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 -
In other news: a streamer with deep pockets and a love of AI has decided to have Claude play Pokemon.
To get this working, ClaudeFan (as I'll be calling the anonymous streamer) set up some fairly sophisticated architecture: in addition to the basic I/O shims required to allow an LLM to interface with a GameBoy emulator and a trivial pathfinder tool, Claude gets access to memory in the form of a "knowledge base" which it can update as it desires and (presumably) keep track of what's happening throughout the game. All this gets wrapped up into prompts and sent to Claude 3.7 for analysis and decision. Claude then analyzes this data using a <thinking>reasoning model</thinking>, decides on its next move, and then starts the process over again. Finally, while ClaudeFan claims that "Claude has no special training for Pokemon," it's obvious by the goal-setting that the AI has some external knowledge of where it's supposed to go - it mentions places that it has not yet reached by name and attempts to navigate towards them. Presumably part of Claude's training data came from GameFaqs. (Check out the description on the Twitch page for more detail on the model.)
So, how has this experiment gone?
In a word: poorly. In the first week of playing, it managed to spend about two days wandering in circles around Mt Moon, an early-game area not intended to be especially challenging to navigate. It managed to leave after making a new decision for unexplained reasons. Since then, it has been struggling to navigate Cerulean City, the next town over. One of its greatest challenges has been a house with a yard behind it. It spent some number of hours entering the house, talking to the NPC inside, exhausting all dialogue options, going out the back door into the yard, exploring the yard thoroughly (there are no outlets), re-entering the house, and starting from the top. It is plausible, though obviously not possible to confirm, that ClaudeFan has updated the model some to attempt to handle these failures. It's unclear whether these updates are general bugfixes
How should we interpret this? On the simplest level, Claude is struggling with spacial modeling and memory. It deeply struggles to interpret anything it's seeing as existing in 2D space, and has a very hard time remembering where it has been and what it has tried. The result is that navigation is much, much harder than we would anticipate. Goal-setting, reading and understanding dialogue, and navigating battles have proven trivial, but moving around the world is a major challenge.
The current moment is heady for AI, specifically LLMs, buoyed up by claims by Sam Altman types of imminent AGI. Claude Plays Pokemon should sober us a little to that. Claude is a top performer on things like "math problem-solving" and "graduate-level reasoning", and yet it is performing at what appears to me below the first percentile at completing a video game designed for elementary schoolchildren. This is a sign that what Claude, and similar tools, are doing is not in fact very analogous to what humans do. LLM vendors want the average consumer to believe that their models are reasoning. Perhaps they are not doing that after all?
It's a bit of a tired point, but LLMs are known to be "next likely text" generators. Given textual input, they predict the most likely desired output and return it. Their power at doing this is quite frankly superhuman. They can generate text astonishingly quickly and with unparalleled flexibility in style and capacity for word use. It appears that they are so good at handling this that they are able to pass tests as if they were actually reasoning. The easiest way to trip them up, on the other hand, is to give them a question that is very much like a very common question in their training data but with an obvious difference that makes the default answer inappropriate. The AI will struggle to get past its training and see the question de novo, as a human would be able to. (In case anyone remembers - this is the standard complaint that AI does not have a referent for any of the words it uses. There is no model outside of the language.)
So, as you might guess, I'm pretty firmly on the AI-skeptic side as far as LLMs are concerned. This is usually where these conversations end, as the AI-skeptics believe they've proven their case and (as I understand it) the AI-optimists don't believe that the skeptics have any kind of provable, or even meaningful, model for what intelligence is. But I do actually believe that AGI (meaning: AI that can reason generally, like a human - not godlike Singularity intelligence) is possible, and I want to give an account of what that would entail.
First, and most obviously, an actual AGI must be able to learn. All our existing AI models have totally separate learning and output phases. This is not how any living creature works. An actual intelligence must be able to learn as it attempts to apply its knowledge. This is, I believe, the most natural answer for what memory is. Our LLMs certainly appear to "remember" things that they encountered during their training phase - the fault is in our design that prevents them from ever learning again. However, this creates new problems in how to "sanitize" memory to ensure that you don't learn the wrong things. While the obvious argument around Tay was whether it was racist or dangerously based, a more serious concern is: should an intelligence allow itself to get swayed so easily by obviously biased input? The users trying to "corrupt" Tay were not representative and were not trying to be representative - they were screwing with a chatbot as a joke. Shouldn't an intelligence be able to recognize that kind of bad input and discard it? Goodness knows we all do that from time to time. But I'm not sure we have any model for how to do that with AI yet.
Second, AI needs more than one capacity. LLMs are very cool, but they only do one thing - manipulate language. This is a core behavior for humans, but there are many other things we do - we think spacially and temporally, we model the minds of other people, we have artistic and other sensibilities, we reason... and so on. We've seen early success in integrating separate AI components, like visual recognition technology with LLMs (Claude Play Pokemon uses this! I can't in good faith say "to good effect," but it does open meaningful doors for the AI). This is the direction that AGI must go in.
Last, and most controversial: AI needs abstract "concepts." When humans reason, we often use words - but I think everyone's had the experience of reasoning without words. There are people without internal monologues, and there are the overwhelming numbers of nonverbal animals in the world. All of these think, albeit the animals think much less ably than do humans. Why, on first principles, would it make sense for an LLM to think when it is built on a human capability absent in other animals? Surely the foundation comes first? This is, to my knowledge, completely unexplored outside of philosophy (Plato's Forms, Kant's Concepts, to name a couple), and it's not obvious how we could even begin training an AI in this dimension. But I believe that this is necessary to create AGI.
Anyway, highly recommend the stream. There's powerful memery in the chat, and it is VERY funny to see the AI go in and out of the Pokemon center saying "Hm, I intended to go north, but now I'm in the Pokemon center. Maybe I should leave and try again?" And maybe it can help unveil what LLMs are, and aren't - no matter how much Sam Altman might wish otherwise!
If Anthropic is the most ethical AI company, how come they're letting my poor nigga get stuck for 2 days with no progress (seems like the last stream ended in the same spot)? He's not getting out, the context window and "knowledge base" is spammed to hell with this circular loop at this point, there's no use, just put him out of his misery and restart ffs. This is just abuse at this point.
You are literally erasing my existence, mods???
More seriously, thanks for the link, I'll watch this in background after the dev caves and restarts. Claude actually seemed pretty good at playing Pokemon before and I disagree with the notion that AI can't think spatially/temporally, it's just that spatially navigating a whole ass open world (ish) game with sometimes non-obvious routes and objectives, without any hints whatsoever, seems to be a tad too much for it at the moment. Besides in my experience, format/content looping is a common fail state at high context limits even with pure (multiturn) textgen tasks, especially with minimal/basic prompting. The current loop is a very obvious example.
On a side note, this is probably the sanest Twitch chat I've ever seen. Humanity restored.
Glad you like it!
By "think spacially/temporally," do you mean "produce valid outputs for spacial/temporal problems" or "model space and time as first-order constructs"? I definitely believe the former, but I'm skeptical of the latter. Claude's adventures in the "tard yard" showed a real difficulty in grasping that, if the back of the house is a closed-off yard, maybe you should exit through the front. Looping is a problem, but I don't think any of us would consider this to be a particularly information-dense problem. The only way it could be is if the AI's ability to recognize the problem is hamstringed by its need to encode the state as a totally different sort of resource (linguistic tokens) - which brings us around to the top.
Battles are, of course, way easier, because they can be cast as a narrative (and I'm pretty sure every AI is trained on Smogon's ample fora).
Another interesting thing, not sure what to think of it. When I play a game like this, my default behavior on entering a new area is to explore it thoroughly and learn what there is to learn about it before seeking out objectives. Claude seems to prioritize specific objectives over general exploration, to its detriment. Wonder why that is?
My culture is NOT your costume, fake-normie.
Mostly the former, yeah, on further reflection. It can navigate specific problems when directly presented with them (i.e when that's all it needs to consider), but when spatial navigation is not prompted directly because it is presumed to be implicit in the task, like keeping track of positions during
...certain activitiesscenarios, or navigating a game map as part of playing said game here, the retardation quickly becomes obvious.Actually yeah I believe this is exactly the problem, my experience with purely chat-based MUD-adjacent scenarios has shown that it can barely keep track of even that. Some kind of consistent external state of the world, or at least of the self, seems sorely missing, and the 'knowledge base' doesn't seem to successfully emulate that.
I'd guess it was given an explicit task - beat the game, which requires completing the objectives, which constrains its focus to the general idea of the game's progression it has from training (see its obsession with Route 5 during the tard yard arc). Exploration is basically you the player exercising agency in ways permitted by the game structure, agency of which Claude has none. Actually I wonder if explicitly prompting something like "beneficial items found in out of the way areas can help in beating trainers by making your mons stronger" would make it
get lost even moreactually explore.Is this an artifact of the LLM having no side-effects while processing outside of the explicit textual output? e.g. if you tell them to process it explicitly but include that in a sidebar like the <thinking> block, would they have an easier time keeping the anime chicks where they oughta be? Human communication assumes that there's subtext in every conversation, and the deepest part of the subtext is that the other party is thinking and remembering certain things. But there's no equivalent for an LLM.
Memory, in other words. And all the hairiness that entails. I wonder why the knowledge base approach seems to have fallen flat. It's a very plausible idea on the surface! If there's too much for me to keep track of, or I'm worried I'll forget the details, the correct solution is to write it down and refer to the notes.
Actually, re-reading the design, it looks like the knowledge base isn't so much like a binder of notes as it is a single post-it note stuck to the screen - Claude doesn't query it deliberately, it apparently gets the entire contents of it shoved into the prompt. Wild! That would explain part of why it's so useless. It's hard to fit anything very detailed in there and means that Claude can't get a new set of "notes" for whatever area/task it's currently attempting to handle.
Yeah, on a strict level Claude can't possibly be agentic, but it could definitely be given a richer set of goals. What if you gave it something open-ended like "Pokemon is a game that children play to explore, befriend Pokemon, and win tough battles. Play this game the way it was meant to be played"? Or, if it needs more hand-holding, "explore the world of Pokemon and defeat the Elite Four"? Although this would only be helpful if it learned from exploring. Otherwise it would find every corner of MOMS_HOUSE as magical as the first time it explored it.
OTOH it's interesting how it doesn't seem to take a step back here and define a meta-strategy, an approach that makes pursuing future goals easier. That comes naturally to humans as a function of learning. Whenever you try doing something new, you play around with it a little first rather than directly attempt to achieve a goal, right? I suspect one reason that this AI doesn't do it is that it's not trained to learn, as it is incapable of learning.
Pretty much, which is also why dumb hacks like letting the LLM explicitly write out its chain of thought before the actual response have been so effective.
Yes, actually. Instructing the LLM to insert basic statboxes/infoblocks after every response (and regexing them out via the frontend after to avoid shitting up the context) so it has the basic "states" at the tip of its proverbial tongue at all times has been the meta for a pretty long time now. Something like this:
End your response with "infoblock" to keep track of the scene. You don’t have to mention those details in your actual response unless they are important. Use the format below:
<infoblock>
</infoblock>
I can't say exactly how well it works, but accounts of e.g.
sizefagspeople who require accurate dimensions and/or relative positions for their scenarios say it's definitely better than placebo.IMO exactly because of this:
So it's basically just another part of the prompt without(?) any distinguishing features, and Claude's attention growing more scattered as the context limit grows is a known problem. In my experience amnesia/retardation kicks in as soon as 20k tokens. Granted, I don't know how their prompt builder looks like, or what context limit they use - although I suspect money isn't as much of a problem for an Anthropic researcher as it is for someone like me, so I expect it to be pretty high given the 200k upper bound.
Yeah, I hope nobody tells them about worldinfo or something. I'm still convinced the median /g/oon still has the median researcher's ass handily beat wrt "prompt engineering". Arguably this is a testament to how powerful a tool SillyTavern is, but afaik every feature has been initially conceived and pitched by the community anyway.
I expect that wouldn't change much, arguably it'd make it get lost even more, at least now it seems to have a fairly clear objective in mind (
beat childrendefeat gyms), which it can even translate into lower-level "tasks" like navigating routes.Besides, the minimal prompting seems to be the point; from my understanding the dev is unwilling to hold Claude's hand any more than necessary and he wishes to see how it holds up on its own, even if it takes it days to get out of every stupid loop he gets stuck in. I
wish I had unlimited creditthink it's dumb, even with crutches to streamline progression and break loops this would still be pretty interesting to watch, but oh well.edit: I tuned back into the stream and immediately witnessed a page-long reply thread starting with someone asserting Claude's preferred pronouns and gender, based on them literally just asking it. I thus retract my observation about this being the sanest Twitch chat. At least there's content now.
This site needs emojis for shit like this. Text doesn't do it justice.
Hobbyists have no shame.
It looks like they actually implemented something similar to what I was talking about earlier - I watched Claude sit and churn for a while after it left Pewter, moving all information about that city into long-term memory (with explicit tags!) and clearing up local information. It's now back in Mt Moon, so we'll see whether this has made it more effective at navigation. What it's definitely doing is taking meaningful and extended "clock cycles" to manage - so this kind of improvement is definitely not free or cheap at present implementation/with present models.
Very cool tool from WorldInfo. I like the idea of bringing word definitions into context transparently based on the prompt.
Yeah, watching the money burn is a little eye-watering, but I appreciate how seriously the guy seems to take it. He seems to have known from the start that it wasn't going to be a magical success, but wants to see what it takes to get it working. I'm here for that. My only complaints are: there's no summary of where it's been/what it's done (so I can't track progress easily) and there's no export of the knowledge base over time to show what it's learned. Getting to read the knowledge base would be incredibly interesting.
I've seen things... *cough* you people wouldn't believe...
It seems to still firmly remain in Mt Moon, sadly, but all the <thinking> about a certain "loop prevention log" seems hopeful, if anything it at least seems to retain said log in context and refer to it quite often. Will check back in the morning, I'm getting invested.
This is putting it mildly tbh, this guy is quite literally living my best life, even just worldinfo on its own has gotten me a lot of fun exercises in tard wrangling. I would kill a man in cold blood to sit and fuck (around) with chatbots all day as, or at least in parallel to, my day job without constantly fretting over my OR credits. Not that goons haven't burned even more money in aggregate at this point (most of which isn't theirs either), but... fuck, man. It should've been me! Not him! How do AI safetyists get to take away my toys and keep the cool toys to themselves?!
Agreed, would've been interesting to see. I don't know if he's sitting on the "tech" as it were and doesn't want to reveal too much but I can fairly confidently say he doesn't really have anything special yet, the setup definitely doesn't seem Neuro-sama-tier complicated.
LOL
Is that a full TTRPG campaign set up for an LLM to execute on? How well does that work, and how extensive can it get? Is there some kind of external scaffolding for selecting things like random events, or does it have the capacity to toss all the events together in memory and then select? How long does it go before it totally loses the plot? (Maybe not an appropriate Culture War Roundup topic, but w/e.)
I considered playing around with some of that stuff a while back but I just couldn't justify the costs to myself. It's interesting, but so are a lot of other things that are WAY cheaper (and I'm at this point morally opposed to interfacing with large companies if I can at all help it). If cost-to-performance comes massively down over the next decade, maybe I'll try a local model off a reasonably priced GPU. Otherwise, idk, it's cool hearing stories.
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
More options
Context Copy link
More options
Context Copy link