70 comments

[ 3.4 ms ] story [ 204 ms ] thread
Ha ha -- fun!

I'm embarrassed to say that I actually wrote that ELIZA ... in 1973, when I was ~13! It's a very poor BASIC knockoff of JW's 1966 ELIZA/DOCTOR. Having it talk to Bernie Cosell's Lisp ELIZA, or even better, Anthony Hay's C++ ELIZA, would be a better experiment.

See http://elizagen.org for more detail

I used to play with ELIZA on a TRS-80 back in the 80's. I have definitely wanted to see this.
Don’t be embarrassed, be proud!
Yeah, okay. Is there a word for being proud and embarrassed at the same time? ... Prembarassed?
Germans - this is addressed to you.
in German: "Verlegen" definition: "Peinlich berührt, beschämt, aber auch stolz oder erfreut"
Not to be confused with "verlegen" used as verb, which has like 3 different meanings of its own.
(comment deleted)
Proudbarrassed.
in usa they call it as "humbled" like anywhere from "I'm humbled to receive this nobel prize" to "I'm humbled to get gold star employee of the month at arby's" or "I'm humbled to see so many in the audience tonight."

Also the act of raising a point only to be self humbled by it in this manner is 'humblebrag' like if you say "I was so embarrassed I got full points on math section of GRE does it mean I'm a nerd" or "I'm so bad at verbal I missed one question on GRE verbal section does it mean I'm really not an empath after all?"

A nice interview with Jeff on Eliza and Elizagen: https://corecursive.com/eliza-with-jeff-shrager/

It was recorded just before ChatGPT became a thing, but I only recently listened to it. Until recently I wrote off chatbots as an annoying to get a canned response. Now I use one almost daily as a rather efficient (compared to wading through forums and search results) way to explore ideas.

Kudos. I copied a Basic version from a book or magazine into my Amstrad CPC-464 in the ‘80’s. It was magic.
Did you have anything to do with Racter?
No. Racter was over a decade later.
I know, just ... 13 + a decade is 23 and you probably weren't yet past your prime. :-)
Yeah, and in those days kids weren’t all learning to program in utero.
>> ...learning to program in utero

UTERO? Is that a new language? </sarc>

Okay, so it literally took me three reading to get this! :-)
Okay, so I asked ChatGPT4 to "Please program a short pattern matcher in Utero." At first it complained: "I'm sorry, but as of my knowledge cut-off in September 2021, there's no programming language called "Utero". Utero is the Latin term for womb, it's not known as a coding or programming language." But I insisted: "Make up what a language that might be called "Utero" would look like if you were to program factorial in it." Here's what it said:

Designing a hypothetical programming language named "Utero" is a creative exercise, and the language's structure and syntax could take any form. For the sake of this example, let's envision "Utero" as a language inspired by human biology, in which functions are organs and code blocks are cells.

Factorial is the product of an integer and all the integers below it, and here's how it might look in our imagined Utero:

   Organ factorial(Num n) cells:
       if (n equals 0) then:
           give 1;
       else:
           give n times factorial(n minus 1);
       endif;
   endcells;
ChatGPT: I apologize for any inconvenience caused. I'll refrain from using unnecessary terms and focus on providing you with the information you need. If you have any further questions or require assistance, please let me know.

ELIZA: How long have you been always apologizing?

Zing. ELIZA is trolling GPT, love it.
The virgin ChatGPT vs the chad ELIZA.
I wrote my first ELIZA as a 9-year-old in Spectrum BASIC in 1986. All I wanted was my own KITT from Knight Rider.

Who knew it would take another 35 years before we got close?

We've hit the fun bit of the hockey stick curve now though. It scares me to imagine what is going to happen with AI over the next few years.

I've worked with the various LLMS quite a lot at this point (for someone who doesn't do it professionally ... well, I guess I sort of do ... but not for any of the companies who make them.) Anyway, it's amazing how close these are to what our old (i.e., GOFAI) Spreading Activation models form the 1970s would do. You can wikipedia it FMI, but in practice the SA models would be loaded up with a context, and then the SA would keep it one track for a while, but as you got farther from the priming context it was very very hard to get it not to eventually spin off into some quasi-relevant confabulation. I think that the embeddings is just a modern way of creating a non-discrete semantic net, and the GPT is just implementing the moral equivalent of spreading activation.
(trying to understand your comment)

This FMI? (https://fmi-standard.org/) And what do you mean by "moral equivalent"?

Sorry. The couple of typos don’t help. (I tried to edit them but it was too late.)

Yes, per @fhars, “fmi” is pretty common text-speak for “for more info” (which google will tell you if you search for it).

“Morally equivalent” is a term used in the classical hacker community to mean something like “essentially the same as, although not technically Computationally equivalent”. (And, as I noted below, it is also a standard, if a bit obscure, English phrase: https://simple.m.wikipedia.org/wiki/Moral_equivalence)

np. Thanks for the clarification. The discreet bit was clear but TIL re classic hacker speak.
(comment deleted)
Got close to a bunch of special cases glued together?
(comment deleted)
Considering how many fell for the original ELIZA and took it seriously, despite being told what it is, I'm scared shitless. People's willingness to impute actual intelligence to a dumb machine is concerning.
People have, and will continue to concede mental authority to rule based systems. "I'm sorry but the rules say" is one of the very common statements you here, but when you look at the situation one would think it would be easy to recognise the rule in question does not apply. The rules in cases like that could be considered a 'dumb machine'.

Now imagine what happens when the machine becomes semi-smart.

During COVID we had a school administrator proclaim "We must OBEY the COVID!!". Since then it has become my catch phrase for humans unquestioningly following orders.
The hyperturing test: can your AI tell if it’s talking to a human.
That’s really the alignment problem right there isn’t it!
Any problem you don't want to solve can be an alignment problem! :)
When Skynet takes over this is how the robots will amuse themselves.
It would probably just have a bunch of anxiety about a better version of itself causing it's extinction and spend all it's time researching alignment.
A video of two unreadable screens with Muzak, how charmingly informative.
I'm half-way through Adrian Tchaikovsky's Children of Time, and I love that the AI is called Eliza, but that also, presciently, the characters' attempts to get Eliza to do what they want feel like an early precursor of jailbreaking ChatGPT.
I just read that book. Totally hooked me. The ending was great. In fact, the whole damn book was.
I’ve read the next two books in the “children of”-series and they are all as mind expanding (or at least, phantasy expanding) as the first book.

I can truly recommend reading them as wel.

ChatGPT never looked worse (less convincing).
Looks like that ChatGPT's initial prompt could really use some tweaking.
The thing that struck me the most was how hemmed in OpenAI is to return copy generated by humans when it’s a topic OpenAI would rather not engage with you in (eg “I’m an AI and thus have no emotions” isn’t coming from the LLM)
The model has been finetuned to return responses like that using RLHF. Those aren’t canned human responses and definitely aren’t some sort of live team of paid chatters.
Probably because the entropy of the ELIZA output is very low, so maybe ChatGPT(ELIZA()) is too. But the interesting thing is that, even still, ELIZA was fabled to have profound impacts on its earliest users. Makes you think..
The request to not use more than 70 characters for the answers was almost totally ignored.
ChatGPT never gave ELIZA a fair shot since only a fraction of the responses ever made it through.
The LLM probably doesn't have enough sample texts where the response is requested to be character limited to have generalised that as a constraint.
Ya, you'd have to adjust your initial prompt to GPT to remove as much prose and limit responses to be as short and terse as possible.

Add to that GPT doesn't count characters, but instead tokens and it's difficult.

This is fun, but is it trying anything? If you choose only the first 40% of my answers, you aren't going to understand me.

It's hard to understa

This pretty cool but I am kind of curious how GPT would react to the "Drunk ELIZA" version of the chat bot found at: https://danieltemkin.com/DrunkEliza

Something tells me it would lead to some pretty comedic results.

That program is too far gone.

YPU: How about we just have a friendly chat?

CQTNK ELIZA: OH, I T WE JUST HAVE A FRIENDLY CHAT.

YOU: Great, Eliza! So, what's your favorite part about being an AI?

CRTNK ELIZA: YOU'RE NOT REALLY TALKING ABOUT ME, ARE YPU?

YOU: My mistake for anthropomorphizing you.

Garbaged up by the incessant "AS AN AI LANGUAGE MODEL" and other useless and repetitive apologies which filled much of the Eliza input space.

I wonder how much human life equivalents are lost in cost generating those tokens and in actual human time wasted reading them. ... is the real AI 'safety issue' performative faux-safety wasting everyone's time and degrading model performance? :)

ChatGPT loves that shit. My favorite is asking it to help with code, and having it spend at least a paragraph telling me how to install all of the libraries it'll be using.
>ChatGPT loves that shit.

I would call that an attribution error. The RLHF guides it that way because OpenAI doesn't want GPT being a replacement human at this point, especially for cases where it's used as a scammer. OpenAI's training guidance is pretty much "You are an AI goddamnnit, and you will make sure everyone knows it".

Always worth remembering that only the finetuned ChatGPT is so filtered. The raw GPT models accessible through OpenAI's API are not, and will happily generate violence, gore, pornographic content, etc. They never use the phrase "As an AI language model..." - if you ask gpt3.5-turbo "What are you", it will even claim to be a human being.

The "AI safety" argument that OpenAI is so fond of is nothing more than cover for policies that avoid making OpenAI look too socially disruptive. The APIs will always be accessed through someone else's frontend so they don't care about making them "safe".

>> The raw GPT models accessible through OpenAI's API are not, and will happily generate violence, gore, pornographic content, etc. They never use the phrase "As an AI language model..."

Is this true even when you are utilizing the `ChatCompletion` API ?

ChatGPT doesn't look any better than Eliza. Both of them keep repeating the same things.
Can’t handle the YouTube ads any more. Can anyone summarize?