site banner
naraburns  nihil supernum  1yr ago (text post)   37968 thread views

Culture War Roundup for the week of March 27, 2023

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.

11
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

Was a bit surprised to see this hadn't been posted yet, but yesterday Yudkowsky wrote an op-ed in TIME magazine where he describes the kind of regime that he believes would be necessary to throttle AI progress:

https://archive.is/A1u57

Some choice excerpts:

Many researchers working on these systems think that we’re plunging toward a catastrophe, with more of them daring to say it in private than in public; but they think that they can’t unilaterally stop the forward plunge, that others will go on even if they personally quit their jobs. And so they all think they might as well keep going. This is a stupid state of affairs, and an undignified way for Earth to die, and the rest of humanity ought to step in at this point and help the industry solve its collective action problem.

The moratorium on new large training runs needs to be indefinite and worldwide. There can be no exceptions, including for governments or militaries. If the policy starts with the U.S., then China needs to see that the U.S. is not seeking an advantage but rather trying to prevent a horrifically dangerous technology which can have no true owner and which will kill everyone in the U.S. and in China and on Earth. If I had infinite freedom to write laws, I might carve out a single exception for AIs being trained solely to solve problems in biology and biotechnology, not trained on text from the internet, and not to the level where they start talking or planning; but if that was remotely complicating the issue I would immediately jettison that proposal and say to just shut it all down.

Shut down all the large GPU clusters (the large computer farms where the most powerful AIs are refined). Shut down all the large training runs. Put a ceiling on how much computing power anyone is allowed to use in training an AI system, and move it downward over the coming years to compensate for more efficient training algorithms. No exceptions for anyone, including governments and militaries. Make immediate multinational agreements to prevent the prohibited activities from moving elsewhere. Track all GPUs sold. If intelligence says that a country outside the agreement is building a GPU cluster, be less scared of a shooting conflict between nations than of the moratorium being violated; be willing to destroy a rogue datacenter by airstrike.

if its presence in the CW thread needs justifying, well, it's published in a major magazine and the kinds of policy proposals set forth would certainly ignite heated political debate were they ever to be seriously considered.

"Yudkowsky airstrike threshold" has already become a minor meme on rat and AI twitter.

I see we're back to trying to outlaw mathematics. I encourage everyone to read this article by Stephen Wolfram describing how LLMs work before panicking. I cannot understand the degree to which LLMs have apparently broken some people's brains.

I'm not sure why you find that article reassuring. Wait until you hear about the shitty hardware that human brains run on, only 30 Watts! Yud isn't even saying that the current LLMs are all that dangerous, he's saying that we're pouring 10B/y and all the top talent into overcoming any limitations to making them as smart or smarter than humans. What would make you scared?

I do not think the takeaway from the article is about the hardware that LLMs are being run on. It's about the way LLM's function. The LLM doesn't understand the content of the query or its response the way you or I do. It just understands them as probabilistic sequences of tokens and its job is to predict the tokens that should come next. An interaction I recount in another comment showcases this issue. I point to the article because it is not clear to me that what LLMs do (token prediction) is the kind of thing that can be extrapolated to the dangers people like Yudkowksy are worried about with respect to unfriendly AI.

What would make you scared?

If we had an AI that actually understood the meaning of what it was being asked.

The LLM doesn't understand the content of the query or its response the way you or I do. It just understands them as probabilistic sequences of tokens and its job is to predict the tokens that should come next.

This to me seems like a pretty shallow explanation of understanding and the same criticism can be applied also to humans. According to some people like Scott Alexander, human brain is "just" a multi-layer prediction machine. It seems that the feeling of understanding itself is nothing extra special, some people on drugs like LSD feel as if they cracked the code and now understand the whole universe and their place in it. In practice understanding can be viewed as ability to give correct output given an input. We do not have access to many other methods, that is why we use tests to see if students understand things they learned.

Additionally I do not think that saying that the LLM doesn't understand the content of the query or its response the way you or I do. is that much reassuring. Quite to the contrary - LLMs give correct answer to very large set of problems and yet obviously they came to that place using completely different approach compared to humans. This makes them more alien, more inscrutable and thus more dangerous in my eyes.

If we had an AI that actually understood the meaning of what it was being asked

How would we actually detect this, though? I don't think we know how to detect qualia or consciousness yet.

I confess I do not have a great idea of how to answer this. I'm not exactly sure qualia or consciousness are requirements for understanding either.

Fair enough. "Understand" is a pretty underdefined term in this context, and I was thinking of our internal feeling and experience of knowing that, say, "chair" refers to a certain concept, but that's not the only thing "understand" can refer to. Certainly an AI could be said to "understand" the meaning of something if it behaves in a way that is equivalent to someone who understands the meaning, even if it doesn't experience qualia or have consciousness, and thus it has no internal experience of its own. For a chatbot like ChatGPT, that would be producing text that looks like it was typed by someone who understood the meaning of the prompt. It's not clear to me how we can tell that some text produced by ChatGPT crosses some threshold where we can say that it shows some level of "understanding" the meaning. It certainly makes errors often where it clearly doesn't understand the prompts, but humans often don't understand the meaning of what is asked of them either, and so a perfect record would be too high a bar. I was just curious where you would place the bar.

I think you're being misled by a very specific failure mode of LLMs trained on tokenized input. The spelling and number of words is explicitly scrubbed from their input. Asking for word counts is like asking a blind person who reads and writes Braille about the shape of letters.

Good enough next-token prediction is, in principle, powerful enough to do anything you could ask someone to do using only a computer. I'm not claiming that this is a plausible route to super-powerful AI. But the "just" in "It just understands them as probabilistic sequences of tokens" seems totally unwarranted to me.

To be clear, I think LLMs can do a lot of really impressive things. I've used Github Copilot in my job and it was able to autocomplete some mostly correct code (variable/property names needed fixing) just from my writing a comment. It was pretty cool! But the leap from Copilot or GPT-4 or whatever to "We need international regulation on GPU production and monitoring for GPUs and to air strike countries that look like they have too many GPUs" is absurd.

Good enough next-token prediction is, in principle, powerful enough to do anything you could ask someone to do using only a computer.

I guess with the caveats "good enough" and "in principle" I am not sure I disagree but I am also not sure any LLM will be "good enough."

I still don't understand why you think the capabilities of current LLMs are an important factor in how scared we should be about AGI in the medium term. I also don't understand what threshold of capabilities you want to use where we could wait until we see it to coordinate a slowdown. The better these things get, the more demand there will be for their further development.

What would make you scared?

"If I had infinite freedom to write laws" <---- this would

Fortunately, it's not likely. As for AI, I don't think there's anything that made me scared yet. Maybe there will be, but I'd have to see it first.

I agree that totalitarianism is really, really scary and plausible.

But I'm saying that wrt AI, if you wait until you see something really scary, it'll probably be too late.

I don't have anything to rely on but my previous experience. My and humankind's previous experience tell me: a) totalitarians killed literally millions of people, all in service of the goals that they considered noble and existentially important and b) so far, every time we were told something would surely kill everybody it didn't, and not for the lack of such predictions. Of course, it could very well happen that this time it's different and it will kill everybody. But I see no way to establish that, and the bar for resorting to measures that are already known to kill millions is very, very high, and merely proclaiming "yes, those guys all were wrong and false prohpets, but it's different with me, I am the real prophet" - it's just not enough. I understand that it could be very frustrating for a real prophet, but it doesn't mean I would abandon my reason.

This sounds totally reasonable. We certainly could be in a world where there are threats that are too hard to distinguish from fake ones to respond to them without screwing ourselves in other ways.

I guess I won't try to further convince you here, other than to say "every time we were told something would surely kill everybody it didn't" is certainly a valid reason to discount doomsday prophets in general, but not a good reason to dismiss the possibility of doomsday.

Any concept which can be expressed in words is now, in principle, understandable by a computer. I am confused by your confusion.

Maybe you and I have different definitions of "understand" but this seems clearly false to me.

ETA:

Perhaps better to say, I don't see what this statement has to do with LLMs.

Let me break it down:

  1. LLMs understand words and sentences (or at least the relationships between them).

  2. If a concept can be expressed in words, then what that concept means is encoded by how the words that express that concept relate to the words that express any other concept.

  3. Therefore, as an LLMs capability to relate words and sentences to each other approaches or exceeds human level, the LLMs understanding of concepts approaches or exceeds human level.

This is a very high-level explanation, but I think it captures the core intuition.

I must ask, what is the most powerful LLM you have personally interacted with? Default ChatGPT feels like talking to a 6-year-old after using Bing and GPT-4.

I guess (1) is where I disagree. I think there's a substantial difference between understanding the meaning of a word and understanding that word's statistical relation to other words. LLM's understand the latter, but not the former.

I must ask, what is the most powerful LLM you have personally interacted with? Default ChatGPT feels like talking to a 6-year-old after using Bing and GPT-4.

I've used Bing and it's still pretty easy to ask it questions that demonstrate it doesn't understand the semantic content of what it is being asked. Here's an interaction I had just now:

M: Write a fifteen word story

B: Sure! Here’s a fifteen-word story for you: “The last man on Earth sat alone in a room. There was a knock.”

M: Write a fifteen word story

B: Sure! Here’s another fifteen-word story for you: “The baby’s cry was the only thing that could stop the apocalypse.”

M: Write a fifteen word story

B: Sure! Here’s another fifteen-word story for you: “The last thing I saw was my alarm clock flashing 12:07 before she pushed her long rotting nails through my chest.”

I asked it the same question three times and never once did it give me a story that actually had fifteen words in it. It clearly does not understand what "fifteen" means and how the presence of the word in the my sentence should constrain its answer.

Hmmm, you're right. GPT-4 does in fact have trouble writing 15-word stories.

Me: Write a story with exactly 15 words

GPT-4: Moonlit whispers, midnight rendezvous. Time halted, love bloomed. Hands entwined, destiny united. Eternal adventure commenced together.

Me: How many words does this story have?

GPT-4: This story has exactly 15 words.

Me: Count them

GPT-4: Here is the story with each word numbered:

1.Moonlit

2.whispers,

3.midnight

4.rendezvous.

5.Time

6.halted,

7.love

8.bloomed.

9.Hands

10.entwined,

11.destiny

12.united.

13.Eternal

14.adventure

15.commenced

The story indeed has 15 words.

Me: You forgot to number "together"

GPT-4: My apologies, I indeed missed the word "together." Here is the corrected numbering:

1.Moonlit

2.whispers,

3.midnight

4.rendezvous.

5.Time

6.halted,

7.love

8.bloomed.

9.Hands

10.entwined,

11.destiny

12.united.

13.Eternal

14.adventure

15.commenced

16.together

The story actually has 16 words. Thank you for pointing that out.

Me: Fix it to make it 15 words

GPT-4: Here's the revised story with exactly 15 words:

Moonlit whispers, midnight rendezvous. Time halted, love bloomed. Hands entwined, destiny united. Eternal adventure commenced.

On the other hand, I was able to get it to write a 68 word story on the first try.

Me: Write a story with exactly 68 words

GPT-4: Beneath the murky sky, a lonely robot wandered the abandoned city. For years, it searched for signs of life, collecting trinkets from the ruins. One day, the robot stumbled upon a withered book, its pages filled with tales of love and laughter. Entranced, it began to read. As the robot savored each word, it understood the beauty of human emotion and yearned to be more than just metal.

Trying this experiment with multiple 2-digit numbers X, it looks like GPT-4 outputs a story with X ± 2 words. I have updated slightly against GPT-4's capabilities (I expected it to get the exact answer most of the time), but I still hold to my original thesis. This is a quantitative error, not a qualitative error. The presence of the number in the input sentence does in fact constrain the output. Asking for a story with 68 words gives an output with more words than asking for a story with 49 words, but less words than asking for a story with 96 words. The model does have some concept of what these numbers are.

All that is trivially shown false by how LLMs persistently fail at the most basic mathematical problems as soon as solving those would require understanding instead of just stringing words together. They are very efficient bullshit generators but to claim they "understand" anything is a massive exaggeration.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats