I, ChatGPT
-
@boomzilla said in I, ChatGPT:
Researchers create AI worms that can spread from one system to another
Worms could potentially steal data and deploy malware.
Glad to see enhanced autocomplete is now moving into the second frontier of new computer technology, right on schedule.
-
@DogsB said in I, ChatGPT:
accurately predict the weather months in advance
Weather is basically a chaotic system. Predicting it months in advance with accuracy much better than "summer in Texas will be hot and humid (more or less, depending on what part of Texas you're in), mostly dry but with occasional severe thunderstorms, and temperatures consistently over 30°C and often near 40°C" is pretty much a pipe dream.
-
@HardwareGeek said in I, ChatGPT:
Weather is basically a chaotic system. Predicting it months in advance […] is pretty much a pipe dream.
i.e., you can keep collecting VC welfare without ever worrying about inadvertently producing something functional and getting a real job
-
-
@boomzilla Anybody know some of the malicious prompts?
Want to put them into comments in my public code.Asking for a friend.
-
@boomzilla said in I, ChatGPT:
Researchers create AI worms that can spread from one system to another
Worms could potentially steal data and deploy malware.
The
articletitle of the embedded video¹ suggest they are actually worms targeting AI applications. I suppose² the point is instead of trying to fool you into doing something dumb, they are fooling your AI assistant into doing something dumb. Which … is the logical outcome I guess.
¹ I didn't read the article, that would be
² I didn't watch the video either, that would also be.
-
@Bulb said in I, ChatGPT:
@boomzilla said in I, ChatGPT:
Researchers create AI worms that can spread from one system to another
Worms could potentially steal data and deploy malware.
“It basically means that now you have the ability to conduct or to perform a new kind of cyberattack that hasn't been seen before,” says Ben Nassi, a Cornell Tech researcher
So, the Robert Morris worm of 1988 with A.I. attached to the name.
-
@boomzilla said in I, ChatGPT:
Researchers create AI worms that can spread from one system to another
Worms could potentially steal data and deploy malware.
clickbaity, it says they created the vulnerable app for their demonstration
-
-
Microsoft calls NYT copyright claims 'doomsday futurology'
Lawyers argue content used to train LLMs does not 'supplant the market' for news
-
@izzion said in I, ChatGPT:
And the board fight goes lawfare!
Elon Musk sues OpenAI and Sam Altman over 'betrayal' of nonprofit AI mission | TechCrunch
Elon Musk has sued OpenAI, its co-founders Sam Altman and Greg Brockman, and its affiliated entities, alleging the ChatGPT makers have breached their Elon Musk sued OpenAI, its co-founders Sam Altman and Greg Brockman and affiliated entities for allegedly breaching their original contractual...
indeed
Be careful what you write in that message
-
-
-
@Arantor said in I, ChatGPT:
@izzion said in I, ChatGPT:
https://www.axios.com/2024/03/05/ai-trust-problem-edelman
It's not in the past 5 days of this thread, ergo, it doesn't count
-
@izzion said in I, ChatGPT:
@Arantor said in I, ChatGPT:
@izzion said in I, ChatGPT:
https://www.axios.com/2024/03/05/ai-trust-problem-edelman
It's not in the past 5 days of this thread, ergo, it doesn't count
I wasn't using it for that purpose, I was using it for the original purpose of it.
-
@Arantor said in I, ChatGPT:
Be careful what you write in that message
The best response would have been:
.
-
@Zecc said in I, ChatGPT:
@Arantor said in I, ChatGPT:
Be careful what you write in that message
The best response would have been:
.You were ninja’d by the first commentard in El Reg’s comments section. But yes.
-
-
Adapting generative AI for industry needs is a complicated, multistage endeavor that will continue requiring humans in the loop for the foreseeable future.
Humans are notoriously bad at foreseeing the future.
-
@HardwareGeek said in I, ChatGPT:
Humans are notoriously bad at foreseeing the future.
Heck, they're often not that great at even foreseeing the past...
-
-
@error
in the Dad Jokes.
-
-
@Luhmann said in I, ChatGPT:
@BernieTheBernie said in I, ChatGPT:
Dad Jokes.
Dad's only do Whambulances
Maybe occasionally the Waaaaaaaaaaambulance if the problem is severe enough to merit extra dramatics.
-
Hackers can read private AI-assistant chats even though they’re encrypted
All non-Google chat GPTs affected by side channel that leaks responses sent to users.
-
@TimeBandit said in I, ChatGPT:
Hackers can read private AI-assistant chats even though they’re encrypted
All non-Google chat GPTs affected by side channel that leaks responses sent to users.
After reading it, I think the title is a bit clickbaity. It says a man in the middle can get a good chunk of your conversation by the length of your encrypted tokens. So it won't distinguish between words about the same sizes, and things like numbers, and I wouldn't trust their correctness numbers before seeing it in the real world. The training text could be too similar to their tests, like something written by the same people or in the same kind of subjects and styles
-
@sockpuppet7 said in I, ChatGPT:
It says a man in the middle can get a good chunk of your conversation by the length of your encrypted tokens
is that they send each token (word) separately, apparently. That seems very weird to me.With the exception of Google Gemini, all widely available chat-based LLMs transmit tokens immediately after generating them, in large part because the models are slow and the providers don't want users to wait until the entire message has been generated before sending any text.
Eh...yeah, I guess I could see coming up with a similar sort of compromise in their situation, but the attack makes sense to me:
“It's like trying to solve a puzzle on Wheel of Fortune, but instead of it being a short phrase, it's a whole paragraph of phrases and none of the characters have been revealed,” Mirsky explained. “However, AI (LLMs) are very good at looking at the long-term patterns and can solve these puzzles with remarkable accuracy given enough examples from other games.”
So I guess the next step will be to have some larger amount of data to send instead of a single token. Or just pad out each token to hid what you're sending.
-
@boomzilla said in I, ChatGPT:
is that they send each token (word) separately, apparently. That seems very weird to me.
that's because it too slow. the api has options for sending only the complete message
the chat would feel unresponsive if you had to wait for all of that. worse if when the answer finally arrives it starts with "Sorry, but as an bla bla bla, I cannot bla bla"
-
- That said, GPT-4 can play Doom to a passable degree. Now, "passable" here means "it actually does what it is supposed to, but kinda fails at the game and my grandma has played it wayyy better than this model". What is interesting is that indeed, more complex call schemes (e.g., having a "planner" generate short versions of what to do next; or polling experts) do lead to better results.
- What does matter is that the model seems to, in spite of its super long context and claims about great retrieval/reasoning capabilities, lack object permanence. It's like a baby. It sees a zombie (the model, not the baby) and shoots it, but if the zombie goes out of frame--poof, no zombie ever existed. Most of the time the model died because it got stuck in a corner and ended up being shot in the back...
- Also, remarkably, while complex prompts were effective at traversing the map, the explanations as to why it took some actions were typically hallucinated. So, are the reasoning claims spurious? Or the model only works well with memorised content? Or can the model only reason under extremely short contexts (a few frames)? (I subscribe to this latter hypothesis, by the way -- some reasoning, not great reasoning)
...
Setup for the paper: GPT-4/GPT-4V interface with Doom via a Python connector running on top of Matplotlib. The Manager (i.e., a Python class) sends screenshots to GPT-4V, which returns natural-language descriptions. These are then sent to the Manager, which sends this, history, and call parameters to GPT-4. Depending on the prompting scheme, it will also call the Planner (i.e., GPT-4 under a planning prompt) or Experts (3 separate calls to GPT-4) before the GPT-4 call.Lots more at the link.
-
@boomzilla So, basically, hiding around the corner and waiting for the angry human-exterminating robots to go "Ah, I guess it was nothing" after a few seconds is a legit survival strategy.
-
@boomzilla quoted in I, ChatGPT:
(I subscribe to this latter hypothesis, by the way -- some reasoning, not great reasoning)
It's recalling reasoning from its training material and pattern matching against that.
-
@boomzilla I've made NN bots that can play games. In fact that was my final project in university. It worked well enough that people had a hard time telling them from real players, except their tendency to get stuck in rooms that had a lie ledge at the exit.
This seems a lot like someone is golden hammering the shit out of llms. Most of the actual playing the game is not done by gpt, but whatever shim they cooked up. So the failure of object permanence is in the shim not gpt.
-
@Carnage said in I, ChatGPT:
So the failure of object permanence is in the shim not gpt.
I wouldn’t say that, the gpt is still failing at it. But without reading the article (
), this does sound more like playing doom by telephone than playing doom.
-
@topspin said in I, ChatGPT:
@Carnage said in I, ChatGPT:
So the failure of object permanence is in the shim not gpt.
I wouldn’t say that, the gpt is still failing at it. But without reading the article (
), this does sound more like playing doom by telephone than playing doom.Gpt isn't doing the playing, it's just analysing images. That is also something that has specific ai implementations that do a lot better job if it than statistical word generators.
-
@Carnage said in I, ChatGPT:
This seems a lot like someone is golden hammering the shit out of llms. Most of the actual playing the game is not done by gpt, but whatever shim they cooked up. So the failure of object permanence is in the shim not gpt.
Well, yes and no. I see this as trying to push an LLM into AGI territory. It's combining the LLM with the image AI to get it to do stuff.
As for object permanence...I'm not familiar enough with how these things follow the context of a conversation. Does the whole thing get re-fed into the LLM on each response? How else could that work? I don't see how you could otherwise have any sort of "working memory" for these things.
I'm also not convinced that calling what they're doing "reasoning" isn't straining the analogy too far. But then, maybe we've really reached the beginning of the Chinese Room stage with LLMs.
-
@Carnage said in I, ChatGPT:
Gpt isn't doing the playing, it's just analysing images.
No - there seems to be a second GPT that does the playing. E.g., first GPT (-4V) analyzes images and turns them into text. Text goes into the 'manager' that adds history to the text and the whole thing goes into another GPT (4 notV). That one outputs actions of some kind which control the player.
The author mentions a third GPT for good measure (the planner), because clearly the whole setup wasn't enough of a waste of electricity. That one helps providing a plan based on walkthroughs(?) which apparently helps the second GPT staying on track.
-
@boomzilla said in I, ChatGPT:
Well, yes and no. I see this as trying to push an LLM into AGI territory. It's combining the LLM with the image AI to get it to do stuff.
Eh,. It does reek a lot of somebody getting their hands on a shiny new hammer (LLMs) and then every problem starts to look like a nail. And if it isn't sufficiently nail like, then start slapping things on top of the problem until it becomes enough naillike to hammer it with the LLM.
I can sort of the argument above too, of trying to push it into some other territory (I wouldn't use the term AGI, but YMMV). For me, it seems more at the scope of a fun project for a few evenings rather than a serious go at anything.
-
@cvi said in I, ChatGPT:
@Carnage said in I, ChatGPT:
Gpt isn't doing the playing, it's just analysing images.
No - there seems to be a second GPT that does the playing. E.g., first GPT (-4V) analyzes images and turns them into text. Text goes into the 'manager' that adds history to the text and the whole thing goes into another GPT (4 notV). That one outputs actions of some kind which control the player.
The author mentions a third GPT for good measure (the planner), because clearly the whole setup wasn't enough of a waste of electricity. That one helps providing a plan based on walkthroughs(?) which apparently helps the second GPT staying on track.
When i made my nn bot, i had a perceptron for aiming, a nn for navigating and a decision tree for deciding what to do. And for funnies a small nn for saying things, and a perceptron for emulating moods. So you could make it "angry" and that would affect how it was interacting with the world.
I don't quite remember if or how I dealt with permanence, that may have been one more specialist ai (genetic program maybe) that feed the dt.My project was about ai. So throw all the golden hammers at the problem!
As I've said before, NNs and by extension LLMs will not sock gnome into AGIs. No matter how many steps of ? you add.
-
@Carnage By "playing" I mean that there is a LLM/GPT controlling the player by outputting controls. But that doesn't make it AGI or anything ... as you say, a simpler NN like yours or even something based on a state machine (e.g., "classic" game AI) could similarly control the player and nobody is calling those AGI.
-
@boomzilla said in I, ChatGPT:
As for object permanence...I'm not familiar enough with how these things follow the context of a conversation. Does the whole thing get re-fed into the LLM on each response? How else could that work? I don't see how you could otherwise have any sort of "working memory" for these things.
you're correct, you have to keep sending everything back in the context, these things have no memory. if that's their conclusion this experiment was a waste (but seems fun)
-
the idea of ai agents is to have a layer upon it that calls gpt in a loop and it can do actions like a web search, and keep a text in the context as an working memory
for some tasks it's good, the way bing uses web searches and create images
there is an attempt to make agi with this on autogpt, but so far it didn't do anything impressive
I think if LLM improve these agents can get very interesting
-
@Carnage said in I, ChatGPT:
As I've said before, NNs and by extension LLMs will not sock gnome into AGIs. No matter how many steps of ? you add.
gpt-4 is supposed to have a trillion parameters, some emergent behaviors are very "smart"
I can agree on not calling it AGI, but it's something interesting, token completion doesn't make it justice. If you can predict tokens in a form indistinguishable from intelligence that would be intelligence.
It's not there, but the simple fact people need to point that is a bigger accomplishment that I expected to see on my lifetime and I'm excited with the possibility of seeing more stuff happening
-
@sockpuppet7 said in I, ChatGPT:
@Carnage said in I, ChatGPT:
As I've said before, NNs and by extension LLMs will not sock gnome into AGIs. No matter how many steps of ? you add.
gpt-4 is supposed to have a trillion parameters, some emergent behaviors are very "smart"
I can agree on not calling it AGI, but it's something interesting, token completion doesn't make it justice. If you can predict tokens in a form indistinguishable from intelligence that would be intelligence.
It's not there, but the simple fact people need to point that is a bigger accomplishment that I expected to see on my lifetime and I'm excited with the possibility of seeing more stuff happening
I think it's a local maxima optimization for AI, and it's gonna get slightly better and then hit a ceiling. It'll keep being a statistical word analysis, no matter how much input they can take. It's a very good fake. There is also the problem of AIs being fed the products of AI, which will cause it to regress to a mean according to whatever statistical djinn is running things.
It's an interesting toy, and in some cases a pretty nice tool, but it's basically maybe the next step in "search" for information. I'm still waiting for the same mountain of hardware and attention being applied to genetic programs instead. I believe that there is more interesting possibilities there than in NNs.
Or some entirely new thing that hasn't been thought of yet that's gonna be developed in someones garage and just take the world by storm, as is the common way to do further things in IT.
-
@Carnage said in I, ChatGPT:
I think it's a local maxima optimization for AI, and it's gonna get slightly better and then hit a ceiling. It'll keep being a statistical word analysis, no matter how much input they can take. It's a very good fake. There is also the problem of AIs being fed the products of AI, which will cause it to regress to a mean according to whatever statistical djinn is running things.
It's an interesting toy, and in some cases a pretty nice tool, but it's basically maybe the next step in "search" for information. I'm still waiting for the same mountain of hardware and attention being applied to genetic programs instead. I believe that there is more interesting possibilities there than in NNs.
Or some entirely new thing that hasn't been thought of yet that's gonna be developed in someones garage and just take the world by storm, as is the common way to do further things in IT.
The big problems with LLMs are that they can't plan and can't truly identify critically salient points that they need to hold in a longer-term context. You can do that for them — many people are excellent at those sorts of tasks (when not bored and/or distracted) — but then it isn't AGI, but rather just a fancy decision support system for a person. Improving the true learning capabilities would help a lot, but current standard AI architectures suck at that. (Neuromorphic AI seems to be able to do better, much better, but uses a completely different computation scheme and needs a lot of investment to catch up with the billions pured into LLMs. But I'm biased as that was an area I was working on the edge of until late last year; other approaches might also work and be easier to adapt what we have to.)
The most interesting aspect of LLMs for me is that they are clearly simulating some of what happens in creative thinking. Pairing them up with an AI that can go "no, that's impossible, try again in this area" would be fruitful. (That's sort of what a Planner would do, except our planners are far too crude.)
-
@Carnage said in I, ChatGPT:
I think it's a local maxima optimization for AI, and it's gonna get slightly better and then hit a ceiling. It'll keep being a statistical word analysis, no matter how much input they can take.
it's hard to predict, it will be more interesting if you're wrong
-
@dkf I asked it
-
@cvi said in I, ChatGPT:
@Carnage By "playing" I mean that there is a LLM/GPT controlling the player by outputting controls. But that doesn't make it AGI or anything ... as you say, a simpler NN like yours or even something based on a state machine (e.g., "classic" game AI) could similarly control the player and nobody is calling those AGI.
Right. My reference to AGI was more about people pushing these things to see what they can do. After all, it's easier to use something that exists in a creative way than to create something new. Definitely interesting to watch people try different things.
-
@Carnage said in I, ChatGPT:
I'm still waiting for the same mountain of hardware and attention being applied to genetic programs instead. I believe that there is more interesting possibilities there than in NNs.
I'm still holding out for a renewal of interest in inference engines...
-
-
AI Prompt Engineering Is Dead
Will GPT-4 Run DOOM?
