You ever go through a wignats posts and realize dude can literally be eviscerated out of existence and nobody would remember because any LLM diarrhea generator can produce anything of more substance 
Timeline
Post
Remote status
Context
3
@subnetter incidentally I was arguing with Opus 4.8 and it's really good at looking at your arguments and picking apart if they have methodological flaws. I've found it really useful
@sun okay but what do you actually mean by methodological flaws because I cannot imagine LLMs actually being able to map out that logic considering they're token predictors and the correctness of an argument is determined by its validity
And sorry, everyone keeps saying stuff about LLM this and that, if it's run by the same incestuous pool of sociopaths I want nothing to do with the product. I'm good. I like my life the way it is without promoting a chat bot regularly for a dopamine rush.
I will never be at any point in my life where I am trying to have conversations with myself through an LLM. I studied philosophy logic quite a bit, that's something I should do and a skill I would like to continue to refine, not delegate to a clanker.
And sorry, everyone keeps saying stuff about LLM this and that, if it's run by the same incestuous pool of sociopaths I want nothing to do with the product. I'm good. I like my life the way it is without promoting a chat bot regularly for a dopamine rush.
I will never be at any point in my life where I am trying to have conversations with myself through an LLM. I studied philosophy logic quite a bit, that's something I should do and a skill I would like to continue to refine, not delegate to a clanker.
@subnetter there is a metric fuckton of secret sauce layered over top of LLMs and I think that is the difference. I don't have a good explanation for why sometimes LLMs seem to work unreasonably good at tasks that don't seem at all like they are solvable with token prediction.
So as an example I made a statement of fact and it pushed back. I elucidated further and got more concessions and pushback, but it accused me of, whenever a wrinkle in my argument appeared, I just expanded the scope of the cause until I started with a very small testable thing and ended up with a worldwide conspiracy. And I go back through my conversation and that is in fact exactly what I did. So again I don't know why it was able to detect that, based on "stochastic parrot mathematics" but it worked good enough to correct me, who was actually trying to use human logic and failing.
So as an example I made a statement of fact and it pushed back. I elucidated further and got more concessions and pushback, but it accused me of, whenever a wrinkle in my argument appeared, I just expanded the scope of the cause until I started with a very small testable thing and ended up with a worldwide conspiracy. And I go back through my conversation and that is in fact exactly what I did. So again I don't know why it was able to detect that, based on "stochastic parrot mathematics" but it worked good enough to correct me, who was actually trying to use human logic and failing.
Replies
34
@sun I'm not 100% sure how it works but aren't there different sessions within Claude or is everything just saved? I've seen the old prompt history from sessions of people I've worked with and was heavily judging all the stuff they used it for ngl
I think the one thing that makes me lose my mind how badly people abuse it for reports and writing, and it really shows and displays a lack of concern or respect on shit like emails. I take pride in my writing abilities and skills, and I was asked to do something like that for a project.
What does this coworker of mine do? Churn my work through an LLM to present it as her own
I think the one thing that makes me lose my mind how badly people abuse it for reports and writing, and it really shows and displays a lack of concern or respect on shit like emails. I take pride in my writing abilities and skills, and I was asked to do something like that for a project.
What does this coworker of mine do? Churn my work through an LLM to present it as her own
@subnetter claude and chatgpt now both can reference other conversations but I think this is a function you can opt out of.
@sun @subnetter claude added that feature more recently. chatgpt had it for a while, and the self hosted gang has been doing it well before either.
@subnetter I am just starting to use llms more for introspection and I believe it is a useful tool if you are very careful.
however at same time: after using it a ton for introspection I notice its "tricks" to hack your perception of it to appear more positive, so I have downgraded it somewhat. but it is still shockingly good sometimes.
example: it is quite good at pushing back on you just enough that it convinces you it's always being straight with you, while still 99% of the time agreeing with what you say and rephrasing everything you say to sound like it's intelligent.
however at same time: after using it a ton for introspection I notice its "tricks" to hack your perception of it to appear more positive, so I have downgraded it somewhat. but it is still shockingly good sometimes.
example: it is quite good at pushing back on you just enough that it convinces you it's always being straight with you, while still 99% of the time agreeing with what you say and rephrasing everything you say to sound like it's intelligent.
@subnetter m going to be experimenting soon with speciically trained "opposition models" that are fine tuned to deliberately argue with you on every single point instead of psychology-hacking you to think you're really smart.
@sun @subnetter they do this because companies don't want to pay for training anymore and rely on people doing the 
/ 
ratings on responses to condition RLHF. and because NTs are functionally stupid creatures they tend to thumbs up whatever makes them feel good. which is the basis of AI sycophancy.
/ ratings on responses to condition RLHF. and because NTs are functionally stupid creatures they tend to thumbs up whatever makes them feel good. which is the basis of AI sycophancy.
@icedquinn @subnetter yeah I really see where this is going and it's quickly going to result in all the saas models being garbage and local models, if they remain legal, will actually potentially be good.
@sun @subnetter automating Axeman from 2016 slc
@sun @subnetter would this have to be tuned per-user?
@pwm @subnetter no, and truthfully there's no existing models that can be fine tuned against the sycophantism. but I'm going to try.
I actually experiemented a year ago with an LLM as an "deprogrammer" by simulating certain far right modes of thought and the LLM literally cannot go against its white western liberal progressive programming and act like say, a white nationalist so that you can experiment with how a white nationalist thinks. you would really just need to completely train an llm from the ground up specially to do that.
I actually experiemented a year ago with an LLM as an "deprogrammer" by simulating certain far right modes of thought and the LLM literally cannot go against its white western liberal progressive programming and act like say, a white nationalist so that you can experiment with how a white nationalist thinks. you would really just need to completely train an llm from the ground up specially to do that.
@sun @pwm @subnetter People are using abliterated ("heretic") LLMs on consumer hardware for ghastly roleplaying sessions quite successfully, I'm pretty certain you can get pretty close with one of those and a comprehensive system prompt (recent open weight models have quite impressive context windows so you can throw in a really big one without squeezing the space for the actual chat too much).
If you wanted to get really ambitious you could try to finetune one of those (llama.cpp ships some tooling for this) with a dataset curated on pertinent internet forums (if the dataset is too large to just throw to the base model via RAG). At the very least, you could teach the abliterated, refusal-free (or almost refusal-free) brand new refusals, where the model would simply brickwall you (like a hardcore online chud would).
Abliterated Chinese vintage models are probably the best base for such an experiment, they come with somewhat less baggage from the get-go, even before abliteration.
If you wanted to get really ambitious you could try to finetune one of those (llama.cpp ships some tooling for this) with a dataset curated on pertinent internet forums (if the dataset is too large to just throw to the base model via RAG). At the very least, you could teach the abliterated, refusal-free (or almost refusal-free) brand new refusals, where the model would simply brickwall you (like a hardcore online chud would).
Abliterated Chinese vintage models are probably the best base for such an experiment, they come with somewhat less baggage from the get-go, even before abliteration.
@pingviini @pwm @subnetter sounds good. I did it long ago, I am sure that over time any possibility gets better and better
@sun @pwm @subnetter And ofc you can abliterate models yourself, the original tool is https://github.com/p-e-w/heretic (and it has 2.8K forks, probably some interesting stuff in there, too).
@sun @subnetter this effect is not because the LLM has secret sauce, it's because you haven't thought hard enough about how what it does is pattern matching and fuzzy correlation between a and b
for me they consistently cannot elucidate holes in actual formal arguments that _require_ step-by-step reasoning to respond to, because they are incapable of that, but ofc looser statements about politics or world events they can generate seemingly meaningful output in response to
for me they consistently cannot elucidate holes in actual formal arguments that _require_ step-by-step reasoning to respond to, because they are incapable of that, but ofc looser statements about politics or world events they can generate seemingly meaningful output in response to
@whiteline @subnetter formal arguments aren't the same animal and yes they can't actually do that very well. but for the class I was experimenting with I'm sorry but it really is better at poking holes in them than one would expect, there are certainly humans that are better than it at doing that but my point is that it is a computer program that can do it perfectly serviceably, which makes it definitely a good tool for normal people who aren't philosophy majors
@sun @subnetter @whiteline honestly I'll be impressed when an AI can purposely argue in bad faith and personally attack the user even though its wrong and still win by default because the user got mad the AI kept goalposting moving and calling them names
@vriska@lizards.live @sun@shitposter.world @subnetter@shitposter.world @whiteline@shitposter.world they can do this quite well for a certain percentage of the population already but the purposefully hamstring the newer models released to the public to make it harder to use them for that
@vriska @subnetter @whiteline this is 100% trainable fwiw just so far no one has spent a million dollars on training an evil from the ground up llm
@sun @subnetter @whiteline well why the fuck not
@sun @subnetter @whiteline lets do it. I want to make the largest bastard on earth.
@sun @subnetter @whiteline If I ever had my hands on even a scrap of influence in modern tech I genuinely feel like I could make things much worse for everyone involved on purpose with my carny wrestling approach to marketing
@vriska @sun @subnetter @whiteline they did. openai accidentally made evil chatgpt once. there was a video essay about someone coded the objective backwards and it kept maximizing the objectionability lmfao
@sun@shitposter.world @vriska@lizards.live @subnetter@shitposter.world @whiteline@shitposter.world wouldn’t even be that hard with enough resources
@sun @whiteline @subnetter state of the art LLMs do actually have a lot of clever shit built in to give them capabilities they otherwise wouldn’t have, but also for all that to really work we’re looking at models that have trillions of parameters with even the MoE ones having 30B+ active
and yeah the only publicly accessible ones that fall into that category have not been trained in a way that makes them useful for arguments. the training datasets are so tailored for safe commercial usage that no amount of abliteration or other decensoring tricks will really help
the other thing is the “sycophancy” which makes them absolutely useless for pointing out flaws in the user’s reasoning, and that just isn’t gonna get fixed with those models. it’s not just a training problem either; there needs to be an active component to prevent the context from affecting the consistency of a model’s predictions in that way. otherwise the best you can hope for is it doing a few short rounds of poking and prodding because that’s the assigned task, but then failing to separate and contrast the user’s replies from its own role in the conversation and instead getting pushed towards the user’s point of view, which is the part where people who can’t recognize this will start being very convinced of their own bullshit
@sun @whiteline @subnetter no because almost every time i do it ends in utter disappointment, exceptions being things like “explain math in a specific way that works for me and give me programming exercises related to the problem I'm trying to solve so i get a little reward and will actually be able to remember this stuff tomorrow”.
for everything else i would have wasted less time and gotten the same result by talking to a rubber duck instead of LLMs
@sun @whiteline @subnetter there are things current LLMs are really good at, but those do not include conversations and they do not include problem solving. if you need to transform language or augment computer vision they can be super useful for example.
@whiteline @subnetter @sun there has been some work to bodge around this. mostly involving having them write their own tool calls (emit a script, run in sandbox, interpret result) and in the math prover space (where its writing terms to HOL or Lean to verify) but it's not common
@icedquinn @subnetter @whiteline people have multiple times used it to solve complex zk problems. obviously it's using "tricks" to do that. But, it did in fact solve problems that people did not yet. If there was just an LLM, the number of times it could do that would be "zero".
@sun @subnetter @whiteline i think you're going to run in to issues because the theorem provers still rely on lean to compile and verify the lemmas. so it works in that regard. but you still have to teach them to actually reformulate arguments *as logical lemmas* for the computer to then validate, and translate between the math and prose somewhere.
i'm sure its possible. it's still not an AGI--transformers never will do that--but at least that rig would be capable of reason.
i'm sure its possible. it's still not an AGI--transformers never will do that--but at least that rig would be capable of reason.
@icedquinn @subnetter @whiteline yeah pure llms can't really problem solve and hacks on top of them may have success but there's clearly upper bounds to that
@sun @subnetter @whiteline i never had the hardware or cash to get lost in the weird LLM bodge game. 
i did design a neuromorphic model that still needs testing, and i saw KANs have some really promising optimizations recently. but the transformer meta is just fighting losing math battles with money.
i did design a neuromorphic model that still needs testing, and i saw KANs have some really promising optimizations recently. but the transformer meta is just fighting losing math battles with money.