Vitalik Buterin pfp
Vitalik Buterin
@vitalik.eth
My impression of chatgpt so far is that it's good at replicating existing human behavior, but bad at doing anything out-of-sample. It anchors quickly onto patterns that have been done by many people before, but there's definitely something that we call "original thinking" that it's not doing at all.
16 replies
0 recast
3 reactions

Vitalik Buterin pfp
Vitalik Buterin
@vitalik.eth
As a random example, consider these AI-generated debates (credit to pmarca for the format). Notice that the Chinese perspective is not even consistent across the two debates! It's pattern matching to the usual yes/no debate, and... https://i.imgur.com/OebQ19c.jpg
3 replies
0 recast
1 reaction

Venkatesh Rao ☀️ pfp
Venkatesh Rao ☀️
@vgr
This is to be expected given how transformers work I think. I was not expecting creativity except by accident.
1 reply
0 recast
0 reaction

Vitalik Buterin pfp
Vitalik Buterin
@vitalik.eth
There definitely is the "scale is all you need" school of thought that says that if you just scale up the param count by another 1000x or 1000000x then insight and creativity will appear. Seems false so far, but... I guess we'll see!
1 reply
0 recast
0 reaction

Venkatesh Rao ☀️ pfp
Venkatesh Rao ☀️
@vgr
I’ve been arguing embodiment is necessary for this reason. It’s how human brains are constantly immersed in sensations for 90% of which there are no words. LLMs don’t even have true image data really. The latent diffusion models use tagged image training data so kinda stay within verbal worlds.
3 replies
0 recast
0 reaction

Stefan Lesser pfp
Stefan Lesser
@stefanlesser
I’ve been on this 4E Cognition trip for a while and found the Relevance Realization framework quite a good perspective on why scaling isn’t going to magically solve this, but I can’t tell if it’s not well known or if it’s been refuted somewhere I didn’t look…? http://www.ipsi.utoronto.ca/sdis/Relevance-Pu
1 reply
0 recast
0 reaction

Stefan Lesser pfp
Stefan Lesser
@stefanlesser
But even if Relevance Realization turns out to be the wrong model, just reading Mark Johnson and George Lakoff seems to also suggest that it’ll need more than just more processing power. I’m no expert at all, but it seems quite obvious.
1 reply
0 recast
0 reaction