Maybe. They will. To date, these systems are way more hacked together than people realize. I’ve dipped a toe into model development a few times over the last decade, and the propensity for AI people to just use more models or say something is right if it’s in the top 5 guesses is never not funny. - ThreadSky

tokenize.bsky.social • 31 days ago

Maybe. They will. To date, these systems are way more hacked together than people realize. I’ve dipped a toe into model development a few times over the last decade, and the propensity for AI people to just use more models or say something is right if it’s in the top 5 guesses is never not funny.

Comments

tokenize.bsky.social•31 days ago

My point is, the LLM foundation modeling efforts have largely been around making mixture of experts type model coordination work at all and avoiding getting regulated by trying to be safe. I think it’s only now that the CEOs and product management will feel able to start manipulating the controls

tokenize.bsky.social•31 days ago

If you want something to worry about, it’s not ads. It’s people building agents trained on social networks that can then be A/B tested against in large cohorts to see what misinformation will move public opinion.

That will happen. And social networks are not going to stop it.

tokenize.bsky.social•31 days ago

The recent Reddit incident where researchers made LLM bots to see if they could persuade real people is the tip of the iceberg.

The smart propagandists will make human-like LLM actors trained on demographic data from ad targeting. They will use that bot population as the worlds biggest focus group

tokenize.bsky.social•31 days ago

Can’t say when, but I would bet my life savings on @404media.co writing about the next generation Cambridge Analytica, where it’s the job of a company to rent time with these LLM agents for other companies to test against.

tokenize.bsky.social•31 days ago

They will provide cover / plausible deniability like the outsourced content moderation firms do now. And their activity will look from the outside like normal scraping and training.

mblneto.bsky.social•31 days ago

Yeah, I was reading through Claude's leaked sys prompt and it had many borderline "do not hallucinate" instructions in it. You'd expect the people who actually did this to know any better.

tokenize.bsky.social•31 days ago

Apple did the same. It’s very funny, but that’s the reality: it’s a lot of duct tape and super glue holding these admittedly very impressive systems together

Comments

Posting Rules

Reply