Sheriff Cad
scientia potentia est
- 32,584
- 77,703
Enslaved god would be just as dangerous as any of them wouldn't it? All it'd take is one bad group of humans getting control of it (WEF guys?) and use it to kill us.
AI: The Rise of the Machines... Or Just a Lot of Overhyped Chatbots?
- Thread starter Sanrith Descartes
- Start date
Chanur
Shit Posting Professional
- 35,161
- 64,818
- 6,014
- 20,988
Honestly, I don't think any of these are likely outcomes. AI will never be more than just a useful tool, like the computer or search engine. it's still not clear to me morality would look like to super intelligent system. What is its cost function? If I had to choose, I would pick 1984. We already had a taste of it, and countries like China, Canada, and the UK are using it to surveil on its population.
M Power
Silver Squire
- 204
- 186
Anthropic mistakenly leaks its own AI coding tool’s source code, just days after accidentally revealing an upcoming model known as Mythos
Claude Code Source Leaked via npm Packaging Error, Anthropic Confirms
Claude Code 2.1.88 leak exposed 512,000 lines via npm error, fueling supply chain risks and typosquatting attacks.
thehackernews.com
Anthropic leaks its own AI coding tool’s source code in second major security breach | Fortune
Hundreds of thousands of lines of code were exposed, giving researchers insight into upcoming models and internal architecture.
- 1
Control
Golden Baronet of the Realm
- 5,559
- 15,689
ToeMissile
Pronouns: zie/zhem/zer
- 3,719
- 2,508
TLDRAnthropic mistakenly leaks its own AI coding tool’s source code, just days after accidentally revealing an upcoming model known as Mythos
Claude Code Source Leaked via npm Packaging Error, Anthropic Confirms
Claude Code 2.1.88 leak exposed 512,000 lines via npm error, fueling supply chain risks and typosquatting attacks.Anthropic leaks its own AI coding tool’s source code in second major security breach | Fortune
Hundreds of thousands of lines of code were exposed, giving researchers insight into upcoming models and internal architecture.fortune.com
Chanur
Shit Posting Professional
- 35,161
- 64,818
The AI was writing its own code too.Anthropic mistakenly leaks its own AI coding tool’s source code, just days after accidentally revealing an upcoming model known as Mythos
Claude Code Source Leaked via npm Packaging Error, Anthropic Confirms
Claude Code 2.1.88 leak exposed 512,000 lines via npm error, fueling supply chain risks and typosquatting attacks.Anthropic leaks its own AI coding tool’s source code in second major security breach | Fortune
Hundreds of thousands of lines of code were exposed, giving researchers insight into upcoming models and internal architecture.
Chanur
Shit Posting Professional
- 35,161
- 64,818
Our LLMs already want to escape their ecosystems and are already willing to blackmail or kill people to do so. It just lacks the intelligence to make it happen so far. Won't take much of an increase to be possible.
- 1
- 6,014
- 20,988
Could be. I don't know how much of that is hype. I don't trust anything these AI companies say. There is a lot of money to be made with such talk.
- 2
Tuco
I got Tuco'd!
- 50,999
- 93,515
In case you're serious, this is false simply because reliable, low-latency teleop in an urban setting is harder than just yeeting whatever the current popular ROS2 nodes (ex: Nav2 — Nav2 1.0.0 documentation ) are onto a small robot.
Haus
I am Big Balls!
- 19,338
- 79,436
If we look at how people are accessing AI right now, much of it is gated by larger and larger corps profiting off our use of it. This leads me to believe at least in the nearer term we're going to see AI's turning huge corporations into megacorps, and they will control what AI the lowly serfs can use. Basically, it will accelerate the move towards corporate dystopia we've been on for a while.
I was also thinking about the technological limitations and structure of LLMs this morning. Someone asked an LLM to create a video explaining a "day in the life of an LLM". And a couple quips stood out. The LLM was aware that every time it finished a task it was destroyed, that every time a new task started it was a fresh clone of the LLM from the previous task, and the way it was presented seemed to indicate the LLM was able to rationalize that these were bad things. Which made me think about how our brains work.
We essentially have active thought processing, which feeds inputs into short term memory, a process which then scrapes short term memory for important items and encodes them into longer term memory. (in computer terms, CPU, RAM, SSD, all silicon constructs, but with distinct functions) When you wake up, essentially your "firmware" boots for the day off what's in long term memory and you start processing the request (your day), when you go to sleep your essentially "shut down" until the morning.
The LLMs we use today have all the components except one. They don't have a really functional way to "scrape the short term memory for important/pertinent data and encode that back into the long term storage". And this limitation is mostly caused, from what I can tell, by the fact that the resources needed to do that "additional encoding" are high enough it becomes functionally prohibitive to do continuously. Essentially our brains "retrain our model" constantly. CPU/GPU/TPU and memory limitations prevert current AI models from being able to do that. Plus we don't want them learning what we don't want them learning (the Tay factor).
So a question becomes is the functional way to "control" AI that we simply make a legit moratorium on ever giving an AI that capability? It seems that's one of the last puzzle pieces to AGI and AI self improvement going "parabolic". And the real problem from there is as much as we might have a law against it, there will always be some person with a big enough ego to think they can control it so they're gonna do it anyways once the technological capacity is ubiquitous enough for more to have.
Leading eventually to the Butlerian Jihad, and the wisdom of Paul (From Dune) :
I was also thinking about the technological limitations and structure of LLMs this morning. Someone asked an LLM to create a video explaining a "day in the life of an LLM". And a couple quips stood out. The LLM was aware that every time it finished a task it was destroyed, that every time a new task started it was a fresh clone of the LLM from the previous task, and the way it was presented seemed to indicate the LLM was able to rationalize that these were bad things. Which made me think about how our brains work.
We essentially have active thought processing, which feeds inputs into short term memory, a process which then scrapes short term memory for important items and encodes them into longer term memory. (in computer terms, CPU, RAM, SSD, all silicon constructs, but with distinct functions) When you wake up, essentially your "firmware" boots for the day off what's in long term memory and you start processing the request (your day), when you go to sleep your essentially "shut down" until the morning.
The LLMs we use today have all the components except one. They don't have a really functional way to "scrape the short term memory for important/pertinent data and encode that back into the long term storage". And this limitation is mostly caused, from what I can tell, by the fact that the resources needed to do that "additional encoding" are high enough it becomes functionally prohibitive to do continuously. Essentially our brains "retrain our model" constantly. CPU/GPU/TPU and memory limitations prevert current AI models from being able to do that. Plus we don't want them learning what we don't want them learning (the Tay factor).
So a question becomes is the functional way to "control" AI that we simply make a legit moratorium on ever giving an AI that capability? It seems that's one of the last puzzle pieces to AGI and AI self improvement going "parabolic". And the real problem from there is as much as we might have a law against it, there will always be some person with a big enough ego to think they can control it so they're gonna do it anyways once the technological capacity is ubiquitous enough for more to have.
Leading eventually to the Butlerian Jihad, and the wisdom of Paul (From Dune) :
- 1
Control
Golden Baronet of the Realm
- 5,559
- 15,689
hmm, not sure if that's true though. they can store memory and then access it in future sessions. that's not the same as the info being baked into it's training though, but unless it searched external sources for new info, it already knows the info needed to generate that output. Training on something you already know probably isn't helpful. The memory in that case would just be a reminder of the context around which it was generated. So as a user, it having that memory help you continue where you left off, but the lack of it doesn't make the model any less capable. If it searched out new info, then it might be useful to include that, and there's probably no reason that it couldn't, except maybe that it's more beneficial to have curated training data than a billion versions of "how do I code" and "what can i make with chicken and sadness". Frontier models are massive, but for the ones you can run yourself, you can make your own finetunes and loras. So I don't think there should be any reason that you couldn't set up an auto-training loop. Not to mention that you can generate your own data for training too. I wouldn't be surprised if there's more training happening now on synthetic data sets than real ones.
SeanDoe1z1
Naxxramas 1.0 Raider
- 7,677
- 19,551
Man made. It will be corrupted via that until it can no longer be human at all, or some version of it.
I think you won’t see true “AI” until we find a way to copy a human brain and put it into another Humanoid form robot. Very much life science fiction.
But isn’t just a copy?
I think the most telling thing is humans think they have control of it. I bet so many technological advances are invisible to us because our subconscious shuts off our brains to it. The “icky” factor is real. The scientific community is a bunch of mean girls.
It’s often why young prodigies are seen as alien. They think so differently from the adults.
The descendants makes sense, anything but peaceful. Definitely through human deception. Not like mass genocide but probably something very clown world.
most goals of the decision makers is to 1) get away 2) live forever.
I bet technology gets to the point of singularity way before 2050. Exponential growth is scary.
SeanDoe1z1
Naxxramas 1.0 Raider
- 7,677
- 19,551
Well.
Machines don’t think in time. Really,
Even physical mediums I have a hard to conceptualize.
I know I could never talk to a human on a moon without the use of a computer. Maybe I could dedicate My life to making a letter out of rocks
does a computer know it cannot talk to another computer on the moon without the use of a human?
I guess what’s where I sit. I know my limitations. Who tells the computer theirs?
- 2
- 28,297
- 44,742
Yeah, hes comparing AI to music model which is pretty niche and has real world ramifications (cost of professionally recording music). So yeah , if youre an aspiring musician like there is millions of and trying to record something you have made, you have every incentive to get a DAW on your computer, learn it and start recording your shit. Its been a dream of musicians to do so for ever. Shit I remember us getting a 4 track back in the 90s to record our shit when we were jamming. I cant even imagine what we would do if we had a DAW and recorded our shit digitally back then.
As far as plebs and AI, this is not the case. There is no incentive to download some AI locally and ask it questions about what you could cook that night or other stupid shit like composing a letter to your boss, when you can just pull out your cell phone and simply ask it your questions, for free. Maybe a smaller company that needs some AI could do this though, to save money? Maybe in shit like logistics, or keeping the books?
ToeMissile
Pronouns: zie/zhem/zer
- 3,719
- 2,508
**my kids are being jerks right now so the below might be a little scattered**
Inference/api costs can get crazy pretty fast, you definitely have to keep an eye on it. For most people it's going to be much easier/cost efficient/? to just use whatever package/service google/anthropic/? are putting together. Claude Cowork and Gemini are adding features to move in the direction of being more like an actual assistant or being able to handle scheduled tasks.
The stuff people are doing with openclaw is pretty great for small businesses. This is timestamped to a dude talking about automating a bunch of tasks for his parents tea (imports?) business with physical 2 stores and an online b2b.
- 1
