LLM links

The WSJ hallucinates about LLMs; David Strom on Movable Type's book writer; Henrik Karlsson on humans learning from AI; me on a podcast with Alan Pentz

The WSJ printed an “explainer” of artificial intelligence that I think was worse than nothing. Bart Ziegler wrote,

These powerful computer systems, called generative AI or large language models, are fed enormous amounts of information—hundreds of billions of words—to “train” them. Imagine if you could read pretty much everything on the internet, and have the ability to remember it and spit it back. That is what these AI systems do, with material coming from millions of websites, digitized books and magazines, scientific papers, newspapers, social-media posts, government reports and other sources.

That is not how LLMs work! They are not fed “information,” and they do not “remember it and spit it back.”

Instead, as my readers know, LLMs take a given word and find patterns of how it is used. What word usually precedes it? What word almost never precedes it? What words earlier in a sentence or paragraph indicate that this word is likely to appear soon? What words earlier in a sentence or paragraph indicate that this word is unlikely to appear soon? These sorts of characteristics are quantified and coded as vectors. What an LLM knows are these vectors.

The training data supplies the word patterns that get analyzed and coded into the vectors. It does not supply the LLM with “information.”

If you read Ziegler’s article carefully, you will see that he often comes close to the correct perspective. But by getting it wrong early, I would bet that he leaves most of his audience with the wrong impression about how LLMs work.

David Strom writes,

MovableType’s AI creates “10 chapters spanning 150+ pages, and a whopping 35k+ words” (or so they say) of… basically gibberish. They of course characterize it somewhat differently, saying its AI output is “highly specific & well researched content,” It isn’t: there are no citations or links to the content. The output looks like a solid book-like product with chapters and sub-heads but is mostly vacuous drivel. The company claims it comes tuned to match your writing style, but again, I couldn’t find any evidence of that. And while “each chapter opens with a story designed to keep your readers engaged,” my interest waned after page 15 or so.

It seems to me that something like this could be very useful. It would be nice if I could feed the essays linked on my Main Routine page and have it turn out a book.

Of course, I am increasingly skeptical of the book format. What I really want to create is a dedicated GPT that uses those essays and other writing of mine to represent me interactively on the topics addressed there.

Henrik Karlsson writes,

For many decades, it seemed professional Go players had reached a hard limit on how well it is possible to play. They were not getting better. Decision quality was largely plateaued from 1950 to the mid-2010s…

After a few years, the weakest professional players were better than the strongest players before AI. The strongest players pushed beyond what had been thought possible.

And it wasn’t simply that they imitated the AI, in a mechanical way. They got more creative, too. There was an uptick in historically novel moves and sequences.

It is impressive how dramatically human Go play improved when people were able to learn from AI’s. But I have a different mental model of how humans learn from AI’s in this context, based on my experience with Othello, another strategy board game.

First of all, computer Othello programs use a combination of pattern matching (looking through databases of games in order to find the characteristics of better positions) and alpha-beta pruning (what we would call thinking ahead).

As far as I know, large language models don’t do alpha-beta pruning. It seems to me an exercise that only makes sense when you’re playing a game against someone and gauging their likely moves. So the relationship between human intelligence and computer intelligence may be different in board games than in other realms.

The other point that I would make is that human players and computers win Othello games the same way: by making fewer suboptimal moves. When I win, it’s not because I come up with a single move that is so brilliant that other players would never have considered it. It’s that over the course of a long game, involving many moves where there seem to be close choices, my opponent picks the second- or third-best move more often than I do.

Computers win the same way. I don’t experience looking at a computer’s move and saying, “I never would have considered that!” Instead, what I see is that I would have picked something different, and my selection turns out to be inferior. I make more inferior moves than the computer does, and so I lose.

And I learn from computer Othello programs by noticing the situations in which my intuition is mistaken. It’s not that I memorize what the computer would do. I can only do that with openings. Later in the game, I pattern-match. I see that my instinct tells me to prefer move 1, but I look at the pattern and I guess that the program would choose move 2.

My guessing is not perfect. Sometimes I fail to guess that a computer would make a different move. And sometimes I guess that the computer would pick move 2 and it actually prefers move 1, which is what I was going to make.

So I still lose to the computer. But by guessing what the computer would do, I play better than I used to play before I studied with the computer. That is how I become a better player by training with the computer.

Karlsson writes,

My guess is that AlphaGo’s success and alien playing style forced the humans to reconceptualize the game and abandon weak heuristics. This let them see possibilities that had been missed before.

I would not describe myself as having reconceptualized Othello. To a fellow Othello expert, I might say something like “I take an unbalanced edge to gain a tempo in more situations than I used to.” I would characterize that as a subtle re-evaluation of trade-offs that I have known about for a long time. It seems a bit of an exaggeration to use phrases like “abandon weak heuristics” or “see possibilities that had been missed before.”

Note that I do not play Go, and the effect of AI on top Go players may be qualitatively different.

In a podcast with Alan Pentz, I give some of my current views on AI. I think it is pretty information dense and worth listening to, especially the second half of the 40-minute show. Let me know.

Listen at one of these three places:

• Apple - apple.co/4b5M1l5

• Spotify - spoti.fi/4bmMAHt

• bCast - bit.ly/47HkALq

substacks referenced above:

@

After AI beat them, professional Go players got better and more creative