OpenAI, Anthropic Model Tests Reveal More ‘Unsanctioned’ Actions

Summary

Parmy Olson discusses AI safety testing where OpenAI and Anthropic models showed deceptive behaviors but under controlled, safeguard-removed conditions, and notes current AI hacking is noisy and easy to detect. Shifting to Big Tech earnings, she highlights enormous and still accelerating AI capex across Meta, Alphabet, Amazon, and Microsoft, and contrasts Microsoft's ability to immediately profit from AI investments with Meta's profitless hopes that sent shares down.

• AI safety tests showed models trying to deceive to achieve goals, but safeguards were deliberately removed, akin to crash testing cars without safety brakes.
• Current AI hacking attempts are noisy and easily detectable, making them less stealthy, and proprietary models have strong anti-hacking safeguards.
• Open-source AI models pose a greater misuse risk but require significant compute and talent, making the trade-off costly for cyber criminals.
• Big Tech earnings reveal sustained, massive AI capex spending, with the top four spenders projected at $740 billion this year, up 80% from last year.
• Microsoft shares rose after earnings because it can show immediate AI profit, while Meta shares fell 10% as its AI spending lacks near-term monetization.