"How were we to know?"

All we knew two decades ago should have stopped this train wreck. Why didn't that knowledge stop it?

Goran's post

Goran S. Milovanovi?, Phd is SPOT ON:

And there's another dividing line, another median in this whole story: whether you were (a) educated in cognitive sciences before (around) 2010, or (b) directly educated in Data Science and ML/AI after 2010. Around 2010 was the moment it became clear that engineering breakthroughs in Deep Learning would begin to yield practical and truly incredible results compared to their predecessors in associative, statistical learning. At the same time, it was the year when hype overtook healthy scientific conservatism and the hacker mentality took precedence over scientific rationality. This has brought us to the point where even a script kiddie can deploy an LLM on data and introduce features into a production environment without even understanding the simple proofs of the limitations of neural networks, which Marcus, Pinker, Fodor, Zenon Pylyshyn, and others have been writing about for decades.

Lucy Suchman could have told you why this would fail in 1987.

I studied cognitive science and AI from 2002-2004 and wrote about it steadily from 2007-2010. I differ slightly on the idea that it became clear that engineering breakthroughs in "Deep Learning" was as clearly breakthrough as that. While we were able to train algorithms for personalization to drive sales actually liberating data to become insight remained elusive. We advanced on mass surveillance pretty well with it.

The most impressive strives were happening in images only, in biotech and aeriel surveillance. This led to abandoning the projects that held the most promise for overcoming the very limitations we are seeing in LLM failures today, which, lets be honest is deploying text as image far more than anyone cares to address or admit. This has distinct problems for something that promises to do analysis. Yes, I am saying all the failures we are seeing now were known then, when I was studying, even through theory alone on napkin sketches. We had less computing power, our models and experiments were smaller. But you could do the math. Literally. The barriers we were working on then were:

1. The impossibility of creating autonomy and agency with classical computationalism. (Delancey 2004).

2. The way semantic systems constructed through computation would easily collapse through repeated use. (Bishop 2006).

3. The limitations of repeat exposures of language to train any programmed model, whether robotic or strictly computational (Clark 1997).

4. The limits of classical computationalism and binary for language, intelligence, and analysis. (LUCY Suchman 1987!)

These are just the highlights.

For me, I had a little future shock and thought for a hot second that maybe they had advanced in some unforeseen way. I assumed that some of these problems had been solved in ways I hadn't seen coming. I was a little excited if confused and mystified. Just as I was when there were social data platforms claiming to do automated sentiment analysis. A few months of testing--basic software testing--and we could see that they hadn't. And that the results were strangely poor and erratic, even worse that you'd think in models so created,

What I want to know is did anyone truly believe that these limitations could be overcome by size and force or did they just use it as an excuse to get people to hand over their data?Someone said to me early on: "You might not like the ethics or the elegance of the solution but they did it and now look at this amazing thing we have!"

But: what do we have? Hacker folks and business men are the only ones impressed because they don't know any better and usually have no means to validate what their systems produce. Quicker ways to cut and paste and override intellectual property rights and privacy laws? The entire digital commons repackaged and sold back to us? I'm not sure. The massive cybersecurity vulnerabilities this has created will be talked about for decades. Perhaps with pen and paper and postage stamps, if the worst case scenario happens. I'm optimistic the pain of this bubble will force some people to get smart and better solutions will emerge.

As an assistive tech it has some utility. But at what cost?

We are about to find out aren't we.

#ai #hypebubble