DON’T TRUST ChatGPT’s CODE! [youtube / 18 min] https://www.youtube.com/watch?v=YiokTYzA6BI
"He explores the code that ChatGPT can write, and gets ChatGPT to practice Test Driven Development (TDD)."
That's currently. The problem I suspect is the form of the input. With AlphaGo/Zero/MuZero, they may be thinking 'fast' but is indistinguishable or better than humans thinking 'slow'.
With AlphaGo/etc there is an internal representation/language being created directly from the board state.
In the case of ChatGPT we are using natural (or programming) language which is once removed from the 'thing' that's discussed. A good illustration of this would be if the AlphaGo folks were trying to beat Atari 2600 games from English descriptions of what's happening onscreen, written out in normal style as if being explained to a person. The machine of course can get this information at high speed so the lack of bandwidth is not the problem. We wouldn't be surprised it doesn't do as well as getting a direct view without human involvement in observation.
As an example AlphaGo was trained on high-level human games and self-play, AlphaZero only learned from self-play. The latter learns faster and plays better, without the human coloring of the data.
Btw, last I heard/read there are still some Atari 2600 games that the machine still can't beat.
If AI may not be reasoning well by self - from choosing the 'right' base points to building only on it (to do this 'right' it would need to have its own opinions and know the difference beetwen light and darkness - how?).. and to get 100% reasoning ALL not reasonable paths shall be replaced.. - but it would be the nicest if the reasoning methods and paths could be extendable and customisable (POVs or emulating differrent ways and coherent) and some of them could be objectivly choosen as canonical in given circumstances with consensus - then, we could let it grow.
There must be a reasoning for some reasoning.
If AI may not be reasoning well by self - from choosing the 'right' base points to building only on it (to do this 'right' it would need to have its own opinions and know the difference beetwen light and darkness, taste ('style transfer') - how?) - and to get 100% reasoning ALL not reasonable paths shall be replaced or explained.. ('human labeling .. ranking.. options..') 'reinforcement'..
.. it would be the nicest if the reasoning methods and paths could be just extendable and customisable (POVs or emulating differrent ways, coherent/emergent..) but some of them objectively choosen as canonical, in given circumstances, with consensus ('fine-tuning .. which one seems the best so ranking'), - _then_ we could let it grow ('generate .. output').. - otherwise it will become a disconnected from truth alternate hallucinated stretched reality which will eat itself in the best case.