I’m confused why there seems to be a dismissal of the most basic ‘strange loop’ of the LLM - the fact that it’s evaluating a context to choose the next word, then reevaluating in a context where that word has been appended.
That always seemed to me like the essence of a Hofstadterish strange loop, so the emergence of Hofstadterish phenomena (self rep, etc) doesn’t seem surprising.
That's a really interesting point. I hadn't drawn that connection before.
And if anyone's reading this and hasn't read Hofstadter, you're making a mistake, it's utterly perspective-changing stuff. Well, it was for me, at least.
And don’t feel you need to go straight to GEB, either. I think his collection of Scientific American columns, Metamagical Themas, is also great and much more accessible.
I read it when I was about 9 or 10, I didn't understand any of it at all but I used to pick it up and look at the pictures of the Escher drawings, which were amazing.
Then, after I graduated from college, a friend of mine who was finishing his physics degree mentioned that he had just finished it and that it was life-changing. I read it again and deeply regretted having put it down for those 12 intervening years.
One of my elementary school classmates brought printouts of Escher's Waterfall to share. I accepted one, took it home, and taped it to my bedroom wall. My father saw it and loaned me his copy of Gödel, Escher, Bach.
At first, I flipped through the book looking at the Escher prints. My second time through, I read the dialogues. Next I had a go at reading the chapters and got lost. Maybe on my fifth read-through (years later) I actually attained an understanding of Gödel's Theorem (at least for a brief moment before it vanished again).
I credit GEB (and its dialogues like Little Harmonic Labyrinth) for helping prepare me to deal with the kind of multi-level abstraction involved in developing a language interpreter or machine simulator. Especially that time I had to debug a crash that only occurred when running my simulator inside a simulated instance of itself (i.e. when three levels deep, but not two).
With computer generated maps, you can just always use a minimally distorting conical or polar projection centered on the exact point you’re focusing on.
That also happens to be broadly what you would get looking straight down at the earth from above that point, which seems like it would also be the least confusing option.
The prime meridian being the least skewed area of the equal earth and Robinson projections, with their bulgy nature, actually serves to make them both offer somewhat inconvenient depictions of places like Alaska and New Zealand and Japan.
Mercator is far more globally evenhanded in that regard, being a cylindrical projection.
To test this, change the history in the context to indicate that the model did or recommended something completely different than it actually did, and then ask it to explain why. You’ll still get a plausible explanation.
Possibly an inevitability, since LLM text is specifically generated to be low entropy. It’s smooth, optimized for minimal perplexity.
And if you don’t find any of the word choices a little bit perplexing, your brain can just gloss over it. There is no meaning here. It’s like smalltalk.
Interesting — do you happen to have any suggestions for something I could read about this?
It really bugs me, this aspect of LLMs, and yet I find myself sounding like a lunatic when I describe it, until others chime in and say they have the same experience.
Chicago is lowest for adjusted risk because it just scores very low for physical risk on their scale. Compared to other cities the adjustment to its risk score for resilience is about typical - not great and not terrible - which suggests that yes, for what physical risks it does face (summer heat) it could probably do more to reduce those risks.
A lot of professional skill nowadays is misdirected from ‘knowing how to do the thing’ into ‘learning how to use a computer to do the thing’.
The role of music producer includes a large chunk of ‘knowing how to program synths and samples and drums and effects’. The role of industrial designer includes a lot of ‘knowing how to model things in CAD’. Even roles like academic researchers require skills like ‘formatting LaTeX’.
I’m genuinely quite excited for a shift in human computer interaction which reduces that ‘knowing how to do things on the computer’ tax.
But there are a whole swathe of jobs where a lot of what people do is ‘know how to do the thing on a computer’. Some of that is really because the bureaucracy has been embodied in a computer system so for example the job of handling an insurance claim seems to involve a lot of ‘knowing how to operate the claims computer system’ - actually that person’s skill is ‘knowing how to navigate the claims bureaucracy’.
Completely agree. How many people over the past 20 years have been musically gifted, but found ProTools or Logic impenetrable? The AI unlock is about enabling that creativity, not replacing it.
Really? You’ve got this guy Dave Fester on there saying he was going to own the website issues and was coordinating everything to do with how the products surfaced on the web. That seemed like pretty clear accountability and ownership to me. What do you want? Some more performative mea culpas?
Point wasn't about fixing the exact issues, the point was how it even got that bad. Of course the easiest route to comprehending the complaint by Bill is to immediately fix the specified issues, making Bill a simple QA instead of concerned leader.
I think accountability would involve trying to understand how it got that bad, and figuring out a new process for preventing it from happening again. Though I'm sure that never happened.
In my humble opinion, some of the blame has to land on Bill himself as he has the full power and agency to build a style of management in his company that would've prevented this from happening. Instead he just scolded a bunch of senior execs who usually find a way to just delegate some work to tuck away any visible issues and go on with their merry lives.
I’m confused why there seems to be a dismissal of the most basic ‘strange loop’ of the LLM - the fact that it’s evaluating a context to choose the next word, then reevaluating in a context where that word has been appended.
That always seemed to me like the essence of a Hofstadterish strange loop, so the emergence of Hofstadterish phenomena (self rep, etc) doesn’t seem surprising.
reply