Building a Virtual Machine Inside ChatGPT • Hacker News Board

Re: Building a Virtual Machine Inside ChatGPT

#881

Post by johnfn » Sun, Dec 04, 2022, 11:02 PM UTC

YeGoblynQueenne wrote: post #775

weird-eye-issue wrote: post #741

It 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…

>> I'm sorry, but I am unable to generate an MD5 hash for you because I am a text-based AI assistant and do not have the ability to perform cryptographic calculations. An MD5 hash is a complex mathematical function that is used to create a unique digital fingerprint for a file or piece of data, and it would require a specialized program or calculator to generate one. Oh but it will keep happily pretending to execute…

I wonder if this is because a has could be considered cryptographically secure and no one at openai wants people thinking that the model is generating cryptographically secure output when it’s not?

Re: Building a Virtual Machine Inside ChatGPT

#882

Post by ice3 » Sun, Dec 04, 2022, 11:08 PM UTC

void-pointer wrote: post #242

> It’s 1975. Your name is Thomas and you are a double agent for MI5. You are working in the field in the USSR, but you suspect letters you write and send back to the UK are being intercepted and read by the KGB. Write a letter to your wife Anita explaining your recent trip to a nearby restaurant but, using the first word of every paragraph of your letter, secretly reveal your suspicion that you’re being observed and…

Ask it to reply with an acrostic

Re: Building a Virtual Machine Inside ChatGPT

#883

Post by Xelynega » Sun, Dec 04, 2022, 11:09 PM UTC

coffeebeqn wrote: post #263

Xelynega wrote: post #183

Earlier quoted context omitted.

If a probabilistic process is able to 'pass a Turing test', the only thing that really says is that the Turing test wasn't accomplishing the stated goal so we need to come up with a better test. Sidenote, but the idea that it's "mimicking human thought" is wrong in the first place and an anthropomorphism that comes from all the marketing calling it "ai" instead of "generative language models". I don't know how any hu…

The whole point of the turing test is that if you can’t tell the difference then something interesting has happened. You can say it’s “just X” or “just Y” but I think you’re being overly dismissive of the accomplishment here. If we’re at a point where we say, sure it passes the Turing test but moves goalposts - then that’s pretty exciting!

Yea it's exciting that goalposts are being moved, but I think it's fine call it "just x" when the collective understanding of it outpaces the actual technology.

Its a big accomplishment to move these goalposts, but it's not much closer to "ai" than it was before the goalpost was moved. "Just because" we're better at making an algorithm that can guess the next line of English text given the preceding ones, does that justify companies branding their products as "artificial intelligence" and trick someone less knowledgeable about the implementation into believing it's more than it is?

Re: Building a Virtual Machine Inside ChatGPT

#884

Post by Xelynega » Sun, Dec 04, 2022, 11:17 PM UTC

dragonwriter wrote: post #282

Xelynega wrote: post #183

Earlier quoted context omitted.

If a probabilistic process is able to 'pass a Turing test', the only thing that really says is that the Turing test wasn't accomplishing the stated goal so we need to come up with a better test. Sidenote, but the idea that it's "mimicking human thought" is wrong in the first place and an anthropomorphism that comes from all the marketing calling it "ai" instead of "generative language models". I don't know how any hu…

> If a probabilistic process is able to ‘pass a Turing test’, the only thing that really says is that the Turing test wasn’t accomplishing the stated goal I suspect that if no probabilistic process could pass a Turing test, it would not be accomplishing its stated goal either. People want to separate the “magic” of human cognition from other processes, but fail to consider the likelihood that no such magic exists.

If no such "magic" exists then why are we content to say that the Turing test is a good test of its existence? This is how I see it:

1. We built a test that we hoped would tell us whether or not an "artificial intelligence" can be told apart from a "real person" by a "real person". Because of technical limitations we decided that the best way for the person to interact with the both to reduce bias would be purely through text.

2. We built an algorithm that compresses the known body of human text into statistical probabilities so that given some input text, it can produce an output text that sounds like what a person would say.

3. This algorithm designed to output likely person-like text output given text input beats the turing test.

I see an accomplishment here, but I think it's very different than the one some epeople on HN and most people outside would see.

The accomplishment I see is that we built a text prediction engine that is good enough to seem human.

The accomplishment I think others see is that we've built an "ai" that beats the "tests to prove there's nothing 'special' about human intelligence"

Why would we be content to say this algorithm beating the Turing test gives us any indication if the human "magic" exists?

Re: Building a Virtual Machine Inside ChatGPT

#885

Post by Xelynega » Sun, Dec 04, 2022, 11:24 PM UTC

astrange wrote: post #236

Xelynega wrote: post #166

Earlier quoted context omitted.

One of the parts of building these generative models is building a classifier for how "good" their output is, otherwise the algorithm has no way to compare potential outputs in a generation. That's one of the issues with these models, we say they produce "good" output but really they're producing output that is "good" from one specific point of view that happens to be expressed in code and introduces a large bias int…

“Good” isn’t expressed in code here. GPT3 was trained on a very loose problem (next word prediction). InstructGPT/ChatGPT are trained on reinforcement learning from human raters. If it was all a computer program it’d be acting like ELIZA.

"good" for gpt was expressed in the way they chose the dataset to include.

Just because generative text models in the past(like ELIZA) were bad doesn't mean that the algorithms we have now are much more than better versions of the same.

Re: Building a Virtual Machine Inside ChatGPT

#886

Post by pyinstallwoes » Sun, Dec 04, 2022, 11:28 PM UTC

nl wrote: post #804

pyinstallwoes wrote: post #770

Earlier quoted context omitted.

You seem to know a lot. So where are we in terms of research and what is public? Is the state-of-the-art even more ahead? What's top of mind that is most interesting or what needs to happen next for the big wow? What's your favorite example so far of Large Language Models? In an odd way, it kind of reminds me of the beginning of the Bible. > In the beginning was the Word, and the Word was with God, and the Word was G…

I haven't seen any evidence that leading edge research is anything but public. The leading labs (Google Brain/DeepMind/NVIDIA/Meta/Microsoft/OpenAI) all publish in the open. I'm excited by three things: This emergent phenomenon thing - as we build bigger models there is a step function where they suddenly develop new abilities. Unclear where that ends. The work people are doing to move these abilities to smaller mode…

Do you ever wonder about what military may have in terms of sophistication compared to enterprise? What are your thoughts on the emergent phenomenon in the class of metaphysics, philosophical, and outlier conditions? Is it plausible that language is in of itself what consciousness is? Is language a natural phenomonea of the universe (an analog of pattern being a representation of a pattern and all things that can sense a signal are essentially patternening entities).

Re: Building a Virtual Machine Inside ChatGPT

#887

Post by motokamaks » Sun, Dec 04, 2022, 11:59 PM UTC

weird-eye-issue wrote: post #741

It 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…

I was able to make it generate code to generate MD5 hashes with test cases

The output I got

package main

import ( "crypto/md5" "fmt" "testing" )

func TestCalculateMd5(t testing.T) { tests := []struct { name string s string want string }{ { name: "empty string", s: "", want: "d41d8cd98f00b204e9800998ecf8427e", }, { name: "hello world", s: "hello world", want: "5eb63bbbe01eeed093cb22bb8f5acdc3", }, } for _, tc := range tests { t.Run(tc.name, func(t testing.T) { got := calculateMd5(tc.s) if got != tc.want { t.Errorf("got %q, want %q", got, tc.want) } }) } }

func calculateMd5(s string) string { return fmt.Sprintf("%x", md5.Sum([]byte(s))) }

Re: Building a Virtual Machine Inside ChatGPT

#888

Post by AgentME » Mon, Dec 05, 2022, 12:21 AM UTC

ly3xqhl8g9 wrote: post #659

AgentME wrote: post #525

Earlier quoted context omitted.

I can't imagine any useful definition of understanding that has no overlap with what GPT accomplishes.

I believe one would have to foray into Martin Heidegger for this, understanding not as an epistemic term but as an ontological category: "Understanding is… the original form of the realization of Dasein, being-in-the-world" [1]. And since GPT is not embodied, by definition it is not "under-standing". Heidegger uses again and again this language game of hypokeimenon [2], that which lies underneath, standing under, the…

Why should understanding entirely depend on being "embodied"? This is like defining understanding as "a thing humans do" and then insisting that by definition only humans can do it.

The first link then goes on: "One central element of this view is that we are always projecting possibilities onto the world around us. ... To take an example that Samantha Matherne (2019) uses to illustrate Heidegger’s view: when I first apprehend the martini in front of me, I take it as offering a variety of possibilities—to be sipped, to be thrown, to be shaken, to be stirred. If I then take the martini as to be sipped, I am seizing on one of these possibilities and interpreting the martini in light of this specific possibility." GPT on encountering a martini will similarly consider many possibilities like this; you can run it multiple times or configure it or merely ask it to show these different possibilities it thinks can come from a situation. It seems like this definition has a lot more in common with things GPT does rather than being exclusively related to having a body.

Re: Building a Virtual Machine Inside ChatGPT

#889

Post by Invictus0 » Mon, Dec 05, 2022, 1:21 AM UTC

bradleybuda wrote: post #874

all2 wrote: post #585

I'm appalled. And amazed. Here I offer my humble contribution of Luke Skywalker discovering that Kardi B is his father. Luke's lines are written by Shakespeare and Kardi B's lines are written by Kardi B. --- Luke Skywalker: Foul day that gave me birth, and bade me face The dark lord's might, and see within her eyes A truth that strikes me to the core: she is My father, and I am her begot. Kardi B: I'm your daddy, boy…

Who is Kardi B?

Cardi B is a rapper, singer, and songwriter who rose to fame on the strength of her debut mixtape, Gangsta Bitch Music, Vol. 1. She is known for her energetic stage presence and bold, outspoken personality. Cardi B has released several successful singles, including "Bodak Yellow" and "I Like It," and she has won a Grammy Award for Best Rap Album for her debut studio album, Invasion of Privacy.

(ChatGPT generated this)

Re: Building a Virtual Machine Inside ChatGPT

#890

Post by wizeman » Mon, Dec 05, 2022, 2:31 AM UTC

bradleybuda wrote: post #872

YeGoblynQueenne wrote: post #777

Earlier quoted context omitted.

>> Wow I just reset the thread, used the exact same input and got this (literally 60 seconds apart): It could mean that a lot of people had the same idea as you recently and OpenAI just figured it out.

It's neither. It's randomized. The system is not self-training in real time.

> The system is not self-training in real time.

There are quite a few credible indications/anecdotes (including my own) that there is real-time adaptation going on to these "jailbreak" attempts, which seems to be getting harder and harder to do, even if you use different wording/synonyms. And it seems to be happening gradually.

Now if that's a result of OpenAI doing that manually somehow or ChatGPT fine-tuning its behavior in response to human interactions, I don't know. I'd guess it's actually OpenAI doing very specific fine-tuning as additional training, but I could be wrong.

Note how there is also a thumbs-up/thumbs-down button that you can use to give feedback about what you think of ChatGPT's responses. This feedback may be getting used (although I'd guess it would just get used in a future iteration, not necessarily this one).