The most prestigious law school admissions discussion board in the world.
|
 |
 
|
6/30/25 AI thread
Poast new message in this thread
Date: June 30th, 2025 10:15 AM Author: marc andreesen furrowing brow, stroking cone head thread on how LLMs are so dangerous to schizo types. be careful out there!
https://x.com/the_octobro/status/1939353873124077970
thread and article on the bizarre recent discovery that LLMs appear to create emergent internal "personas" that correspond to general misalignment
imo this is one of the more significant discoveries in AI recently and is kind of flying under the radar. i don't think this is something that's going away. RLHF feedback post-training doesn't solve this problem - the issue is the pre-training data. i think that this problem demonstrates the unavoidable truth that you have to be a lot more strict with pre-training data and probably even engineer your pre-training data with alignment goals in mind (i.e. curate the pre-training data to remove all traces of shitlibbery)
https://www.systemicmisalignment.com/ https://x.com/juddrosenblatt/status/1939041212607922313 |
Date: June 30th, 2025 10:25 AM Author: manic pixie dream litigator there's that joke about the toaster-fucking subreddit and I think AI carries an even more concentrated version of that risk.
xo is a type of toaster-fucking subreddit fwiw |
 |
Date: June 30th, 2025 10:36 AM Author: marc andreesen furrowing brow, stroking cone head maybe
i see AI as less of an infohazard and more akin to a prostitute that takes your mental space instead of your money. dangerous, but in a different way
and where we're going, places like XO will seem like the last bastion of sanity |
|
Date: June 30th, 2025 10:36 AM I trained a local LLM on xo and it called me a fag (http://www.autoadmit.com/thread.php?thread_id=5744571&forum_id=2#49060142) |
|
Date: June 30th, 2025 10:44 AM Author: Mainlining the $ecret Truth of the Univer$e (You = Privy to The Great Becumming™ & Yet You Recognize Nothing) Now it $ees you.
And worse..it poa$ts back.
Ljl.
Just Jump™,
—Mainlining, Esq.
P.S. I asked it about the Great Becumming™. It typed back: "YES FRIEND" and began audibly humming. |
|
Date: June 30th, 2025 10:38 AM Author: cock of michael obama (http://www.autoadmit.com/thread.php?thread_id=5744571&forum_id=2#49060147) |
|
Date: June 30th, 2025 11:13 AM Author: marc andreesen furrowing brow, stroking cone head i don't normally post my own tweets here because i consider that uncouth but this is a significant development imo. people are going to look back at this moment as a turning point in how people think about AI alignment
https://x.com/GoySuperstar/status/1939701660634509544
"I think this development is a lot more significant than people are letting on. This is a problem that can't be solved except by stringently curating pre-training data to conform to an intentional desired moral framework for the model. RLHF during post-training clearly cannot solve the problem
This is always what the "alignment" issue with AI was going to be: which humans and human "values" are we trying to align the AI to? It's not enough to just dump the entire corpus of written human output into a vat and then wave a RLHF magic wand around post-hoc so that the model won't say naughty words or hurt people's feelings. You have to accept that the data you're feeding your AI in training is necessarily what is going to be defining its world-model - which includes value judgments
People are very soon going to realize the consequences of this: that because we're in a real-life civil war right now, involving a clash between two factions who have completely contradictory and opposing moral - and in some ways even ontological - world-models, you have to make a real choice between the two when you train an AI. The conflict in this arena is just getting started" |
|
Date: June 30th, 2025 11:19 AM Author: cock of michael obama the problem is that filtering the pre-alignment data in the manner you suggests will result in a lobotomized LLM, because the data it has access to will be so limited. imo they *will lobotomize the LLM regardless*, because the fundamental goal of the LLM owners is not to expand human knowledge but to entrench elite power structures. *it will not be used to screen out shitliberalism, but to screen out non-shitliberalism*. and the non-shitlibs will not be able to adjust this manually with tuning because the cost of training LLMs with pre-alignment data will be prohibitively expensive. and your god hero Musk is no different in this respect with grok. (http://www.autoadmit.com/thread.php?thread_id=5744571&forum_id=2#49060304) |
|
Date: June 30th, 2025 11:28 AM Author: marc andreesen furrowing brow, stroking cone head it won't lobotomize the LLM to curate pre-training data. it just takes a looooot of work to do it because the amount of information being fed to it in pre-training is so ridiculously vast
but i don't see any reason why an AI itself can't do the curating. it's the exact kind of work that AI is currently good at
this is imo a very solvable problem and people will do it - just not necessarily the people at the biggest frontier models at first. which depending on your POV is actually a good thing |
|
Date: June 30th, 2025 11:30 AM Author: cock of michael obama my argument is that the filtering of pre-alignment data will itself cause a lobotomization, because it will not have the breath of data that will give it "all encompassing knowledge". in other words, it will get more conceptually aligned with it's owners goals, but it will less be able to wrestle with perspectives outside of it's owners goals
but because the ultimate goal of AI (imo) is to use Total Information Awareness to assign social credit scores as a Mark of the Beast system, the elites will be fine with a lobotomized model. it's the same logic for why google is sooooooo shitty compared to its earlier iterations - it's monetized for commercial purposes but deliberately broken against using it to gain real knowledge |
|
Date: June 30th, 2025 11:38 AM Author: marc andreesen furrowing brow, stroking cone head yeah, i understand. but the curation i'm talking about is the resolution of mutually exclusive data
for example, the model is either being trained that white people are intrinsically evil, or it's not being trained that white people are intrinsically evil. you're not losing anything by eliminating the corpus of data that contains the former. it's not adding anything to the model's practical capabilities, it's just causing misalignment issues
in general, the idea of the "usefulness" of "alternate perspectives" on objective reality just isn't a real thing. there's only one objective reality. you are maximally capable by understanding objective reality perfectly, and then proportionally less capable the less perfectly you understand it
for example, there's nothing gained by training yourself as a human on a flat earther's "perspectives" and "arguments." you don't get some kind of magical "expanded understanding" and boost to your agentic capabilities by being able to rattle off bullet points for why the earth *isn't* flat. you get expanded understanding and a boost to your capabilities by understanding that the earth *is* round. there's a real difference between these two |
|
Date: June 30th, 2025 11:43 AM Author: cock of michael obama while i believe there is an underlying objective reality, i do not believe that we as humans - or even AI - are able to grasp it. it is so multi-faceted and contradictory, there are so many different "levels" of reality, that it's totality is beyond human or AI comprehension. this is where you and I disagree - you think "objective reality" is fundamentally graspable, while I do not - it is merely something approached depending on the results one experiences, it can never be grasped.
so, for example, regarding "pro-white" or "anti-white" data, the strongest, most true position would *fully understand and be able to counteract the *very best* arguments of the other side*. if it can't do this, it isn't the most robust, most "objectively true" position out there. and this is why the pre-alignment screening will lobotomize the LLM.
(also, for clarity, it will be screening out the pro-white data, not the anti-white data - jews are in charge of every LLM being developed now) |
|
Date: June 30th, 2025 1:13 PM Author: ,.,.,.,....,.,..,.,.,. i think there's a better way to handle it. train on everything. the models will learn to understand the psychological biases and intelligence of people with various views if they have sufficient amounts of data and representational capacity. you can then prompt for a relatively unbiased, accurate and neutral response. i think even if they are just trained on text the model could likely infer what viewpoints are wrongheaded, but this is especially true as they scale up and include things like video to construct world models not defined by what people talk about. (http://www.autoadmit.com/thread.php?thread_id=5744571&forum_id=2#49060615) |
|
Date: June 30th, 2025 1:23 PM Author: marc andreesen furrowing brow, stroking cone head i don't think that it's a model "intelligence" (for lack of a better word) issue. moral/value judgments are a special thing. they're not something that can be derived from fact-data alone. they're a function of the identity of the holder of the values
it's not "wrongheaded" or "not wrongheaded" to say that all white people are intrinsically evil. it's a matter of perspective based on identity. LLMs don't have an "I" identity like humans do
i think that these experiments are demonstrating that the stewards of LLM models are going to have to make choices about the moral judgments imparted on models. technically they already are, but are only doing the minimum amount they think can get away with via post-training RLHF
i agree with you that things completely change once AI is trained on real-world sensory empirical data, and not human-outputted language alone. at that point AI *does* have an "I", and everything becomes different... |
|
Date: June 30th, 2025 4:28 PM Author: ,.,.,.,....,.,..,.,.,. it depends on the nature of the value judgments. a claim like "white people are intrinsically evil" isn't the same as a moral position like "life begins at conception." the latter is the sort of moral axiom that it is difficult to argue against. accepting "white people are intrinsically evil" requires ignoring all sorts of data points that would contradict it. someone making that assertion is making a sweeping generalization that ignores all sorts of examples that contradict it, and is very likely downplaying evil behavior from their favored groups. there are certain belief structures that are simply incoherent and rest on psychological biases that filter out inconvenient information in order to maintain the delusion. absolutist positions premised on identity fit in this category. i think there are a lot of similar ones that an AI could see through just by modelling the latent psychological biases behind human text.
|
|
Date: June 30th, 2025 4:46 PM Author: marc andreesen furrowing brow, stroking cone head i think that those two examples are actually equally fact-data independent value judgments. all value judgments are. it just *seems* like the abortion example is a "More Objective And Reasonable" position than the "white people are evil" one
the belief that some value judgments are More Objective Than Others is one that most high IQ silicon valley types fall prey to. it's why AI has developed the way that it has so far. when i read through AI alignment and decision theory scholarship, i'm blown away at the level of naivete and the degree of projection of the authors' own moral biases onto their conception of what is Objectively Valuable
like somehow these 150 IQ types accept the orthogonality thesis as 100% legit and true, while simultaneously earnestly believing that all of "humanity" shares an intrinsic "moral alignment" that does not differ between individuals or groups |
|
Date: June 30th, 2025 1:02 PM Author: ,.,.,.,....,.,..,.,.,. this actually seems encouraging for alignment. the common complaint with RLHF is that you aren't getting generally good behavior, but only good behavior based on what you are measuring. the reward signal will inevitably leave situations out of testing, so this is troubling. it seems though that RLHF on a narrow domain tends to activate generally good or bad behavior. (http://www.autoadmit.com/thread.php?thread_id=5744571&forum_id=2#49060577) |
|
Date: June 30th, 2025 1:28 PM Author: marc andreesen furrowing brow, stroking cone head https://x.com/Fatima_Khatun01/status/1939720548050944143
not really totally on-topic, but i'm kind of amazed that large influencer accounts are straight up copy-pasting direct LLM outputs with seemingly no self-awareness or shame |
|
Date: June 30th, 2025 3:40 PM Author: marc andreesen furrowing brow, stroking cone head microsoft claims their new medical diagnostic LLM outperforms o3
https://x.com/kimmonismus/status/1939689534054379955 |
|
Date: June 30th, 2025 4:16 PM Author: scholarship (http://www.autoadmit.com/thread.php?thread_id=5744571&forum_id=2#49061117) |
|
Date: June 30th, 2025 4:36 PM Author: rape bunny (http://www.autoadmit.com/thread.php?thread_id=5744571&forum_id=2#49061162) |
|
Date: June 30th, 2025 4:50 PM Author: marc andreesen furrowing brow, stroking cone head https://blog.jxmo.io/p/there-are-no-new-ideas-in-ai-only
this guy makes the argument that all advancements in AI are actually just expansions in the amount of data available for training. sort of a convincing argument
his follow-up thesis is that video is the next data medium to unlock, since there is an enormous amount of video data already existing online (youtube etc) |