> If the goal is achieved, which is highly unlikely, then we get very very close to AGI and all bets are off.
Question on two fronts:
1. Why do you think, considering the current rate of progress think it is very unlikely that LLM output becomes indistinguishable from expert creatives? Especially considering a lot of tells people claim to see are easily alleviated by prompting.
2. Why do you think a model whose output reaches that goal would rise in any way to what we’d consider AGI?
Personally, I feel the opposite. The output is likely to reach that level in the coming years, yet AGI is still far away from being reached once that has happened.
1. The progress is there but it's been slowing down yet the downsides have largely remained.
1.1. With the LLMs, while thanks to the larger context window (mostly achieved via hardware, not software), the models can keep track of the longer conversations better, the hallucinations are as bad as ever; I use them eagerly yet I haven't felt any significant improvements to the outputs in a long time. Anecdotally, a couple days ago I decided to try my luck and vibe-code a primitive messaging library and it led me in the wrong path even though I was challenging it along the way; it was so convincing that I wouldn't have noticed hadn't my colleague told me there was a better way. Granted, the colleague is extremely smart, but LLM should have told me what was the right approach because I was specifically questioning it.
1.2. The image generation has also barely improved. The biggest improvement during the past year has been with 4o, which can be largely attributed to move from diffusion to autoregression but it's far from perfect and still suffers from hallucinations even more than LLMs.
1.3. I don't think video models are even worth discussing because you just can't get a decent video if you can't get a decent still in the first place.
2. That's speculation, of course. Let me explain my thought process. A truly expert level AI should be able to avoid mistakes and create novel writings or research just by the human asking it to do it. In order to validate the research, it can also invent the experiments that need to be done by humans. But if it can do all this, then it could/should find the way to build a better AI, which after an iteration or two should lead to AGI. So, it's basically a genius that, upon human request, can break itself out of the confines.
Question on two fronts:
1. Why do you think, considering the current rate of progress think it is very unlikely that LLM output becomes indistinguishable from expert creatives? Especially considering a lot of tells people claim to see are easily alleviated by prompting.
2. Why do you think a model whose output reaches that goal would rise in any way to what we’d consider AGI?
Personally, I feel the opposite. The output is likely to reach that level in the coming years, yet AGI is still far away from being reached once that has happened.