in the ass to AI companies is a real job now | Rob Wiblin
- Anthropic's proposed federal legislation would significantly restrict government ability to halt AI training, requiring Commerce Secretary personal approval, a specific court case proving imminent catastrophic harm, and issuing orders that expire in 45 days without renewal; the bill would preempt state safety regulations in California, Illinois, and New York, leaving AI companies to self-regulate on dangerous actions.
- Anthropic's CEO estimates a 25% probability of AI causing significant harm, while the company faces potential hostility and political concessions from the Trump Administration, prompting a strategic view that retaining regulatory power at the state level or accepting self-regulation are the only viable paths.
- OpenAI recently fired three safety researchers over a suspected leak, a decision characterized as likely mistaken and pretextual, which risks eroding trust and deterring candid safety dissent; Google DeepMind has similarly introduced lifetime non-disparagement clauses in employment contracts that prohibit public criticism.
- The prospect of automating most staff roles in the coming years threatens to eliminate employee leverage, though past advocacy campaigns like "The Not For Private Gain" and "Right to Warn" have successfully forced policy changes and influenced state legislation like California's SB53.
- Emerging external pressures include investigations into AI swarms by the Nightingale Collective and hypocrisy disclosures by the Midas Project, with 400 OpenAI staff previously petitioning to slow AI development and UK Parliament members signing letters criticizing DeepMind.
- Public trust and regulatory oversight are expected to become increasingly salient over the next few years, with former AI employees possessing unique credibility to generate earned media and exert influence on corporate behavior.
- Risks associated with concentrated control involve the US military and intelligence services possessing the power to halt models, creating a combination described as potent, threatening, and open to abuse, alongside the risk of reckless companies pursuing recursive self-improvement without sufficient checks.
- Future impact strategies emphasize collective coordination, public speaking within company channels, and leveraging the "extremely hostile" environment to force changes, as individual marginal employees may achieve greater positive impact at other organizations or through external advocacy.