This is orthogonal to the topic at hand. How does the chemistry of biological synapses alone result in a different type of learned model that therefore requires different types of legal treatment?

The overarching (and relevant) similarity between biological and artificial nets is the concept of connectionist distributed representations, and the projection of data onto lower dimensional manifolds. Whether the network achieves its final connectome through backpropagation or a more biologically plausible method is beside the point.

That is a used method for training other LLMs, I believe OpenAssistant did this.

I get that, but what I’m saying is that calling deep learning “just fancy comparison engine” frames the concept in an unnecessarily pessimistic and sneery way. It’s more illuminating to look at the considerable mileage that “just pattern matching” yields, not only for the practical engineering applications, but for the cognitive scientist and theoretician.

Furthermore, what constitutes being “actually creative”? Consider DeepMind’s AlphaGo Zero model:

Mok Jin-seok, who directs the South Korean national Go team, said the Go world has already been imitating the playing styles of previous versions of AlphaGo and creating new ideas from them, and he is hopeful that new ideas will come out from AlphaGo Zero. Mok also added that general trends in the Go world are now being influenced by AlphaGo’s playing style.

Professional Go players and champions concede that the model developed novel styles and strategies that now influence how humans approach the game. If that can’t be considered a true spark of creativity, what can?

To counter the grandiose claims that present-day LLMs are almost AGI, people go too far in the opposite direction. Dismissing it as being only “line of best fit analysis” fails to recognize the power, significance, and difficulty of extracting meaningful insights and capabilities from data.

Aside from the fact that many modern theories in human cognitive science are actually deeply related to statistical analysis and machine learning (such as embodied cognition, Bayesian predictive coding, and connectionism), referring to it as a “line” of best fit is disingenuous because it downplays the important fact that the relationships found in these data are not lines, but rather highly non-linear high-dimensional manifolds. The development of techniques to efficiently discover these relationships in giant datasets is genuinely a HUGE achievement in humanity’s mastery of the sciences, as they’ve allowed us to create programs for things it would be impossible to write out explicitly as a classical program. In particular, our current ability to create classifiers and generators for unstructured data like images would have been unimaginable a couple of decades ago, yet we’ve already begun to take it for granted.

So while it’s important to temper expectations that we are a long way from ever seeing anything resembling AGI as it’s typically conceived of, oversimplifying all neural models as being “just” line fitting blinds you to the true power and generality that such a framework of manifold learning through optimization represents - as it relates to information theory, energy and entropy in the brain, engineering applications, and the nature of knowledge itself.

The real problem is folks who know nothing about it weighing in like they’re the world’s foremost authority. You can arbitrarily shuffle around definitions and call it “Poo Poo Head Intelligence” if you really want, but it won’t stop ignorance and hype reigning supreme.

To me, it’s hard to see what cowtowing to ignorance by “rebranding” this academic field would achieve. Throwing your hands up and saying “fuck it, the average Joe will always just find this term too misleading, we must use another” seems defeatist and even patronizing. Seems like it would instead be better to try to ensure that half-assed science journalism and science “popularizers” actually do their jobs.