[gabelschlager]

Well, Chomsky already dismissed corpus based linguistics in the 90s and 2000s, because a corpus (large collection of text documents, e.g., newspaper, blog post, literature or everything mixed together) is never a good enough approximation of the true underlying distribution of all words/constructs in a language. For example, a newspaper-based corpus might have frequent occurences of city names or names of politicians, whereas they might not occur that often in real everyday speech, because many people don't actually talk about those politicians all day long. Or, alternatively, names of small cities might have a frequency of 0.

Naturally, he will, and does, also dismiss anything that occured in the ML field in the past decade.

But I agree with the article. Dealing with language only in a theoretical/mathematical way, not even trying to evaluate your theories with real data, is just not very efficient and ignores that language models do seem to work to some degree.

[blululu]

This is a bit lateral, but there is a parallel where Marvin Minsky will most likely be best remembered for dismissing neural networks (a 1 layer perceptron can't even handle an xor!). We are now sufficiently removed from his heyday where I can't really recall anything he did besides the book Perceptrons with Seymour Papert (who went on to do some very interesting work in education). There is a chart out there about ML progress that makes a conjecture about how small the gap is between what we would consider that smartest and dumbest levels of human intelligence (in the grand scheme of information processing systems). It is a purely qualitative vibes sort of chart, but it is not unreasonable that even the smartest tenured professors at MIT might not be that much beyond the rest of us.

[masswerk]

This dismissal of Minsky misses that Minsky had actually extensive experience with neural nets (starting in the 1950s, with neural nets in hardware) and was around 1960 probably the most experienced person in the field. Also, in Jan 1961, he published “Steps Toward Artificial Intelligence” [0], where we not only find a description of gradient descend (then "hill climbing", compare sect. B in “Steps”, as this was still measured towards a success parameter and not against an error function), but also a summary of experiences with this. (Also, the eventual reversal of success into a quantifiable error function may provide some answer to the question of success in statistical models.)

[0] Minsky, Marvin, “Steps Toward Artificial Intelligence”, Proceedings of the IRE, Volume: 49, Issue: 1, Jan. 1961: https://courses.csail.mit.edu/6.803/pdf/steps.pdf

[riku_iki]

Gradient descent was invented before Minsky. Imo, Minsky produced some vague writings, with no significant practical impact, but this is enough for some people to claim his founder's role in the field.

[masswerk]

Minsky was actually a pioneer in the field, when it came to working with real networks. Compare

[0] “A Neural-Analogue Calculator Based upon a Probability Model of Reinforcement”, Harvard University Psychological Laboratories, Cambridge, MA, January 8, 1952

[1] “Neural Nets and the Brain Model Problem”, Princeton Ph.D dissertation, 1954

In comparison, Frank Rosenblatt's Perceptron at Cornell was only built in 1958. Notably, Minsky's SNARC (1951) was the first learning neural network.

[riku_iki]

> when it came to working with real networks. Compare

my understanding is that that no one knows what that SNARK thing was, he built something on the grant, abandoned it shortly after that, and only many years later he and fanboys started using it as foundation of bold claims about his role in the field.

[masswerk]

Well, his papers are out there to read.

[cma]

Take the amount of language a blind 6 year old has been exposed to. It is nothing like the scale of these corpsuses but they can develop a rich use of language.

With current models if you increased parameters but gave it a similar amount of data it would overfit.

[riku_iki]

It could be because kids are gradually and structurally trained through trials, errors and manual corrections, which we somehow don't do with NN. He wouldn't be able learn language if only exercises he would be doing is to guess next word in sentence.

[aaroninsf]

For me this is a prototypical example of compounded cognitive error colliding with Dunning-Kruger.

We (all of us) are very bad at non-linear reasoning, reasoning with orders of magnitude, and (by extension) have no valid intuition about emergent behaviors/properties in complex systems.

In the case of scaled ML this is quite obvious in hindsight. There are many now-classic anecdotes about even those devising contemporary scale LLM being surprised and unsettled by what even their first versions were capable of.

As we work away at optimizations and architectural features and expediencies which render certain classes of complex problem solving tractable by our ML,

we would do well to intentionally filter for further emergent behavior.

Whatever specific claims or notions any member has that may be right or wrong, the LessWrong folks are at least taking this seriously...

[aaroninsf]

To some degree is quite an understatement. :)

My own hobby horse of late is that independent of its tethering to information about reality available through sensorium and testing, LLM are already doing more than building models of language qua language. Write up someone pointed me at: https://thegradient.pub/othello/

[bmitc]

How is an understatement? And what do people mean by language models working well? From what I can tell, these language models are able to form correct grammar quite surprisingly well. However, the content of such, is quite poor and often void of any understanding.

[masswerk]

"seem to work to some degree", appears much like a second order argument to this debate… ;-)