Rather than merely predicting ‘what word most often comes next’, conjoined alternate “world views” are surmised to offer ‘the next levels’ in machine learning, making our tools more anthropomorphic and adept at ‘real-world’ use. Imagine that Japanese- and Kanji-trained models that have also been trained with ‘Samurai moves’ and culturally-specific videos would be ‘tuned’ to better achieve certain results with magnitudes greater efficiencies than LLMs that merely ‘talk’ and ‘reason’ about the tasks, results. Google’s old amassing of textual and also visual content suddenly makes more perfect sense.
It’s very clear to me that he is right about the problem. If his solution will work remains to be seen.
Learn also by what I do, not just by what I say
The self-driving is like machine chess-player
