zlacker

>>kaeruc+(OP)
Goodhart's law: if you rely on a social signal to tell you what's good, you'll break that signal.

Very soon, the domain of bullshit will extend to actual text. We'll be able to buy HN comments by the thousand -- expertly wordsmithed, lucid AI comments -- and you can get them to say "this GitHub repo is the best", or "this startup is the real deal". Won't that be fun?

>>perihe+ca
We'll be back to the 1990's "software agents" craze take two: Needing AI driven agents that seek out and index and evaluate content on our behalf, and seek to negotiate with each other for recommendations with currency being trust based on how "your" agent evaluated prior results.

I'm hoping to put an AI between me and my e-mail inbox this weekend (I had ChatGPT write most of the code; it's not much); not fully automated, but evaluating and summarising and categorising. I might extend that to e.g. give me an "algorithm" for my Mastodon timeline (despite all of the people insisting on reverse chronological, I'm at a few hundred people I follow and already can't keep up), and a number of other sites I visit. For most of these things latency does not matter, so e.g. putting them through llama.cpp rather than something faster is fine, and precision isn't critical (I won't trust it to automatically reply or automatically reject anything, but prioritisation an categorisation where missteps won't have any critical impact.