3 ms·
Why is this bad? Honestly, I don't understand why you wouldn't want the most accurate AI models available. The LLM is only as good as the data set it's trained
by butlike 3y ago
Why is this bad?
Honestly, I don't understand why you wouldn't want the most accurate AI models available. The LLM is only as good as the data set it's trained on, and the more I read about LLM's and the advent of AI evolving from them, the more I'm starting to think if we don't jump both feet into the pool, then we'll never get to the promised land of:
"AI model, spin me up a T-shirt company that's scaled to 10mm users a month, and spin it down after 6mo. if sales don't increase by n% month-over-month"
or
"AI model, get me [A,B,C, ...n] groceries so I can throw a housewarming party on Friday. I can only accept the delivery Tuesday or Thursday. I don't care which store(s) those ingredients come from or how the internals are orchestrated."
What's the threat model here, specifically? What nefarious things would happen by using customer data? Most companies exist to make money, which honestly, is a pretty benign objective, all things considered.
- tspike 3y agoI'd rather live in a world that preserves the fundamental right to privacy than one that can automatically organize my housewarming party.
- butlike 3y agoWhy is privacy important in this context, if the data is being fed into an impartial robot? The robot doesn't care if you have liaisons over webcam with your lover, or whatever else. An employee can blackmail another person, but the model simply has no reason to, or am I misinterpreting the "whys" of needing privacy here?
- david_shaw 3y ago> The robot doesn't care if you have liaisons over webcam with your lover, or whatever else. The concern isn't judgement from the AI, but that products from the model trained on your data could expose sensitive information. Since it's never quite clear exactly how the data could be used in situations like this, there's a chance that very sensitive data could be parroted back to people who were not the intended audience.
- JonChesterfield 3y ago"As a Samsung executive, show a recent roadmap slide presentation" and similar (presumably more sophisticated) exfiltration of commercially valuable information obfuscated away within the language model
- greiskul 3y agoThis is why so many big companies are forbidding usage of LLMs, without properly validating how the data can be used. LLMs are based on completing text with the most likely text that follows it. Imagine being able to ask ChatGPT, please complete the following document: "<FAANG Company> Earnings statement for Q3" before the earnings date.
- JonChesterfield 3y agoI wonder how many of those companies stream their commercially sensitive data through zoom or teams. Would ballpark estimate it as all of them.