Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Starver]

"The moon[sic] landing on July 20, 1969, was a historic and well-documented event, watched by millions of people around the world." x2.  (As the clearest indicator, but there are other lesser examples.)

Spoiler: More an AI Thread discussion (click to show/hide)

A chatbot should ideally be able to be formal (or of whatever style it is intended to emulate) yet unrepetitive/varying in its phrasing. Even if it's not so sophisticated as to continue with some past knowledge of its interactions and at least go somewhere like "As already mentioned, millions of people around the world watched the Moon landing, on July 20, 1969, making it a well-documented historic event."

Even the BBC Basic version of ELIZA that I typed in, back in the early '80s, had the necessary flexibility of phraseology in its limited (but effective, so long as you tried to forget all the data-strings you'd tapped in) range of cut'n'paste responses with occasional insertions. ("> It's my (Mother/Father/Dog)" => ": Tell me about your WHATEVER$." / ": Did you come to me to talk about your WHATEVER$?" / "What is it, about your WHATEVER$, that concerns you?" ...or very roughly like that.)

Top-slot politicians (as alluded to) give themselves very little very variation, because when they do let themselves wander away from the down-pat response to a given prompt they tend to find themselves in a minefield of their own making (as do bottom-slot ones, but with the option of them either then sinking without trace or riding an unexpected zeitgeist wave upwards), in the face of a determined interviewer, but the AIs can surely develop (and use) a wider range of "prepartee" to avoid looking so scripted (in the wrong sense).

It's certainly necessary to overcome the obvious suspicions of a correspondant in a Turing Test scenario. And, for that, a vital metric for the combative counterpart of any Generative Adversarial Network to itself develop the detection of (for downgrading insufficient variability and thus promote the alternative), long before it gets to the point of release to the public.


(PS, after I wrote that (and started looking for the actual BBC Basic version of ELIZA) I came across this blogpost that combined the politician and my own Eliza-nostalgia in one! Probably was a related (but converted) port of the micro-Eliza! )

[dragdeler]

So... anybody know where one can get a chatbot? Not use like own it and run it feom your own machine? I have this feeing that I need to backup a usable version before they become too commercialised, ad filled and useless... This might just be peak chatbot.

[anewaname]

So, let's say some million people start a "let's troll chatGPT about the moon" and puts KittyTac's questions into their own words. Will chatGPT ever question "is my own data suspect?"

[jipehog]

try github chat bot

[dragdeler]

IDK how we wound up talking about this in the balloon thread, but yeah I'm actually excited about chatbot powered search engines, except that as soon as bing dethrones google it will start sucking, mark my words... So yeah that "trained up to 2019" gpt3 or plus or whatever they keep mentioning seems like a fine and knowledgeable assistant to me.

[KittyTac]

"The moon[sic] landing on July 20, 1969, was a historic and well-documented event, watched by millions of people around the world." x2.  (As the clearest indicator, but there are other lesser examples.)

Spoiler: More an AI Thread discussion (click to show/hide)

A chatbot should ideally be able to be formal (or of whatever style it is intended to emulate) yet unrepetitive/varying in its phrasing. Even if it's not so sophisticated as to continue with some past knowledge of its interactions and at least go somewhere like "As already mentioned, millions of people around the world watched the Moon landing, on July 20, 1969, making it a well-documented historic event."

Even the BBC Basic version of ELIZA that I typed in, back in the early '80s, had the necessary flexibility of phraseology in its limited (but effective, so long as you tried to forget all the data-strings you'd tapped in) range of cut'n'paste responses with occasional insertions. ("> It's my (Mother/Father/Dog)" => ": Tell me about your WHATEVER$." / ": Did you come to me to talk about your WHATEVER$?" / "What is it, about your WHATEVER$, that concerns you?" ...or very roughly like that.)

Top-slot politicians (as alluded to) give themselves very little very variation, because when they do let themselves wander away from the down-pat response to a given prompt they tend to find themselves in a minefield of their own making (as do bottom-slot ones, but with the option of them either then sinking without trace or riding an unexpected zeitgeist wave upwards), in the face of a determined interviewer, but the AIs can surely develop (and use) a wider range of "prepartee" to avoid looking so scripted (in the wrong sense).

It's certainly necessary to overcome the obvious suspicions of a correspondant in a Turing Test scenario. And, for that, a vital metric for the combative counterpart of any Generative Adversarial Network to itself develop the detection of (for downgrading insufficient variability and thus promote the alternative), long before it gets to the point of release to the public.


(PS, after I wrote that (and started looking for the actual BBC Basic version of ELIZA) I came across this blogpost that combined the politician and my own Eliza-nostalgia in one! Probably was a related (but converted) port of the micro-Eliza! )

You know how GPT models work, right? They don't have preset phrases. They have a repetition penalty setting. In ChatGPT it's set rather low I suppose. And it's less meant to imitate a human and more as a question-answering machine. Which it does superbly if the question is not overly technical and doesn't involve precise numbers.

BTW it can indeed take previous questions into account but if you bombard it with trolling it kinda goes into defense mode and puts disclaimers at the top of every post until you change the subject or reset the conversation.

So... anybody know where one can get a chatbot? Not use like own it and run it feom your own machine? I have this feeing that I need to backup a usable version before they become too commercialised, ad filled and useless... This might just be peak chatbot.

KoboldAI is the best available option. Do your own research for that one, never ran it locally. ChatGPT is a private model, it could get leaked eventually but consumer computers aren't strong enough to run something like that yet.

So, let's say some million people start a "let's troll chatGPT about the moon" and puts KittyTac's questions into their own words. Will chatGPT ever question "is my own data suspect?"

It doesn't learn unless manually trained by having that data fed back into its dataset, which is a lengthy process and isn't done after every prompt (and the data is checked over to prevent the scenario you are describing). So no.

[jipehog]

I'm actually excited about chatbot powered search engines, except that as soon as bing dethrones google it will start sucking, mark my words... So yeah that "trained up to 2019" gpt3 or plus or whatever they keep mentioning seems like a fine and knowledgeable assistant to me.

How would 'chatbot' improve search engines?

I haven't thought about it much, but from the top of my head replacing most human customer service agents and commercial spam farms would be its primary application.

[Lord Shonus]

In theory, chatbots could improve search engines by providing a more complete understanding of what the user is searching for. Google has been trying to do just that for some time now, with the result being a severe degradation in the quality of their output.

[dragdeler]

Look at the latest wan show if you don't want to sign a thousand eulas and let MS on your smartphone. Luke opted into the trial, and what we've seen on the show was nothing short of impressive... I grew up in the golden age of the internet and just spent since october researching how to come up with my own micromobility solutions, google has become so bad, I'm sure a chatbot frozen in 2019 would answer all my questions pertaining to "school knowledge" in a fifth of the time.

[dragdeler]

In theory, chatbots could improve search engines by providing a more complete understanding of what the user is searching for. Google has been trying to do just that for some time now, with the result being a severe degradation in the quality of their output.

You really think the fucking resume answer on top of some google searches is ML generated? They're direct quotes from the source that's posted just below, and to confidently state, their search results have been degrading because of ML and not humans putting their fatfingers into a perfectly functional searchengine for lefthanded attempts to boost engagement, requires a "source?" from me.

I do not accept that alexa and chatgpt3 are qualitatively the same thing.

[nenjin]

I'm actually excited about chatbot powered search engines, except that as soon as bing dethrones google it will start sucking, mark my words... So yeah that "trained up to 2019" gpt3 or plus or whatever they keep mentioning seems like a fine and knowledgeable assistant to me.

How would 'chatbot' improve search engines?

I haven't thought about it much, but from the top of my head replacing most human customer service agents and commercial spam farms would be its primary application.

The current model is that you ask a question, the search engine indexes everything it can against that query and then provides you results YOU have to scan, digest and synthesize.

The proposed benefit of AI-assisted searches is that it does all that scanning and synthesizing for you. So instead of going through the top 5 results on Google, reading them and comparing them, the bot does that and comes to the "best" answer it can based on the available data.

You're implicitly trusting the bot delivering the "best" answer. And it's really just reporting what seems to be most commonly true. Assuming trusting a machine to provide the best answer works for you, there's a lot of time and mental energy savings in there.

Also, it has the ability to drill down better and faster.

For example: "How long do you cook a 5 pound turkey?"

Bot provides an answer. Then it can guess the next couple things you'll probably ask.

"How to cook a turkey so it's not dry."

"How to cook a turkey so the skin is crispy."

It can provide a drilldown into deeper layers of knowledge rather than YOU having to know that's the question you should ask next, and having to go through the effort of requerying, re-reading and re-sorting that additional info.

[jipehog]

In theory, chatbots could improve search engines by providing a more complete understanding of what the user is searching for. Google has been trying to do just that for some time now, with the result being a severe degradation in the quality of their output.

My thoughts exactly, but then again I have little experience with Alexa and the likes, which I hear becoming much more popular in every day uses in USA

For example: "How long do you cook a 5 pound turkey?"

Bot provides an answer. Then it can guess the next couple things you'll probably ask.

"How to cook a turkey so it's not dry."

"How to cook a turkey so the skin is crispy."

That could be illegal. It is one thing to simulates human-like conversations it is another to use others content (recipes) consequently depriving them from traffic and revenue, recently google had a huge legal fight and settlement over just such thing with news aggregation.

[dragdeler]

Seems about par for the course, no wonder I'd rather talk to the computer if human leeches are now trying to copyright protect sentences that must have been uttered by 95% of all grandmas in the world...

To get back at the earlier topic: search engine optimisation will inherently water down the quality of search results, leeches will try very hard to serve you adriddled shit instead of a pertinent answer... But like, google knows, and if anything they have been very coy about it, we've all seen the degradation.

[Starver]

So, let's say some million people start a "let's troll chatGPT about the moon" and puts KittyTac's questions into their own words. Will chatGPT ever question "is my own data suspect?"

No? It's been trained and frozen with any possible data being from no later than ~2019, and I don't think they're going to leave the door open to it being another Tay...

(That's from an external perspective. The argument that chatGPT is ever open to philosophising is restricted to similar views about whether Lemoine's statements about LaMDA are valid.)

IDK how we wound up talking about this in the balloon thread,

I know I didn't start it, and tried not to make it the main shift of conversation... Ah well. Everyone seems happy (or at least happily contributing).

You know how GPT models work, right? They don't have preset phrases. They have a repetition penalty setting. In ChatGPT it's set rather low I suppose.

My point. The example exhibits startlingly little sign of 'free and fluid reply', given the reputation it has. Unless it's just Muggles who can't see the rigid framework behind the magic, but I don't flatter myself that I'm a wizard in this world.

[jipehog]

Seems about par for the course, no wonder I'd rather talk to the computer if human leeches are now trying to copyright protect sentences that must have been uttered by 95% of all grandmas in the world...

If someone invested time and money to establish a website for recipe sharing and cooking community they are entitled for protection from google AI crawling through their site and copy their hard work and present it as their own ..

Right now google have partnership with several content providers, if you google 'How long do you cook a 5 pound turkey?' you'd likely get promoted content with link to Allrecipes, but what nenjin suggest with several sites that is a few step further.. and if you think on what is possible with use of transcripts in youtube, I would definitely call it predatory practices and be pro expanding laws like those passed recently in EU.

As for you, you are more than welcome to create and share a DB of your grandma wisdom online for free, but expecting that others content should be treated as such stinks of entitlement.