🎙️ EP 116: Just 250 Docs Can Hack a 13B AI Model?! & Google Shoe Try-Ons

What if I told you that a few hundred poisoned documents could break models as big as GPT-4 or Claude? 😵 Anthropic just proved it. Their new paper shows that just 250 samples can secretly backdoor any LLM, no matter the size. In today’s episode, we unpack this wild discovery, why it changes AI security forever, and what it means for the future of open-web training.

We’ll talk about:

• How Anthropic’s team used 250 poisoned docs to make 13B-parameter models output gibberish on command
• Why bigger models don’t mean safer models and why scale can’t protect against poison
• The rise of TOUCAN, the open dataset from MIT-IBM that’s changing how AI agents learn real-world tools
• The new AI race: from Jony Ive’s “anti-iPhone” with OpenAI to Amazon’s Quick Suite for busi ...