From 2020 until now, we've gone from crude blurry or clearly generative artefacts to being able to create full professional illustrations based upon textual prompts. That is huge. Classic generative art techniques look like cave paintings compared to what the latest image generation models put out (and I'm not talking about "AI slop" type stuff that DALL-E does).
Similarly, tools could fabricate podcasts years ago that sounded terrible. Now we have NotebookLM doing a "reasonable" job with two cliched-sounding "hosts". In a few years, will they potentially be able to create something akin to a professionally produced podcast given some smart prompting? The progress made so far points to yes, and I haven't seen any evidence so far to be pessimistic about it happening.
Can current techniques be scaled/improved/optimized to do this or do we need new techniques?
It took 30 years in the generative art world to move from cave paintings to the level that we have today because we needed new techniques.
For podcasts we are at the cave paintings level.
If we can get to professional level quality podcasts with the current techniques then we might only be a few years away.
I think it is more likely we will need new techniques which puts us potentially decades away.
If we look at LLMs the improvements over the last 18+ months since gpt4 was released have been minor despite incredible levels of investment.
> "being able to create full professional illustrations based upon textual prompts. That is huge. .. (and I'm not talking about "AI slop" type stuff that DALL-E does)"
Then what are you talking about, where can I get it?