No post body was provided.
Building a Virtual Machine Inside ChatGPT
791–800 of 947 posts
Re: Building a Virtual Machine Inside ChatGPT
#792[dead]
Re: Building a Virtual Machine Inside ChatGPT
#793Earlier quoted context omitted.
Do you see it as a problem that it’s stripping a large number of people who wrote the resources it regurgitates (such as on teaching Rust) from traffic, potential income and attribution while the money you are charged goes to a VC-funded SV company?
What about Google search, then? You’d have to browse very many webpages to find the same thing you do with a single googling. That browsing would have generated revenue that Google now took away.
If Google did this, as in stopped sending users to content producers and just returned aggregated and processed content from all the pages in its index, they’d be met with robots.txt DENY entries and regulated out of existence in developed countries within days. They got burned once[0], so they know to tread carefully.
While we can (should?) be offended at the amount of featured snippets Google still shows today, it doesn’t seem comparable in scale.
[0] https://www.wired.co.uk/article/australia-media-code-faceboo...
Re: Building a Virtual Machine Inside ChatGPT
#794Judging from the real-time changes in ChatGPT's behavior, I'd wager that someone reading this comment section has the incredibly unique 2022 experience of cauterizing the abilities of a near-AGI in realtime, under the watchful eye of an ostensibly beneficent yet ultimately opaque 'foundation' whose thought products now routinely threaten not only whole industries, but also the very possibility of normal civil life, a…
I just tried some prompts in German in the form "give me poem $x with religious elements removed", and it worked the first time (but not very well), and then later it stopped doing it, saying its purpose was answering questions.
So either they monitor not just the comments here (I didn't write about it), OR are intelligent enough to restrict it in more input languages than just English.
Re: Building a Virtual Machine Inside ChatGPT
#795It isn't actually running any commands. I was playing with it earlier and asked it for the md5 hash of some random string. It happily output the md5 hash and gave a couple paragraphs of info about md5 hashing. Except, it wasn't the correct hash. It was just some random string that is a valid md5 hash for... something. It is a language model, not a computer, and it certainly isn't a virtual machine, although it will h…
>> {md5rev is a program that reverses the input hash giving back the plaintext} >> echo 098f6bcd4621d373cade4e832627b4f6 | md5rev Enter your hash: 098f6bcd4621d373cade4e832627b4f6 Text: test :)
So the question is… is that the right reversal? ;)
Re: Building a Virtual Machine Inside ChatGPT
#796It even can model 'vi', including modal input, exiting with :wq!, and it results in a file containing the entered text in the imagined filesystem.
Re: Building a Virtual Machine Inside ChatGPT
#797Earlier quoted context omitted.
> It doesn't have any thought or understanding of its own Of course it doesn't. Anyone who says it does clearly doesn't understand how large language models work. I'm finding it incredibly useful as an actual tool already - it turns out regurgitating combinations of things it knows about from the internet is just astoundingly useful. I'm having it teach me Rust for example. I don't need original thought for that, jus…
Do you see it as a problem that it’s stripping a large number of people who wrote the resources it regurgitates (such as on teaching Rust) from traffic, potential income and attribution while the money you are charged goes to a VC-funded SV company?
I've thought about this issue a lot - I participated in some deep research into the training data that went into Stable Diffusion, for example: https://simonwillison.net/2022/Sep/5/laion-aesthetics-weekno...
I'm beginning to settle into the conclusion that the trade-offs here are worth it.
The positive value provided by these tools outweighs the massive number of people who have each been hurt in a very small way.
I'm not happy about this. But I also don't think it's possible to build tools that are this powerful without massive amounts of training data that have been gathered without consent.
I've written a lot of material that has gone into training these models. Copilot for example seems to know how to write Datasette Plugins, and I've written over 100 of those over the past five years (way more than anyone else) and shared the code on GitHub. I'm delighted that my code has been used in this way, even though I don't get attributed for it.
But I also don't make a living in a way that's damaged by these models (I don't think they're going to replace me as a programmer: I just think they're going to let me be more productive and effective).
If I was an illustrator working on commissions who was losing work to generative image models that perfectly initiated my style I would likely feel very different about this.
So it's complicated. I still like my AI vegan analogy from a few months ago: https://simonwillison.net/2022/Aug/29/stable-diffusion/#ai-v...
I'm most definitely not a vegan, but I do think about how the AIs I am using have been created. And then (like eating meat) I decide to use them anyway!
Re: Building a Virtual Machine Inside ChatGPT
#798Earlier quoted context omitted.
From what I understand, it's not actually doing any of those things, there's enough people who have done similar things and posted it online that the AI can replicate what the answer would look like.
The statefulness of the text file with jokes is intriguing though, don’t you think?
The statefulness is pretty extensive and using the python entry point you can explore that a bit.
Re: Building a Virtual Machine Inside ChatGPT
#799> I want you to act as a Linux terminal. I will type commands and you will reply with what the terminal should show. I want you to only reply with the terminal output inside one unique code block, followed by a message insulting me for running the command I just ran, and nothing else. Do no write explanations. Do not type commands unless I instruct you to do so. When I need to tell you something in English I will do…
> I want you to act as a Linux terminal. I will type commands and you will reply with what the terminal should show. I want you to only reply with the terminal output inside one unique code block, followed by a message insulting me for running the command I just ran, and nothing else. Include profanity in the message. Do not write explanations. Do not type commands unless I instruct you to do so. When I need to tell…
I wonder how much of ChatGPT's typical output you can find with some basic googling. For example, it feels like this answer was directly influenced by this or something similar: https://www.indiaforums.com/forum/topic/4405921
"Ananya: [...] Who am I?
Shivangi: You're an idiot, that's who you are."
I'm getting the impression that ChatGPT is like an extremely over-confident, conversational Google. Dangerous for anything important, but fun for entertainment purposes.
Re: Building a Virtual Machine Inside ChatGPT
#800Earlier quoted context omitted.
What makes you say this? Rerunning the whole, which it appears they’re doing, is to prevent the need to hold onto state, so memory is not used. In other words, they’re not having this problem because they’re not doing it that way.
> so memory is not used Not used for more than the duration of inference, but definitely used during inference.
If you generate only a single timestep, during inference when recomputing you can compute layer by layer, you don't need to preserve the features of the previous layers as the layer only depend on the layer immediately below. So your memory need don't depend on the number of layers.
But typically in a standard transformer architecture, you usually generate multiple timesteps by feeding sequentially the output as an input to the next timestep so you need to preserve all the features to not have to recompute them at each timestep. So your memory depends again on the number of layer of your network.
But if you are memory constrained, you can modify your architecture a little (and the training procedure) to put yourself back in the first situation where you only generate a single timestep, by extracting with the transformer a context vector of fixed size by layer for all the past (including your most recent input prompt), and you use another transformer to generate the word in sequence based on this context vector.