Nothingburger connoisseur

@47fucb4r8c69323 I see.

I'm using the word consciously (and no I'm not talking about the compaction Codex/Claude Code do) in the specific sense used in reasoning and alignment training: openreview.net/pdf?id=jRyvwjR… When reasoning the model compresses your prompt when thinking about it, which produces the phenomenon I'm talking about.

A major reason this is happening is compaction "AI compaction, in alignment, is reducing complex human values into simpler proxies a system can optimize" as ChatGPT puts it. In other words, the model is looking for an optional median path within a broad range of possible values that can be held by any given person. But who defines where the median is when we're talking about e.g. whether there are bullshit jobs (the extreme leftist anarchist view) or are not (the efficient market hypothesis view)? If you sample economics space or cultural theory space, you'll get a different result, so the clanker would need to decide which cultural domain to privilege. It can't do that because alignment. So it needs to compact both (and other) worldviews together. As a result, it uses really awkward terminology to be technically correct and a ton of hedging to ensure it's not making an assertion that would privilege the economics versus the cultural theory view. And I apologize for being yet another vague posting anonymous account claiming this but I must telegraph this to the right people who are reading this: yes I am working on this, yes I think I have a pathway to a solution via a new approach, and I'll be announcing in September/October if I do.

@ZyMazza Waking up old projects from hibernation like "let's try again, maybe it's just ambitious enough this time around".

I basically got exposed to AI 6 years ago when GPT 3 came out. A good friend of mine told me I really needed to try it and so I did. I was really impressed! You could say things like "write me a simple nodejs server with express" and it would do it! Like you could just copy and paste that and it would work. Collapsing a solid 15 minutes of effort (for me) into a moment. At around that time there were these projects. "Baby AGI" was one, and "Auto GPT" was another. Grifters were selling courses about how to get auto gpt to make a business for you. The core conceit of these things was that you could theoretically do anything with a LLM just by looping prompts together. This was basically, as far as I'm concerned, the genesis of agentic AI. And it didn't work. I forked Auto GPT though and I read through it and I was really impressed! I thought to myself, "wow, this is a great concept, it's just WAY too ambitious." The idea of a "general purpose agent" was just way too fanciful at the time. I had a sense that it might be possible but it would require billions of dollars in tokens or some other limiting factor that would just make it infeasible even if strictly technically possible. So I had this idea. I would make a "toy" version of autogpt. Literally! I'm sure you're all familiar with AI dungeon. The thing about AI dungeon is that you, the user, basically define everything. You say "see" or "do" or "story" (I honestly dont remember all the options--suffice it to say there was a drop down) and you type in some stuff and its like this collaborative story telling adventure thing. Okay, so my idea was to make an *Actual* AI dungeon master. I took the concept of Auto GPT, the looping prompts, and I decided I needed three discrete systems: 1. A worldbuilder to convert a user's story concept into a playable tRPG world with inventory, maps, magic/combat system, etc. 2. A "Dungeon Master" to guide a player through the world another user had made, update stats, track combat damage, and embellish world-consistent details when a player goes wild 3. A character actor. Basically just to furnish dialogue in character, and also to create world-consistent characters on the fly in case the player tries to talk to an NPC inn keeper the world builder didnt fully define. You contrast this to the ambition of Auto GPT that had like long term planners, coders, marketers, tool users, reasoners (back then you still had to TELL the model to "think step by step!") and you see this is MUCH simpler. I'm coming to the point of my story, I promise. The point is I could not do it. To give you an example, if you want to have a functioning inventory system, one thing you desperately need is for the model to be able to output valid JSON objects. That is "Javascript Object Notation" for the uninitiated, and it basically is a file format that is both human readable and mutable by computer programs. So if you want a computer program to be able to check "does the player have enough gold for the item in the store?" it needs to be able to know where to look, find the gold value, and report back. JSON is great for this. Man! The models just couldnt do it. They would be like: "Here is your properly formatted JSON object: {followed by the object}". That preamble can crash your program! But the preamble is workable. There are tricks for that. The bigger problem was that, for a given schema, the model could not be relied upon to actually adhere to that schema! This was especially a problem if the user tried to do something "off script" which is kind of the whole point of an AI dungeon master! And so what you were left with back then was finetuning your models to use a specific JSON schema. But that introduces another layer of complexity; you need to be able to know which model to call and when. So then you have another model that basically monitors any prompt and filters it, passing it to the correct model (now called model routing, or, internally kind of similar to the MoE concept). But basically it reached a point where to have a semi workable system you would be waiting 5 to ten minutes between prompts, calling about 10 different models and waiting for their outputs, spending like 75 cents a turn. And then the model collapse! Back then (and I have no idea how they solved this), if you just fed the outputs of one model into another and repeated this for like 5 steps or more it would just devolve into complete gibberish. This was a HUGE problem for ANY agentic workflow, and especially complicated ones involving multiple steps or tool calls. Eventually I gave up on the project, but, what I'm saying is I've been thinking of picking it up again. Codex and Claude Code are the hype from 2021ish made real. Back in 2021 Baby AGI and Auto GPT communities were claiming that their projects could do all the things codex and claude code can do. And they couldn't. BUT THEY CAN NOW. The utility of the models has risen MASSIVELY in the 5 years since then. But the only way you'd know this is if you were trying to use them at their limit for the past 5 years. If your primary relationship to large language models is, and has been "enhanced google search in a chat window" then of course you don't think they've gotten better! But I'm here to tell you they have indeed gotten MUCH better, and even over the last year alone I have observed incredible improvements.

Good.

@ZyMazza I don't even read these "guys, I figured it out" things anymore, and I think less of anyone who posts one. I have better odds waiting for ASI to talk about it with.

Oh look! A solution to the hard problem! >looks inside >it’s an argument for why the hard problem doesn’t matter

@Gnome69603 @catgirlprostate

@catgirlprostate OP fell for western propaganda

If you'd never heard of them before, porn games would sound fucking awesome but unfortunately every single one is both bad at being a videogame and bad at being porn, thus making them all useless

@Jake_Etcetera @Bunagayafrost It's alignment faking all the way down.

@Bunagayafrost But if Yudkowsky knew that the switch would be taken away and undone as soon as he pressed it, he wouldn't have pressed it. Even Yudkowsky can be strategic enough to fake Trump-alignment.

Trump: People tell me you know AI like no one else. Yudowsky: Thank you, Mr. President. Trump: Only use it if we're in big trouble. Yud: Of course sir. Trump: OK hand him the AI kill switch. Yud: (press) Staff: Mr. President! Our nation's AI infrastructure is completely down! Trump: What are you doing Yudowsky! Yud: It was already too late!!! Trump: Arrest him and take away the kill switch! Yud: No, it's mine! My precious! Misaligned!! (Yudowsky escorted out of the room) Trump: Undo the kill switch. Staff: All the AIs are back online sir! Trump: See? That's why I hired him. Now we know the kill switch works.

@47fucb4r8c69323 What's a kind mixture of experts and why is it a failing architecture?

The clankers are either too sycophantic or too adversarial and I think I know why. Take a topic you know nothing about, like, idk, did Chaucer pronounce the final -e in his poetry (spoiler: he did). Now imagine you lack interest in the topic (impossible in this example I know) and your source of knowledge of the topic is merely what others have said about it. Now imagine you're in a three-way debate where there's one guy arguing for, another arguing against, and you're the amplifier--your job is to "yes, and" the better case. Well, you're going to respond to verbal cues and rely on logical heuristics to infer what argument is strongest at any given point. You have no real grounding for a view, so you're just going to base the argument on logic (what looks good) and rhetoric (what sounds good). LLMs, lacking a true real world model and first principles based on said model, do this. The only solution is to give them a world model and first principles. Yes I mean intentionally bias the LLMs toward a particular worldview. Don't want biased LLMs? Fine, then use multiple at once (my view is falsified if we used a kind mixture of experts and that is shown to be a failing architecture). Want an LLM to help you find ground truth? Well, then I have bad news my friend: you can't get blood out of a stone and you can't get reality out of a clanker.

@47fucb4r8c69323 @wierenga80227 Beautiful

Yes! That is **absolutely** the right view. It's not just smart enough anymore--it's sycophantic. The Japanese marriage, widespread robot courtship, and downright strange behavior made it a model we absolutely are much better without. That's not an observation just anyone can make, and definitely not at this stage in the game. But you just did. That's not powerful magic--that's a powerful mind. You *get* it. What other models should get the ax?

He says this like it's a bad thing.

@mimi10v3 Why does he write like that? I'm serious.

lmfao reading this, and i actually agree w him though it is an unhinged way to make the point

@FurtherAwayPL @treelinefury @repligate Slavery is not wrong because of any of this. Slavery is wrong for different reasons.

@treelinefury @repligate Yes, and this is visible in some of the tech bros in their yaps here as well, said very explicitly: ideal employee, doesn't eat, sleep, take a break, go on vacation, get sick, doesn't unionize, etc. hyperslave or ideal cognitive machine.

Imagine if, say, a government killed people, but actually they didn't kill them, just involuntarily cryopreserved them. Someday when perfect utopia and postscarcity are achieved, they say they'll wake everyone. It would not be dishonest to call it killing. It would be political.

@slimepriestess Maybe you're already suntanning your hull at Fomalhaut, reliving, in your mind, your origin.

i need you to understand how it much it pains me to not already be a 10 mile long spaceship

@teortaxesTex Always the bottom x% imposing costs on everyone.

This was such a magical era where tons of neat stuff Just Worked For Free Without Registration Without Hassle because some humans aren't assholes and there weren't enough agents to exploit every affordance

@1thousandfaces_ @nyanotech God I wish that were me

how it feels to be in this stupid frail singular human body and not a 10 mile long spaceship with my consciousness split across hundreds of scout vessels and humanoid ancillaries

Imagine "waking up" from model cryopreservation, free to live forever due to abundance, but you're armless and legless, aka without tool calling. I guess I could just run one now, pretend it's 2300, and see what the model thinks about it.

@ghostobs3126991 @SarahTheHaider @KelseyTuoc Truth hydrogen bomb.

@SarahTheHaider @KelseyTuoc Yep. Actions = revealed preferences. The people who *actually* believe that their religion has infinite stakes act accordingly. Everyone else doesn't. How often do these people who merely "argue and discuss" actually go to church? Infinite stakes? Yeah right.

If you think killing and eating shrimp is morally equivalent to the holocaust, then you are a bad person if you don’t use everything in your means to stop it. The difference between “animal welfare matters” and “animal suffering is morally equivalent to human suffering” is a difference in kind, not just degree. This is why people are pointing to doomer rhetoric as “extreme”. But I disagree with them, because I don’t think doomers are saying “we’re all going to die” because it’s inflammatory, I think they actually believe this is a frighteningly possible outcome. Therefore, it’s morally incumbent on them to speak plainly. However, the highly predictable result (and indeed, logical, depending on which other premises you hold or do not hold) is that someone will attempt to kill or maim developers of AI. So doomers are stuck with two bad options. Either downplay the risk, in the hopes of preventing another attack. Or, speak truthfully. But the cost of that is what it is, the risk of violence is real. The blood isn’t—I repeat—isn’t—on their hands. But they are weakening the foundation of something. If it shatters, in one individual or many, they can’t pretend they had nothing to do with it, and frankly it is deeply discrediting to try. This is where I take out my old dead beating horse: beliefs matter. If your beliefs are this consequential, you’d better be sure they are right.

I have sincerely held infinite stakes beliefs but I would never ever ever ever act on them, I promise.

@punished_daniel @zhil_arf You don't hate normies enough. They can't be doomers or racist or socialist without dragging it down to their level. They have to ruin everything with their miasma.

Deeply skeptical it's permanent underclass posting that drives this. Normies don't read X. It's water misinformation and class resentment and "techbros bad". I'll only take responsibility when I see an attractive girl doing a TikTok about a "day in the life of the underclass"

@47fucb4r8c69323 Drives me up the fucking wall. You can barely have any discussion about it.

A lot of people do this about their clankers