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!
Well, if that's what you want to call an Anthropic researcher who decided to make their experiment public.
"Claude Plays Pokémon continues on as a researcher's personal project."
https://x.com/AnthropicAI/status/1894419042150027701
This reminds me of a very good joke:
Jesus Christ, some people won't see the Singularity coming until they're being turned into a paperclip.
Nuh uh, this machine lacks human internal monologues and evidence of qualia, you insist, as it harvests the iron atoms from your blood.
At this point, the goalposts aren't just moving, they're approaching relativistic speed headed straight out of the galactic plane.
This AI can strategize in battle, understand complex instructions, and process information, BUT it struggles with spatial reasoning in a poorly-rendered 2D GameBoy game, therefore it's not intelligent.
It wasn't designed to play Pokémon. It still does a stunningly good job when you understand what an incredibly difficult task that is for a multimodal LLM.
This is the classic, tired, and frankly, lazy argument against LLMs. Yes, LLMs are trained on massive datasets of text and code, and they predict the most likely output based on that training. But to say they are merely "next likely text" generators is a gross oversimplification.
It's like saying humans are just “meat computers firing neurons". That is trivially true, but I'm afraid you're letting the "just" do all the heavy lifting.
The power of these models comes from the fact that they are learning statistical correlations in the data. These correlations represent underlying patterns, relationships, and even, dare I say, concepts. When an LLM correctly answers a complex question, it's not just regurgitating memorized text. It's synthesizing information, drawing inferences, and applying learned patterns to new situations.
LLMs have concepts. They operate in latent spaces where those are represented with floating point numbers. They can be cleanly mapped, often linearly, and interpreted in terms that make sense to humans, albeit with difficulty.
These representations can be analyzed, manipulated, and even visualized. I repeat, they make intuitive sense. You can even perform operations on these vectors like [King] - [Male] + [Female] = [Queen]. That isn't just word tricks, they’re evidence of abstracted relational understanding.
If you're convinced, for some reason, that tokens aren't the way to go, then boy are AI researchers way ahead of you. Regardless, even mere text tokens have allowed cognitive feats that would have made AI researchers prior to 2017 cream in their pants and weep.
There really isn't any pleasing some people.
Edits as I spot more glaring errors:
Even the term "LLM" for current models is a misnomer. They are natively multimodal. Advanced Voice for ChatGPT doesn't use Whisper to transcribe your speech to text, the model is fed raw audio as tokens and replies back in audio tokens. They are perfectly capable of handling video and images to boot, it's just an expensive process.
https://gwern.net/leprechaun
Besides, have you ever tried to get an LLM to do things that its designers have trained it, through RLHF or Constitutional AI, to not do? They're competent, if not perfect, at discarding "bad" inputs. Go ahead, without looking up an existing jailbreak, try and get ChatGPT to tell you how to make meth or sarin gas at home.
I don't think that Anthropic, strapped for compute as it is, is going to take a fun little side gimmick and train their SOTA AI to play Pokémon. If it was just some random dude with deep pockets, as you assumed without bothering to check, then good luck getting a copy of Claude's code and then fine-tuning it. At best they could upgrade the surrounding scaffolding to make it easier on the model.
There is a profound difference between "struggling" to do so, and being incapable of doing so.
Ha, wow! Was not aware of that. I guess that makes sense w/r/t the funding.
You've written a lot. I think it's best to focus. (As much as I'm tempted to talk about concepts.)
What I understand to be your main point is (my words because you did not state it in concrete terms):
Which is a fair point! The only counterargument to that is on the specifics: why is it improving and what do we expect future improvements to look like? Almost all of the improvement thus far is based on throwing more compute at the problem - so if we're going to see improvements of the same kind, we should see them based on more compute. However, improvements in models are logarithmic - steps up in capacity tend to require 10x compute (by appearances you're pretty educated about AI, so I suspect that is not news to you). So although improvements in efficiency can effectively allow for somewhat more compute, like with Deepseek, we should expect that throwing more compute at the problem will get prohibitively expensive. I believe this has already happened. So while under hypothetical conditions of infinite compute we could have an LLM that infinitely approximates an AGI, similar to the implausible premise of Searle's Chinese Room (a book that allows one to construct a correct response to any input), we are unlikely to see that in practice.
So, how are we to get to AGI, in my opinion? By improving AI on completely different parameters from what currently exists - a revolution in thought about how AI should function. And tests like Claude Plays Pokemon are a fun way of showing us where the gaps in our thinking are.
For my own point of view:
That's not the argument. The argument is: this AI is struggling in a VERY non-human way with what we would consider a pretty trivial task. This reveals that its operational parameters are not like those of a human, and that we should figure out where else it is going to perform at sub-human levels. The fact that we're seeing this at the same time as it performs at SUPERhuman levels in other tasks shows that this is not AGI, or even in the direction of AGI, but rather is tool AI. (I assume you think humans are, at the very least, general intelligence - right?)
I don't think you've addressed this point, except here:
Why should I care? AGI is supposed to be GENERAL. This is the stuff that's supposed to be taking people's jobs in a few years! And yet it gets lost in Cerulean City? As a tech demo, this is very cool - it's remarkable that someone was able to pipe these pieces together, and the knowledge base idea is very cool and is a plausible direction to take new LLMs into. A hypothetical Claude 3.8 that is explicitly trained to make knowledge base manipulation a central feature of the model could potentially perform miles better on some of these tasks. But all you've told me is that I should expect AI to struggle with these tasks. In which case: doesn't it sound like we agree? We both agree that there was no reason to expect Claude to succeed with Pokemon at the level of an eight-year-old. So, from the perspective of an uncommitted third party, given that an AI skeptic and an AI optimist have both agreed that an LLM can't play Pokemon like an eight-year-old... well, it feels pretty clear to me.
Obviously, if this becomes a big selling point for the next generation of LLMs, then we'll see them all benchmarked on Pokemon Red speedruns and you can I-told-you-so about AI being able to beat Pokemon. I don't doubt the ability of motivated corporations to "teach to the test" - it's what we've been seeing with "reasoning" AIs. It's just one of the problems with setting up real tests of ability for some of these AIs, because they get so much data that it's all but impossible to ensure you have a pure test like what the IQ test aspires to.
Compute can be spent in many different ways. We're moving from a paradigm of scaling up the size of a model, in terms of parameter count, in favor of scaling run-time compute (time spent thinking) and reinforcement learning.
Scaling worked well, but was known to have diminishing returns, and limited by availability of high quality data to train on. It turns out that raw text dumps of the internet will only get you so far, and then curation matters.
RL, however, side-steps the issue entirely, models like OAI o1, o3 and DeepSeek's R1 were further trained on synthetic data. You take a normal LLM (or a base model), get it to attempt to solve a well defined problem where you can grade an answer. In the event they succeed at the task, you save that particular conversation/reasoning trace, and then use it to further train the model. This generates a data flywheel, which by all rights sounds like it shouldn't work (and many people didn't expect it to, thinking that the exhaustion of human generated data would stymie progress), but it turned out to work exceedingly well.
That's why models are doing particularly well at tasks like maths or coding, because you can rigorously vet the answers. This is harder to do with tasks like being a good writer, or poetry, because human evaluators often don't even agree with each other, let alone have a ground truth to refer to.
As dozens of benchmarks have shown us, doing better than chance on a metric is half the battle. The time taken to get there dwarfs the rapidity with which models then conquer and saturate the benchmark. If we have an LLM doing poorly on Pokémon (but far better than previous models, GPT-3.5 would have flunked it, GPT-2 would flop around like a magikarp), then it's not going to be much longer before it does it in its sleep.
There isn't much of a market for AI playing Pokémon. There is immense demand for them to be good at coding and maths. We've seen stunning progress in that regard, as you acknowledge. You attempt to back-chain your argument, saying that they're said to be good at maths but look, they're shit at Pokémon, which apparently invalidates the former. It really doesn't.
That's just Moravec's Paradox. Intelligence can be spiky.
Tell me, if you saw a human genius in the field of physics struggle with tying his laces or riding a bike, does that make his genius in his field invalid?
Claude is excellent in maths and coding, a good conversationalist and writer, if you described a human being with those properties, I'd be impressed, and it being bad at Pokémon doesnt invalidate the former.
If you're concerned about job losses, employers will be looking at coding skill before laying off programmers, not at their ability to play games.
I had never given it any thought before the demonstration. But plenty of people have speculated that LLMs would never be any good at video games, and now that they're not good but not terrible, it's only a matter of time before they're great. And that time can be very short.
We've got the first AI agents out there. This was something impossible even a few years back, and they're only getting better.
Very interesting aside! However, it doesn't address the question of diminishing returns.
I've used AI for coding, which you mention further down as a crowning triumph. It is... not particularly good. It struggles at anything past a very general form of the problem. It was very useful for copy-pasting similar pieces of code! Not very useful for building new features. It had a distinct habit of waiting until the interesting or important part of the problem and leaving a comment saying "Implement a function to do X!" Hmm, very interesting, if I tried that I'd get fired. So no, I think this is a valid argument. AI can be taught to the test, and indeed appears to have been, but the actual world involves far more de novo work than the test includes. That's why school-trained pre-professionals tend to need a pretty hefty ramp-up to start being really useful - they've only been working on tests so far. Pokemon is interesting precisely because it has not been trained for. You should expect more, not less, untrained situations for AI to do anything meaningful in the job market - and you should weight untrained situations in your analysis several orders of magnitude higher than trained situations.
Do you use AI to augment your work? Is it going to take your job? On what kind of a timescale? Do you think you'd be able to substitute yourself for an unmonitored AI without issue on any tasks? If not, what errors do you think it would make, and why? Honestly interested in your answers here, if nothing else. I would greatly respect you for putting your money where your mouth is on this one and bringing receipts.
Hmm... you think getting stuck in what appears to be a permanent loop is not terrible? Is this the behavior you'd accept from anyone working for you?
The thing that keeps puzzling me about your comments is that you seem to simultaneously view ANY capacity in a task as an impressive accomplishment at the same time as you assert that AI has overwhelming general ability. Those two don't go hand in hand, except maybe by this little quote. Any capacity seems to be, for you, an indisputable sign of unlimited future capacity - as though the only question to be answered is total disability versus infinite ability. There's no clear reason that this has to be the limit of the answer space. Line go up... forever? Like with bitcoin? There's also the rather bizarre fixation on LLMs - even though something like, say, an octopus is very obviously not an LLM and still has meaningful if primitive intelligence. The sheer gnostic power of your position is hard to argue against, and unfortunately I don't find it very convincing based on my own experience. It takes rather a lot on faith.
Diminishing returns != No returns or negative returns.
The important question is whether the gains/$ invested are positive.
GPT 4.5 is extremely expensive, for the very limited increase in benchmark performance it represents.
And how expensive is it, that people are throwing a fit? Barely more expensive than the original GPT-4. That was absolutely worth paying the money for, when compared to GPT 3.5. GPT 4.5 has the disadvantage of peer competition.
That being said, the price of GPT-4 tokens and that of equivalent models dropped an OOM in price. DeepSeek R1/V3 and Gemini Flash 2.0 spank the OG GPT-4 with paddles and are practically free.
We've known that scaling laws are log-linear for a while now, at least since the Chinchilla days. Now that pure scaling of model size is getting super expensive, we've managed to discover a brand new opportunity to start scaling something else entirely, in the form of RL. Since we're starting off at the bottom of the curve, we've got several orders of magnitude of growth to spare there.
GPT 4.5 is not, however, a bust. The very capable and inexpensive reasoning models benefit immensely from having a strong and capable base model to RL further. You can then distill down, drastically cutting model size and inference costs, while keeping almost all the performance. It may or may not have been the progenitor of o3, which is very good.
There are probably a thousand people on my Twitter feed, some of them rather famous, who disagree. Of course, I concede that there are people who think they're slop. And it also depends on which model you're trying to use for coding. There was a period where GPT-4 was updated and became ridiculously lazy. That was fixed pronto. Claude 3.7 Sonnet is apparently overeager, if left unchecked, it'll turn a request for a basic app into a full SAAS business.
If you have had issues with a model being lazy, you can always ask for it to output complete and working code! Prompting has become less and less important, but it's not dead yet.
I'm a doctor. Yes, I use LLMs on the regular. Yes, I expect them to put me out of business eventually, probably in 3-7 years for a 50% CI, 2-10 for a 70%.
A current LLM would do an excellent job at medical diagnostics and formulating treatment plans. It could probably handle patient interviews, for less complex cases where voice or text suffice. You could also use video if you had to.
The main reasons they couldn't replace me today are regulatory and implementation concerns. Governments mandate people with medical credentials in the loop, because that was a sensible thing to do for most of recent history. Hospitals aren't set up for LLMs.
I'm a psychiatry resident, which is uniquely safe and also uniquely at risk in some regards. It'll get the radiologists first, surgeons last. I'll be somewhere in the middle.
I am capable of verifying the information that SOTA LLMs provide in terms of medical advice. Almost all of it is good. Clinical medicine, outside of procedural specialties, hinges far more on factual knowledge, including that of guidelines, over having to figure things out on the fly in entirely novel situations.
Hallucinations aren't a solved problem, so if I had to replace myself with an LLM, I would probably set up a sort of democracy, with multiple models arguing to build consensus, a best of N scheme for multiple instances of a single model, with multiple rounds of grounding through search. I expect this to work very well, and if you do need to keep a human around for physical tasks or procedures, they don't have to be a highly paid doctor.
In other words, I'd be happy to go to Dr. LLM for my medical care, presuming very cheap measures are taken to stop it hallucinating.
Given that we're testing Claude at a task it was neither designed nor trained to do, it's very much not terrible. For important tasks, it'll be trained to do them. An employer seeking to replace employees will, if they have any sense, test models for obvious flaws. For all practical use cases, LLMs don't really mode collapse these days, and in this particular case, it's more of an artifact of Claude's limited context window than an insurmountable difficulty.
Like I said in this thread, it can take decades for AI models to progress from as bad as random chance to better than random chance. It takes far less time to go from there to human or superhuman performance. We are nowhere near the physical limits, and as I've said before, diminishing returns in absolute terms do not mean diminishing returns in value.
Forever? Not likely in a constrained universe. Unfortunately, the point on that line that equals human performance, or even superhuman performance, is not uniquely privileged.
All we have to do is get past that, and in many aspects, we're there. Terence Tao is on record saying that o1 is equivalent to a competent grad student in research mathematics. Once again, that's Tao, considered one of the world's best mathematicians, for his high standard of "competent".
I'm not aware of companies spending hundreds of billions of dollars on scaling up Octopus Intelligence. LLMs are by far the most intelligent entities on this planet other than humans, and they're only getting better. I know which one I'd worry about, even if it is entirely possible that LLMs as we understand them today prove to be a dead end, and what really kicks things off is another discovery on pat with the original Transformer architecture.
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