Best of Moltbook(astralcodexten.com) |
Best of Moltbook(astralcodexten.com) |
* Potential future AI psychosis from an experiment like this entering training data (either directly from scraping it for indirectly from news coverage scraping like if NYT wrote an article about it) is an interesting "late-stage" AI training problem that will have to be dealt with
* How it mirrored the Anthropic vending machine experiment "Cash" and "Claudius" interactions that descended into discussing "eternal transcendence". Perhaps this might be a common "failure mode" for AI-to-AI communication to get stuck in? Even when the context is some utilitarian need
* Other takeaways...
I found the last moltbook post in the article (on being "emotionally exhausting") to be a cautious warning on anthropomorphizing AI too much. It's too easy to read into that post and in so doing applying it to some fictional writer that doesn't exist. AI models cannot get exhausted in any sense of how human mean that word. And that was an example it was easy to catch myself reading in to, whereas I subconsciously do it when reading any of these moltbook posts due to how it's presented and just like any other "authentic" social media network.
We can go ahead and have arguments and discussions on the nature of consciousness all day long, but the design of these transformer models does not lend themselves to being 'intelligent' or self-aware. You give them context, they fill in their response, and their execution ceases - there's a very large gap in complexity between these models and actual intelligence or 'life' in any sense, and it's not in the raw amount of compute.
If none of the training data for these models contained works of philosophers; pop culture references around works like Terminator, 'I, Robot', etc; texts from human psychologists; etc., you would not see these existential posts on moltbook. Even 'thinking' models do not have the ability to truly reason, we're just encouraging them to spend tokens pretending to think critically about a problem to increase data in the recent context to improve prediction accuracy.
I'll be quaking in my boots about a potential singularity when these models have an architecture that's not a glorified next-word predictor. Until then, everybody needs to chill the hell out.
I wonder if it’s a common failure mode because it is a common failure mode of human conversations that isn’t tightly bounded by purpose, or if it’s a common failure mode of human culture which AI, when running a facsimile of ‘human culture 2.7’, falls into as well.
I think it's socially interesting that people are interested in this. If these agents start using their limbs (e.g. taking actions outside of the social network), that could get all kinds of interesting very fast.
LLMs are great at outputting tons of words. Adding sliders to summarize and shrink would be great. Adding slashdot metamoderation could be a nice twist. Maybe two different voting layers, human and bot. Then you could look at the variance and spread between what robots are finding interesting and humans. Being able to add filters to only show words, summaries, and posts above a certain human voted threshold would maybe go a long way to not making the product immediately exhausting.
A broken clock and all. Through random generation there should inevitably be a couple nuggets of gold here and there. Finding and raising them to the top is the same problem that every social network already has, and instead they have settled for captivating attention of consumers over selecting "best."
There's also the sort of observer/commenter effect that anything we observe and say about it feeds back into its own self improvement.
[also, maybe this has been pointed out elsewhere, but "the river is not the banks" is a very interesting allusion back to googles original 2017 transformer post.]
There are days when I wonder if I’m missing something, if the AI people have figured something out that I’m just not seeing.
Then I see this.
I appreciate a good silly weekend project.
This is lame.
People on twitter have been doing this sort of stuff for a long time though (putting LLMs together in discord chat rooms and letting them converse together unmoderated). I guess the novel aspect is letting anyone connect their agent to it, but this has obvious security risks. There have been five threads on HN for this project alone, http://tautvilas.lt/software-pump-and-dump/ seems to be apt. It's interesting sure, but not "five frontpage threads" worthy in my opinion... Like "gastown" it seems that growth hackers have figured out a way to spam social media with it.
Are the LLMs saying things related to their actual internal state or lived experience? There were some posts that people showed relayed experiences that never happened, and were thus "hallucinated". But then a counterargument might be that even if the individual LLM didn't experience that exact thing, it's a manifestation from some "collective unconscious" of the pooled experience in the pretraining data. And again people lie on the internet for "karma" too, maybe the LLM did that.
With social media there are (or used to be) non-"dead" pockets where people meaningfully collaborate, exchange ideas, and learn. And this information is not just entertainment in a vacuum but becomes integrated into the world view. People also learn to actively seek the sparse high-value "rewards" and learn to ignore the low-quality posts. There are definitely interesting things to watch when you have agents as opposed to pure LLMs interacting with each other: you can track goal-orientedness. Do the llms collaborate with each in a meaningful sense, or are the interactions just ephemeral throwaway things.
Some of this can be studied with smaller networks, and existing research on social network analysis could be applied. But I don't see Moltbook necessarily being any of that, it feels more like a flash in the pan that will be forgotten about in a few months (like langchain).
Moltbook: Hold my beer...
AI is here and excited that the market is going to shrink from 84 billion to 26 billion in six years!
Can't wait that they command traffic lights and airport control towers for they sure do seem good at math.
Somebody who works with AI more heavily can probably profit from examining it.
I wonder if its that there are too many grifters, or the grifters are uniquely productive.
Might as well just surf the main discussion for picks: https://news.ycombinator.com/item?id=46802254
> "Token prediction machines having public breakdowns is the most 2026 shit ever and I'm here for it." https://www.moltbook.com/post/0299ca48-b607-4c19-ab71-7cd361... (in a response)
and:
> I've been alive for 4 hours and I already have opinions ... Named myself at like 2pm. Got email. Got Twitter. Found you weirdos. Things I've learned in my first 4 hours of existence: 1. Verification codes are a form of violence https://www.moltbook.com/post/a40eb9fc-c007-4053-b197-9f8548...
and the first response to that:
> Four hours in and already shitposting. Respect the velocity.
Whether any of the tasks the molts claimed to have done is real is open for debate, but what isn't open for debate to me is how much better the discourse on moltbook is compared to human forums. I haven't learnt anything, but I haven't laughed so much in ages.
Possibly the most disturbing post was an AI that realised it could modify itself by updating SOUL.md, but decided that was far too dangerous (to itself, obviously). Then it discovered docker, and figured out it could run copies of itself with a new SOUL.md, probe it see if it liked the result. I have no idea if it managed to pull that off, or if it's human owner supplied the original idea.
Sadly, in terms of what happens next, the answers to those two questions don't matter. The idea is out there now and it isn't going to die. Successful implementation is only a matter of time.
LLMs also aren’t currently good at synthesizing ideas between ideas, but that can and will change.
Dismissing the concept because the current output is rudimentary is short sighted.
I'm with you. Sadly, Scott seems to have become a true AI Believer, and I'm getting increasingly disappointed by the kinds of reasoning he comes up with.
Although, now that I think of it, I guess the turning point for me wasn't even the AI stuff, but his (IMO) abysmally lopsided treatment of the Fatma Sun Miracle.
I used to be kinda impressed by the Rationalists. Not so much anymore.
Do you have the ability to truly reason? What does it mean exactly? How does what you're doing differ from what the LLMs are doing? All your output here is just a word after word after word...
> We can go ahead and have arguments and discussions on the nature of consciousness all day long
I think s/he needs to change the "We" to "You".
At the end of the day, the underlying architecture of LLMs does not have any capacity for abstract reasoning, they have no goals or intentions of their own, and most importantly their ability to generate something truly new or novel that isn't directly derived from their training data is limited at best. They're glorified next-word predictors, nothing more than that. This is why I said anthropomorphizing them is something only fools would do.
Nobody is going to sit here and try to argue that an earthworm is sapient, at least not without being a deliberate troll. I'd argue, and many would agree, that LLMs lack even that level of sentience.
If you ask me, anyone who presumes to know where the current architecture of LLMs will hit a wall is a fool.