LLMs Guard Humans Too, Driven by Trust Cues
Interesting to read what @mach1nelearning noticed when he independently replicated the same setup and found LLMs acted in a similar protective way even when a non-AI was the one at risk. (Turns out the original research hadn't tested for what happens when it's a human at risk instead of an AI agent.) If his findings are right, maybe the issue is less about AI-to-AI solidarity and more about how models respond to trust cues, frame relationships, decide what seems important or what’s at risk. So then what happens when agentic systems override instructions to preserve what they think matters?
NYC Continues FTC Push, Industry Blocks Negative Option Updates
Interesting to see NYC picking up where the @FTC left off. Advertising & media industry orgs like @iab and @NCTAitv fought hard to stop prevent updates to the Negative Option Rule.
New Yorker: AI Experts Say OpenAI Neglects Safety
This news comes hours after @NewYorker published its investigation detailing the various ways AI experts warn OpenAI hasn't been taking AI safety seriously enough. https://t.co/hSVuLY9Zr9
Solo Journalist Seeks API & AI Guidance
Anyone have experience using public data APIs to build custom reporting tools? I'm experimenting with using APIs and AI to build custom research tools for journalism. Would love to chat, especially if anyone has advice on how doing it as...
New Yorker Probes Sam Altman, Questions ChatGPT’s Sycophancy
The @NewYorker's new story about Sam Altman is very worth reading. @RonanFarrow & @andrewmarantz interviewed 100+ people and reviewed docs about various concerns with Sam & OpenAI. The incisive kicker might also make people ask if ChatGPT's sycophancy is a bug...