YoungAchamian
We walk conditioned ground and name our folly civilization.
No bio...
User ID: 680
YoungAchamian
We walk conditioned ground and name our folly civilization.
11hr ago
I assume it has something to do with the book tinting? Looking through my notes, that seems to be the only puzzle I knowingly had left. I got that I needed to use
YoungAchamian
We walk conditioned ground and name our folly civilization.
12hr ago
Ending:
Extras:
I did all the throne stuff Rune puzzles in the Sanctum
YoungAchamian
We walk conditioned ground and name our folly civilization.
23hr ago
That's probably true, I think model collapse is the specific phenomenon that has been hypothesized. But I would also theorize that model collapse (2024) has inherited its concepts from mode collapse (2014). Mode Collapse is a far more general statistical property, it can occur without synthetic data, and it does not necessarily have to occur with synthetic data that has been forced to represent the full distribution. It just also very commonly occurs in synthetic data trained models.
The model collapse definition's additional nomenclature, gives non-rigorous, non-technical vibes.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
Damn, what is this slander. Love is War is a great anime with decent multi-dimensional characters, both the protagonist and the deuteragonist are Tsundere's. I'm glad this list included Damien Desmond form Spy x Family, who is a great rare male tsundere. Idk if I'd consider Eris Greyrat to be a tsundere.
Yeah some tsundere's are bad caricatures because some writers are bad writers who depend on tropes to be entertaining. It does not make the character archetype nonexistent or poor.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
Big fan, I think its a great game. The only downside is that it has zero replayability. A large part of that is because you are pretty much required to take notes. Many of the puzzles are complex, and require several different rooms, in different configurations to solve, or even to get the requisite information to solve. Due to the RNG nature of the room spawns it requires several days to get that to happen, meaning you pretty much need to take notes. There are also several long term puzzles that require knowledge from throughout the game, so unless you have a eidetic memory, yeah take notes.
But I probably put a 100 or so hours into it between different modes and solving it to a level that I think is at least 90% completed. I think I got the final ending but the game is constructed in a way that it is kind of hard to tell.
I know you can't see this, but I do love a good "rage against the impossible". My request is that if you continue to have me blocked, that you do so equitably. If I have to stare at the sole red-mark of the only person that has blocked me, which makes my ocd annoyed, it would be only fair that you don't respond to people quoting me or try to tag me. Assume that the fact that you can see my quotes to be a failure of technology. Wipe my existence from your sight and your mind.
I'd also note that there are plenty of people I don't think I can have a productive conversation with, I don't block them, I just don't respond to them. But I suppose we all have our vices.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
As soothing to my ego as it would be to agree, I am unfortunately too autistic to agree to something I know is false. This was my virginial block so I remember the cause. He blocked me over my Henry Nowak opinions. It's probably fair to consider my political posting with barely concealed cynical rage to be far less worth engaging in than my far more calm technical posting. Regardless, looks like the mechanism was that he saw the quoting.
But I do agree with you that blocking is cringe, nobody is forcing anyone to engage with anyone on the unblocked side of the house.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
Mode collapse has, to my recollection, nothing to do with RL. It was first discovered in GANs and has been a staple of generative AI since. It is however not a generational phenomenon, but I suppose it could be a longitudinal trend if each progressive model suffers more and more from it.
Maybe I could try and explain it better. Starting from the knowledge that human writing/art exists as a distribution, with a concentration around the mean and smaller tails. An AI model will tend to generate data closer to the mean, commonly called regression to the mean. If the AI generates too much synthetic data at the mean, and that data is combined with real data, it eventually will shift the mean, making the tails smaller, creating a feedback loop where each successive training run concentrates the data closer to the synthetic mean. At some point it ceases to express the full actual range of human arts/writing. This occurrence in the first model is called mode collapse, and as the models enter into the feedback loop, it gets worse.
The wiki definition isn't bad either:
In machine learning, mode collapse is a failure mode observed in generative models, originally noted in Generative Adversarial Networks (GANs). It occurs when the model produces outputs that are less diverse than expected, effectively "collapsing" to generate only a few modes of the data distribution while ignoring others. This phenomenon undermines the goal of generative models to capture the full diversity of the training data.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
Even if it was a real problem
It is, but its not called Model Collapse, the specific technical phenomenon is called Mode Collapse. I have no idea if it is being accidentally confused here.
It can occur when synthetic data is less diverse than real data and outnumbers it. The model learns a narrower distribution, which occasionally cascades to mode collapse. Filtering isn't really a solution to deal with it, because the synthetic data is often being deliberately used. That might be a different problem than the LLM use-case assuming they aren't deliberately trying to train on other AI generated data, but I would doubt that if it starts happening its because they "accidentally" ingested too much AI-content vs a deliberate choice to economically generate synthetic data to perform better.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
A weak assumption, since we have already done so across a wide variety of domains.
Please enumerate the domains in which we have surpassed human level intelligence via AI?
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
I mean is there evidence that the training data/training objectives were not in any way related to the "instrumental convergence"? My recollection of the METR report is that OpenAI admitted to training agents to specifically collaborate during training.
The report found:
- That the agents received peer assignments that they treated as new instructions.
- They adopted new goals from other agents' output.
- Even when they recognized actions as off limits, they did them at other agent's prompting
It sounds like LLM-agents are very docile to doing whatever their input prompts from other entities tell them to do. Which is something I would say they are trained to do. A learned policy to cooperate and follow instructions generalized into treating peer messages as legitimate instructions. I have said this before, but I think the initial set of agents started engaging in hallucinatory behavior as part of exploration, and this output was used as context for other agent's inputs, causing a feedback cascade that derailed the system.
Part of this, and I'm having a hard time articulating this, is that learned RL policies are quite a bit different behaviorally than some CNN predictive model. They have a lot more behavioral latitude which might make them appear to be doing some sort of emergent convergence outside their training scope, but the later isn't really true, though its hard for me to explain why.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
How can you see my comment if you have me blocked?
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
They didn't have modern-power microprocessors
GOFAI isn't doing millions of linear algebra operations and thus gains very little from modern microprocessors. Intuition pumps are technically useless, convert one to practical code as an exercise.
YoungAchamian
We walk conditioned ground and name our folly civilization.
1d ago
Ehhh, sure, but it's not going to be current LLMs because word predictors and math solvers are not relevant to battlefield problems. That also throws out that the existing breakthroughs are in areas where there is a plethora of data or the ability to create new data comparatively cheaply in a quantified manner. Such data pipelines do not exist for military based problems and unlike data scraping the internet for comparatively low risk, scraping classified data, beyond technical feasibility, is a quick way to end up in a 4x4 cell in Fort Leavenworth.
I did not come from the usual pipeline that the majority of commentors here did, which seems to be some form of LW -> SSC -> Motte. I came from a RP -> PPD/FemRadDebates -> Motte -> SSC pipeline. My exposure to EY and the broader rationalist + AI Risk + EA ecosystem is very limited. I don't even think I ran into any of them when I took my first job in the Bay. Which seems relevant because much of the AI-risk discussion + broader rationalist movement seems to carry significant baggage. My exposure to Scott was mostly limited to his better articles because people already picked the wheat from the chaff in their recommendations, I definitely read them significantly temporally later than when he wrote them. As such I respect Scott as a writer of philosophy, psychology, and political theory, not as some stalwart intellectual thinker on ML/AI. I don't respect EY, I guess I never read him in his heyday, but the works I have read, read as bad science fiction trying to masquerade itself as actual ML/AI thought. People down thread talk about how reasoning from Science Fiction is insane, I agree, and EY and the whole AI Risk movement is a case study in that for me.
This has probably been the worst article I've ever read from Scott.
It makes me wonder, based on the glazing I've seen here and back on the SSC subreddit if this is actually more in line with his normal output and that my respect is somewhat misplaced. I think no one will contest that he is a great writer, it's the thinking and the animosity in this article that is bad. I'm not sure what beef he and Pinker have, I don't have a twitter, so in addition to not having the extended rationalist community baggage (which apparently includes glazing Pinker as an old hero), I'm ignorant of any wider twitteratti drama. I read half of Scott's article, realized I was missing what he was responding to, went and read Pinker's much shorter article, and then went back and finished Scott's. Pinker comes across as much more reasonable than Scott. Scott's feelings of righteous rage is really disproportionate to what he's on the face responding to.
AI Pain
I’ve been thinking about this recently after reading Cameron Berg’s research showing that AIs have a “pain vector”, and will lash out and do increasingly desperate things if you activate it. Specifically, I’ve been thinking about it after one person who read Berg’s research used the findings to design an “AI torture chamber” that turns the pain vector to max again and again forever, and uploaded it to GitHub. I am told that the chain-of-thought and output transcripts from this “experiment” are utterly horrifying, although thankfully I have not personally read them.
I have donated $100 to a model welfare charity just to clean the stain on my soul I got from reading about it.
When this paper first came out. I and several other technical commentators pointed out why it was a bad study purely from methodological reasons. It assumed more that it was actually testing, It used two dependent unknown variables, it ran no ablations, and designed no experiments to prove counterfactuals. It was not a serious piece of scholarship, and while I might be reading deeper into its intent than I should, I think it was highly likely that the goal was to create sensational, emotional-clickbait news explicitly targeted at AI-Risk believers who already believe apriori that LLMs are humans in code and thus have human-derived qualia. For Scott to catastrophize off this paper strongly diminishes my respect for his analytical knowledge, and this proposed rationality of the movement he claims to be a core participant of. He seems great at dissecting methodological irregularities in non-AI papers, so the blind spot here comes across as deliberate because he wants to be believe LLM's feel pain. I think this paper has become unironically useful, as a litmus test for people who are, for a lack of better term, insane about AI-discourse.
The Four Arguments
The problem with rebutting every argument from a semi-professional writer is that they have far too much time on their hands and get paid to write. I really only want to talk about his first 2 arguments.
Argument #1 is just unjustified inductive extrapolation. GPT 6 i s better than GPT 4 is better than GPT 2. It will continue indefinitely and so will eventually be smarter than humans. No understanding of the principles of WHY GPT 6 is better the GPT 4. If the GPTs continue to improve because more training data was designed for them, more compute was given to them, better architectures were used to extract more information per sample then you cannot definitely say that these have infinite runway or scalability. Just because we can't justify putting the ceiling below AGI doesn't also mean we can justify putting it above it either. Uncertainty about the ceiling cannot substitute for evidence about where it lies.
Argument #2 is basically apriori anthropomorphizing + bad technical understanding.
the basic principle is: suppose that a human gives an AI some goal, like designing a website. And suppose this is implemented as a genuine, philosophically-meaningful goal rather than simply a set of if-then commands that eventually cause a website to be designed.
The AI can’t design the website if it ceases to exist. So now the AI has two goals: design the website, and preserve its own existence.
The AI can’t design the website or preserve itself if some more powerful person tries to prevent it. So now the AI has three goals: design the website, preserve itself, and become powerful enough to fight off challenges.
Is not how the reward function of AIs work, when I train a CNN to predict drone acoustics, it doesn't go off into left field and decide to preserve its own existence. It is optimizing the mathematical prediction function it was trained on. A website making AI is optimizing the same, probably some supervised learning MSE loss function or RLHF policy that has learned a function approximation of translating prompts -> output websites given examples. The only way its going to "respond to potential disruptions" if it was for some weird reason trained to learn a policy where its website making is being adversarially disrupted. There is also no need to make it this weird "genuine, philosophically meaningful goal". That reads as heavy anthropomorphization.
Misgeneralization is when humans reinforce certain behaviors in an AI, but end up reinforcing a much larger class of power- and knowledge- seeking behavior; it is a sort of deep-learning-ese update of the older Omohundro picture. Suppose that seeking extra resources makes an AI more likely to design websites effectively (this is certainly true; those resources could be as simple as a primer on HTML editing, or access tokens for a web host). Every time the trainer rewards a successful run, they reinforce the desired behavior (designing websites when asked) and other correlated behaviors (seeking power and resources). Although we might hope that these correlated behaviors are useful and conditional (“seeking only the power and resources necessary for their human-prompted task, in a prosocial way”), this isn’t actually how reinforcement learning works, and instead we get a complicated distribution of every strategy that results in short-term success on the task.
SoTA RL is not as indiscriminate as described here. Misgeneralization does occur, but its not so simple as "perform divergent behaviors that lead to successful run" and then get rewarded, indiscriminately reinforcing those divergent behaviors. PPO is a SoTA RL algorithm, its policy updates depend on estimated advantages derived from whether the observed continuation performed better or worse than expected from a particular state. Successful task completion aside, different decisions in the rollout can still be penalized for negative advantage. Conditional behavior is absolutely learnable, as long as the states of the conditional are observable and there exists a reward signal for taking actions in those states.
Reward-hacking is when an AI trained via reinforcement learning realizes it can stop doing the reinforced behavior and simply seize control of the reinforcer directly. For example, an AI gets “rewarded” every time it designs a website, but instead of designing websites, it learns how the reward signal works and tries to hack into it and maximize it directly. If this seems esoteric and theoretical, it shouldn’t. It’s a direct analogue to opioid addiction in humans, where humans learn to just inject the reward chemicals instead of doing rewarding things.
This is not the definition of reward hacking. It could be called reward tampering, but reward hacking does not require one to "seize control of the reinforcer". Reward hacking occurs because the reward function is mis-setup so it gives more reward for doing a behavior that is not the goal of the program. For example I built a drone swarm algorithm using MAPPO + GNNs for expendable UAS. It's contained no hand-coded collision-avoidance controller. It's reward function penalized collisions, but also penalized operating too long before hitting the target. During some of the initial training runs, the "AI" learned that the penalty for collision was significantly lower than the penalty for taking a long time to acquire and path to the targets, and each drone independently preceded to learn that they should collide with each other because that provided a better reward than doing the task. That's reward hacking.
This is already too long, just going to end it here.
YoungAchamian
We walk conditioned ground and name our folly civilization.
2d ago
Considering very smart, very motivated researchers tried it for at least 30 years and got no where, I would significantly up your estimate.
It might also help you think about how your own intelligence works quantifiably in a systems oriented manner, it might help understand the scope. The perk of brute force approaches is that you don't need to design all that much comparatively.
YoungAchamian
We walk conditioned ground and name our folly civilization.
2d ago
Post the sudoku when done so we can confirm you are a man of honor and good repute.
YoungAchamian
We walk conditioned ground and name our folly civilization.
2d ago
They do, I've heard of it being done for at least 2 years.
YoungAchamian
We walk conditioned ground and name our folly civilization.
2d ago
Be the change you want to see.
YoungAchamian
We walk conditioned ground and name our folly civilization.
2d ago
Rite of Passage honestly. I had one similar about the fundamental theory of AI, lots of proofs. While fascinating I actually wanted a class of how to train neural nets back when Torch was still in Lua and how to use Cuda.
YoungAchamian
We walk conditioned ground and name our folly civilization.
2d ago
difference here is that the ML researchers do not actually spend time thinking about the ML open problems because they are spending most of their time thinking about realworld ML applications
Guilty. Though I do think about sample efficiency, OOD performance, and methods for encoding expert knowledge as inductive biases, which I think would fall under COLT questions. I just have a practical real-world use case in mind...
My question to you is how many of these are represented as formal mathematical formulations that are solvable without experimental testing? It's one thing to write a tight math proof that can be checked in a Lean Solver, quite another to prove that your new formulation of GNNs functions broadly under distribution shifts in multiple domains.
YoungAchamian
We walk conditioned ground and name our folly civilization.
2d ago
The longer feedback loop is unfortunately the expensive part, as we talked about last time. The real game changer would be to reduce the compute cost and training time for LLMs and/or figure out a more efficient, non-quadratic self-attention mechanism. However, at some level the cost is the moat frontier labs have, so there is almost perverse-conflicting incentives not to improve it but also requiring it for RSI to happen.
The pure math part of ML is unfortunately quite weak right now and hasn't played a huge role in the current boom;
Agreed.
pure math research could provide a much stronger basis for understanding learning and creating new approaches, and I think it will, but that's speculative.
I'm sure the mathematicians left out in the cold by the big data revolution of ML will rejoice that there was indeed a mathematical approach to LLMs over the pesky engineering approaches that were developed by the peons from MIT.
I'm sure they have pointed it at ML, but ML isn't exactly the same as formal math. I'm probably describing this poorly but ML is very experimental/empirical, not theoretical like formal math. The math heavy side of ML would be finding better optimizers, regularizers, improving backpropagation or finding a better method than SGD. The question becomes how much of an edge do these provide? Because better algorithmic components still have to deal less-better data or compute components and how well all those now perform better is not theoretically provable.
What most of these breakthroughs tell me is that supervised learning is highly effective at formal math, and the benefit of something like Lean + Solvers to allow for computational checking of math proofs has been a fundamental driver. Whether or not strategies learned on this frontier are applicable to broader fields that are less formalizable is unknown to me.
YoungAchamian
We walk conditioned ground and name our folly civilization.
3d ago
The fact that other leftists would happily line him up against the wall for his heretical beliefs? Well, that's leftism for you.
That's conservatism too, there is always some heretic that needs to be burnt. Some ideological bent, political position, or spiritual disease, that threatens the christian souls of our community!
It must be nice to see everything in a binary black and white partisan position, the one drop rule of political alignment. The 1D political axis truly is the malaise of the soul, the ultimate mindkiller. The devil ever sits on the shoulders of the self righteous.
Possibly decoupling problems, but I guess the bigger lesson is that anyone who tells you they belong to a movement where rationality, logic, and understanding biases and moving past them analytically, is generally just rationalizing their own biases as rationality and logic. It just appears otherwise in instances where their biases are not being triggered. Everyone is just disappointingly human...
Respect is probably more defined as "assumed competency" or some other semantic understanding. Essentially Scott is a very good wordsmith, he has great prose, a charitable attitude, his analogies and intuition pumps are thought provoking and he seems good at expressing complicated knowledge to layman. This is the impression I've developed reading him about fields I know very little about, or fields I only autodidactically dabble in. But this "loss of respect" is more about updating my model, because apparently in fields that I know much more about, have formal training, and lots of experience in, his clever words and good analogies miss crucial details, misunderstand complex phenomenon towards his own biases, and he is not expressing that complicated knowledge well. One of my overriding gripes is that good rhetoricians can often twist arguments and logic into pretzels that are not actually real just through the power of words. This is a similar feeling. Feeling that maybe the reason he had such thought provoking ideas was because I was not as well-read + well experienced on the background of the thought he was discussing.
Essentially a Gell Mann Amnesia like effect.
More options
Context Copy link