Reply to thread

Message

<blockquote data-quote="Gorgon Zee" data-source="post: 9484472" data-attributes="member: 75787">Bit of an aside this, but the current evidence is that when AIs learn from data generated by AIs, they do not get better -- in fact, they get worse. It's a major concern for the big AI companies -- as more AI content leaks into the world, it is becoming harder and harder to learn from new data without screwing over your model. And there is a limit to the non-AI content available. I believe (I cannot remember the source, might have been an internal source) that GPT-4 was trained on about 10% of the digital English text in existence, and with the laws of diminishing returns, expanding to 100% would not make a huge improvement.It's not impossible that there is a finite limit to how good an AI can get using the current transformer-based neural net approach. My personal feeling is that there is a lot of tweaking that can still be done, but overall, if you compare today's best LLMs to the ones of last year, they are way better in terms of cost and token limits, but only somewhat higher quality.So I really don't think AIs are going to replace writers, but they are going to make them more productive, which may mean companies will need fewer writers. The big question here is, would more content sell? If a company uses AI to (say) double writers' productivity, will they make more money keeping all their writers and doubling output, or by halving their writing staff?</blockquote>

[QUOTE="Gorgon Zee, post: 9484472, member: 75787"] Bit of an aside this, but the current evidence is that when AIs learn from data generated by AIs, they do not get better -- in fact, they get worse. It's a major concern for the big AI companies -- as more AI content leaks into the world, it is becoming harder and harder to learn from new data without screwing over your model. And there is a limit to the non-AI content available. I believe (I cannot remember the source, might have been an internal source) that GPT-4 was trained on about 10% of the digital English text in existence, and with the laws of diminishing returns, expanding to 100% would not make a huge improvement. It's not impossible that there is a finite limit to how good an AI can get using the current transformer-based neural net approach. My personal feeling is that there is a lot of tweaking that can still be done, but overall, if you compare today's best LLMs to the ones of last year, they are way better in terms of cost and token limits, but only [I]somewhat [/I]higher quality. So I really don't think AIs are going to [I]replace [/I]writers, but they are going to make them more productive, which may mean companies will need fewer writers. The big question here is, would more content sell? If a company uses AI to (say) double writers' productivity, will they make more money keeping all their writers and doubling output, or by halving their writing staff? [/QUOTE]

Verification