zlacker

"As of 2025, after more than 80 years of research on Artificial Neural Networks, there remains no evidence of their general intelligence. Even after priming the pump with nearly every book ever written, all available scientific articles, and all publicly available text on the internet (and consuming more power and resources than entire countries), they routinely fail at basic tasks and generate nonsensical output and 'slop' that has already degraded the internet as a source of training data for future systems. The long slow failure of ANNs is a strong indictment against the statistical/neuromorphic approach to AI."

>>musica+(OP)
Exactly.

On top of it, Cyc's whole 40-year run expense of "200 million dollars, 2000 person-years" is... What LLM companies waste each month, for slightly less hallucinations and getting the right number of fingers most of the time?

>>musica+(OP)
There's plenty evidence of the general intelligence of LLMs. And a site that has climbed to the 6th most visited in 2 years is clearly not failing to be generally useful.

I do think the author was more scathing than necessary here though

>>musica+(OP)
LLMs are clearly useful, even if flawed; conversely, the author had to struggle to find uses of Cyc. Additionally, neural networks were already showing promise before the scale-ups; AlexNet was trained on consumer hardware and outperformed all other solutions to date.

I see your point, and I agree that the way that paragraph was initially posed was perhaps more dismissive than it should have been... but to disprove it, you need proof that Cyc was a working approach for anything, and said proof appears to be lacking.