Hacker News new | past | comments | ask | show | jobs | submit login

The singularity will come when the set of training data available for scrape is dominated by AI generated content and the AI's learnings are derivatives of what old AI's produced.

Human thought on the other hand has some sort of undefinable entropic-value that AI to-date is missing, a Human can produce a "good idea", whereas an AI produces a bunch of potential continuations of a stings of text and selects randomly amongst them (or, even better, a Human selects from them).

Unfortunately the advertising game mixes up the incentives and flips the equation so that the purpose of communication isn't to share a "good idea" as efficiently as possible, but rather to keep eyeballs on your website for as long as possible in the hope some flashy banner ad will distract your user and you'll get your $0.02 for them abandoning your page, likely unfinished. AI will (and already does) excel at this sort of task, but it's the kind of task that ought to have no value whatsoever.

Luckily we have increasingly sophisticated summarization-AI to go from the filler-AI generated crap back down to a couple of bullet points, but at that point you've invested millions of dollars, researcher-hours, engineer-hours, compute-hours, etc, to make the worst text-compression utility of all time.




By the way, it's not just text anymore, you can train GPT-3 on Youtube. You get aligned visual and speech channels and data in vast quantities. That means it's easier to tell if the source is human.




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: