This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
A follow-up on last week's discussion of LLMs and AI. [TLDR: I tested ChatGPT and was pretty shocked at how well it performed]
To recap, one of the criticisms of LLMs is that they are unable to create models of the world. So for example, according to one commentator (I believe it's Gary Marcus) an LLM will attempt to play impossible chess moves. Despite having the rules of chess in its training data, as well as large numbers of chess matches, it's (apparently) unable to have a working internal model of chess. By contrast, a reasonably bright teenager can learn pretty quickly how to play perfect chess. (Perfect in the sense of never making an illegal move.)
According to Google's AI (yes, I appreciate the irony here).
I decided to test this idea that LLMs are unable to model the world by creating a very simple game; in order to play the game it's necessary to have a simple model of the game state. As expected, the LLM made numerous errors.
But what was interesting was that I pointed out the errors to the LLM and it told me that it could fix these problems. And it did so in an interesting way: After each move in the game, it spelled out the game state in text. After that, it stopped making errors. Admittedly, this is a very cumbersome way to model the world -- by means of an iterative written description. But it seemed to work well for this very simple game. To my mind, this was rather astonishing and shocking. And if there is a cumbersome way to accomplish something, you can usually count on computers to accomplish it anyway by means of throwing more and more processing power at the situation. (Actually, that's not totally true, since some tasks have exponential or even combinatorial time complexity. But still.)
In the last thread, my opinion was that LLMs are missing something essential. And I still think that, but I wouldn't be surprised at all if LLMs required very little theoretical augmentation to reach AGI.
And that's the worst way to solve this problem. What it should have done is write a tool call (to a state machine in python it itself wrote, mapping your rules to logic; or an open source chess engine; or a geodata software suite - whatever is required), and then just using the language model for interacting with you and the tool. This solves many problems created by the lack of a word model - much better than plain text ever could.
I suspect GPT would have solved the problem this way with a little nudge, or if called from a coding harness instead from the chat window, because many of those capabilities were added over the last few months. Previously difficult world model questions like "I'm driving from Bordeaux to Prague. The drive will take two days. Find a spot to spend the night, roughly in the middle, where I can go rock climbing. Don't add more than 60 minutes of additional driving." are now trivial, since the LLM just creates a hand full of Google Maps calls. Without tool use, arguing its way to correct solution was not possible - the travelling salesman cannot be argued with.
Unfortunately, this only shifts the problem a little. You now need good tools for manipulating the world. Many already exist, but others need to be created now. Some of those tools need a lot of non-LLM AI themselves (lots of robotics is like that, but getting a robot to fold laundry seems to be in reach), others probably need an entirely new approach (getting an LLM to call a CAD kernel to create 3D parts and assemblies seems like a monumental task, both in old-fashioned logic and vision capabilities).
Worst in what way? Time efficiency? Memory efficiency? Accuracy?
Yes, all of those. Instead of burning hundreds/thousands of tokens on game state/memory each turn, you can run a simple state machine for comparatively free compute. This will also be more deterministically correct (you could even do correctness proofs if you really want to, or just have the LLM write a bunch of tests), because if you keep game state as tokens, you currently have no guarantees/tests.
Anything else? Or is that pretty much it?
Having the LLM do expensive things cheaply in e.g. python is doubly beneficial, since a lot of the tool calls (again: game states, chess engines, navigation software, ect.) are not only much cheaper (in FLOP/s), but also are often CPU loads instead of GPU loads. That frees up the GPUs for LLM inference of other customers, while a cheap CPU runs some python scripts on the side.
An again, the results will be much better. Using python for math will get correct results to math questions. Calling a chess engine means almost all humans will loose that game. Only calling a map app unlocks correct world model answers to navigation and spatial reasoning questions.
Ok, but other than time efficiency, memory efficiency, and accuracy, are there any other ways in which what I discussed was the "worst way to solve this problem"?
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link