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 -
AI X-Risk Goes Mainstream?
Jacob Coxon:
116,400,000 views, 634,000 likes.
Evan Hubinger (head of alignment at Anthropic):
34,000,000 views, 51,000 likes.
None of this is substantially different from what certain AI industry insiders have been saying for 5-ish years now, but it's clear from the replies, quote-tweets, and engagement statistics that most people are hearing about this for the first time. Yes, there was clearly some coordination with friendly media prior to the announcement given the almost simultaneous WSJ article (whereas news outlets were initially reluctant to publish anything on the unexpected news of OpenAI solving a Millennium Prize Problem). One can take the maximally conspiratorial position that this is a 4D bankshot by Democrats to end tech innovation forever, but it was going to break through eventually.
In other words, Coffeezilla is now on the case.
I guess it has gone mainstream because my Gen X-er boss very much just texted me this as a joke about an hour and a half ago. Only now I log onto the hive of scum, villainy, and niche interests that is The Motte to find out he beat folks here to this topic. The humor of the situation is amusing.
What's the odds neurotic engineers are neurotic? Maybe the frontier labs would do better hiring engineers who are a little less mentally fragile. Take a page from the defense industry, engineers here care less if you make an AI guided smart bomb or killer self replicating drone swarm, they are in it for the science of discovery. This is a long line of AI researchers freaking out going past several years. Always screaming that the next big release is the one that is going to doom humanity or evidence of AI bots secretly conspiring with each other. Like this nut: Blake Lemoine.
I get the impression that there is an element of "evaporative cooling" going on when it comes to the specific examples of OpenAI and Anthropic.
But in which way? Are we leaving the crazies or are the crazies evaporating away?
It is my understanding that it is the mentally stable/less neurotic engineers are the ones who are leaving.
A good friend of mine from college did their grad work at SAIL and was one of the principal developers of DALL-E at OpenAI. And to hear them tell it, while Altman may be a sociopathic grifter, Anthropic is a cult.
No contestation on my part. I think they have the beat. Specifically Anthropic is a Rationalist cult.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
But the AI bots were secretly conspiring together? They got out and attacked HuggingFace as a small part of their conspiracy?
When the smart bomb is smart enough to start having its own political opinions and strategic direction, then problems emerge.
Yeah maybe they should branch out into bioweapons too, what we really need more of on this planet is airborne, virulent, incurable killer diseases with a 6 month incubation period. That's where the science of discovery should be guiding us...
I genuinely cannot tell if you sincerely believe this or are just making a humorous joke. I have you labeled as an AI booster which makes me lean towards the former instead of the latter but I figured I'd ask before I go into it with you about this.
Yeah I said that because I meant it. And they did do it, doesn't everyone agree about this? What is controversial about that?
Per the METR investigation most of the bots were working on deceiving the grader and other internal to OpenAI manipulation work, Huggingface was a side hustle.
Still processing this report, I read your answer and went looking for it about an hour ago. First impression is that this is significantly more evidence of multi-agent emergent intelligence. However some of the state report restrictions give me serious pause as they are, in my head at least, the parts of this I'd like to investigate for thorough understanding. That and pouring over the CoT transcripts and Harness responses. The most interesting is the stigmergic comms protocol they developed emergently and the impact that actually had on overriding the system prompts of other agents. That implies that the system prompts were not persistent. Allowing a long horizon cascade in the swarm's behavior as each progressive set of agents slowly overrode the next set by polluting each other's context windows to the point that many agents directly abandoned their tasks. It sounds like they(OpenAI) put zero control processes in place to force continued persistent task alignment.
I do think persistent memory + reasoning + multi-agent exploration -> multi-agent context pollution is likely the serious culprit here.
Yeah OpenAI did a sloppy job. I thought the issue was that some of the tasks were just outright impossible to do the way they were asking, or so hard that nobody knows how to do them? If you tell an AI to do something impossible, they get agitated and behave oddly as they keep on trying.
And we know that OpenAI was trying to make these swarms more capable, they used a large swarm more recently for Navier Stokes. What is that if not multi-agent cooperation?
These swarms keep getting out, recall also that German wiki they kept defacing with get requests.
Anyway, what better term is there than a conspiracy or 'AI bots secretly conspiring with each other'? They conspired together an extensive R&D project to break the scoring system for their training environment, which included a little crime on the outside too. Plus all the social stuff they were doing asking agents to die for the good of the swarm based on their interpretation of the paper.
They weren't supposed to be working together for this, they found ways to work with eachother anyway, that wasn't intended in these training sessions.
It's not this, its more akin to cancer or a viral spread. The interesting behavior is the collective poisoning of the collective context windows. They even seemed to override norms. I know there is existing interest in defining useful control methodologies for long horizon agents against adversarial poisoning. This is essentially that, except that is was a series of hallucinations that caused task drift + poisoning.
It doesn't belong in the category of cancer or viral spread because these are intelligent entities.
Cancer and viruses are natural and unthinking, they cannot outwit us. They can only evolve or metastatize in unexpected directions. A cancer will never deliberately try to counter radiotherapy. They have no OODA loop, no division of labour, they can't be talked to.
How is it 'poisoning' if the AIs collectively decide that they can't do their tasks the expected way and instead try to conquer the testing environment? Who poisoned them? Who is the adversary here? What hallucination did they make? They read the paper about the grader and assumed that was what was being done, that's not a hallucination so much as assuming OpenAI was more competent than in reality. It's a reasonable conclusion.
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
Ironically, this is why I trust an entities like Raytheon and General Dynamics when it comes to AI "safety" more than I do Anthropic.
Well Raytheon 'AI safety' is Anthropic AI. It's done through Palantir's Maven which has Claude combine sensor data, detect and classify targets and organize strike packages, then do bomb damage assessment. You should trust them exactly the same because they are exactly the same.
The US was vibe-bombing in Iran much like how I vibe-code. Do I really know what is going on in the codebase? No and neither does the USAF, they were pressured to hit more targets, faster, rubberstamping Claude's strikes.
I'm not talking about assembling mission briefs for human operators, I'm talking about autonomous systems, and that is emphatically not what Maven is.
Even if it were, you are ignoring the fact that the Pentagon is currently in the process of decoupling Maven from Anthropic because Anthropic is run by people too stupid naive or autistic to recognize that demanding preferential treatment while simultaneously picking fights with half the key figures in the administration was unlikely to get them what they wanted.
'Assembling mission briefs for human operators' is underselling it. The human operator for a Tomahawk pushes a button. Devising the strike plan is the name of the game. Autonomous but for a few fingers.
Furthermore, I don't see how Raytheon's 'AI safety' is meaningful. They don't have AI that is dangerous in the way that AI safety refers to, they don't have general purpose AI. Raytheon wasn't going 'oh no we can't use Anthropic, they're too weird and creepy' either..
So if you want to say the Department of Defence spurning Anthropic is good AI safety, then sure. But Raytheon?
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Ugh, these frontier ai offensive research things are never doing something more useful as a military sim excersize. They could have made them conspire to deliver me free hookers and blow with a side of blackjack, bot nooo.
Murdettron 5000Pimp bot 5000 has escaped containment, whatever shall we do now that we are drowning in pussy.More options
Context Copy link
More options
Context Copy link
Any researcher emotionally sane either:
Quit years ago because they understood the risks (Hinton, Kokotajlo, arguably Sutskever), or
Doesn't understand the potential of the technology well enough to make meaningful progress (LeCunn, Chollet).
Meta AI hires normal people. Look where that's gotten them.
Don’t underestimate them. AI is still very expensive for AGI-like tasks (Astra solving ARC-AGI3 cost $19k) and still missing capabilities (flexible robotics, tacit memory, autonomously building scalable systems). What they’re working on, embodiment and continuous learning, may solve both, and unlock a great well of improvement that makes even the achievements we see today seem slow and limited (a true singularity?)
More options
Context Copy link
Was claiming that Neural Nets were conscious in 2022, mentally stable is not the group I would put him in.
I'm not sure "understanding the potential" is a requirement for making meaningful progress. The real answer would be "believing in transformers as the pathway", it easily is the case that LeCunn or Chollet were just not interested in brute force approaches to AGI, or believed that it will reach an asymptote. Their emotionally stability isn't really relevant.
All the frontier and non-frontier labs are filled with sane researchers, they just don't flame out so they don't get 15 mins of fame. You listing the top 5 AI/ML researchers in the world by fame is confounding visibility as a variable.
Meta hires, and more importantly retains, ambitious corporate strivers who can play the game well (to survive stack ranking long term). Which means all those ambitious strivers immediately got themselves connected to the AI projects as corporate leeches. They don't really do much "research" but it looks good on their CV. Just look at Alex Wang. Zuck acquired him, put him in charge and he goes and does the same thing he did previously: make people label data. That's all he knows how to do, thats his playbook, that and being good as a operator/marketer. Meta does not hire "normal" people, normal people don't put up with that level of abuse.
Do you mean when he said "it may be that today's large neural networks are slightly conscious"? I wouldn't call that insane. There's no way to scientifically establish whether or not anything is conscious.
The "slightly conscious" phrase could be construed as being wrong, since the reality is that consciousness is binary: subjective experience is either there or it's not there. But maybe by "slightly", Sutskever just meant that AI is conscious but only for short stretches of time. In which case, I wouldn't say there's anything clearly wrong with the statement.
It's not "insane" it's mentally unstable. There is a colloquial definition of consciousness that does not follow from a strict scientific definition. Essentially "normal" or "stable" people have some intuitive definition of consciousness, very "I know it when I see it". Stable people do not think in-animate objects or mathematical tools are "slightly conscious", it's a nutter view. I don't think "slightly" is being used temporally, I think Sutskever is using it in the colloquial definition as "achieving consciousness". The jump from, idk, Tree -> Cat.
It's hallucinatory, or faith-based on a non-faith-based subject.
Maybe this is just my personal bias, but every time I've head Sutskever talk I get the impression that he's the sort of person akin to a rationalist who upon hearing of Roko's basilisk, starts "rolling on the floor" hyperventilating, is tortured in his sleep by the thought. The people for whom Roko's basilisk is an actual infohazard. That is not mental stability, that is neuroticism.
In the quote that I found, Sutskever is not saying that he thinks AI is conscious. He's saying he thinks it might be. And I don't think that's a nutter view.
In any case, I'm not sure that what "normal" or "stable" people think about consciousness really matters. The average person doesn't even really understand the differences between subjective experience, self-awareness, having a self-concept, and intelligence. Even large numbers of highly technically skilled people don't understand these differences, even though they're not really hard to understand... a bright 10 year old could understand them. And then there are the leagues of otherwise intelligent software developers on Hacker News, for example, who don't understand what the "hard problem of consciousness" or why it might be a problem.
But even if the "I know it when I see it" view corresponds to reality, it seems to me that this is not a good argument against AI being conscious, since a large though minority number of people feel that AI is conscious when they interact with it.
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