ChatGPT

[BrianShaw]

So it seems that ChatGPT knows little about:

• Home
• Forums
• Analog Workflow Forums (100% Analog/Traditional)
• Darkroom
• Silver Gelatin Based Emulsion Making & ...

 

[koraks]

I say all this as technology professional who finds the tech fascinating and uses it regularly.

Then you're also aware that the 'algorithms' behind language models like ChatGPT et al. are not ones that are easily modified other than through the selection of training data. The algorithm as such is a black box, even to those who helped make it. This leaves the question to what extent the choice of training data is deliberately tailored by the agents/institution you mention towards specific preferences. This is likely not a yes/no answer, but a massive grey area, and some biases may turn out not to be all that deliberate, but simply the result of opportunistic selection of training data or the same kind of biases and oversight every human is liable to.

Having said this, from the perspective of a moderator I need to remark that exploring the nature of these AI/language models is OK, but hypothesizing (let alone anything that even remotely approaches proselytizing) about the political force of tech companies is likely to be qualified as 'political' and we may/will cut such debate short for this reason.

 

[Steven Lee]

CharGPT answers are always well formulated - but they are only rarely accurate when it comes to more technically involved topics on which online available information is inconsistent or incomplete. ChatGPT is in the end a language model that's trained on existing and relatively easily available (online) data, which makes it particularly vulnerable to the old adage of Garbage In, Garbage Out. Contrary to what the moniker suggests, AI is not actually intelligent in the sense that it can weigh, interpret and assess information.

This assumption is quite widespread and the role of online datasets is often exaggerated. The OpenAI team has always been very much aware of the "online garbage problem" and the datasets used for training are highly curated. I am not too close to them, but I am fairly sure that online forum data wasn't used for training.

But what's interesting is that LLMs aren't good at admitting and clearly saying: I don't know. Instead, they proceed full speed to advice giving. In that sense they behave exactly like a typical photrio or photo.net regular.

 

[koraks]

This assumption is quite widespread and the role of online datasets is often exaggerated.

Well, online datasets in my book include (esp. open access) academic literature, digitized library content etc. The problem remains that the fact that something was once published, doesn't make it good. I'm sure that training on forum data would have resulted in a proper abomination - and in fact experiments with various Twitter (still called that at the time) bots proved this to be true.

But what's interesting is that LLMs aren't good at admitting and clearly saying: I don't know.

Yes, that's one part of the problem.
Another part is that extremely valuable human trait that we understand so little about, but that we use all the time and that proves to be crucial: intuition. One could argue that intuition is nothing more than the human equivalent of an LLM (or vice versa!), but it's surprisingly useful, despite its inherently fuzzy nature.
And then there's more, of course.

It's an interesting development for sure, and in a way we're facing a very similar situation like we did about 100 years ago when advances in nuclear physics started to accelerate (hah).

 

[RezaLoghme]

All technological advancement aside, I wish that forums and message boards will always be populated by humans who contribute with their diverse opinions and viewpoints, and not by AI-fuelled chatbots.

 

[chuckroast]

Then you're also aware that the 'algorithms' behind language models like ChatGPT et al. are not ones that are easily modified other than through the selection of training data. The algorithm as such is a black box, even to those who helped make it. This leaves the question to what extent the choice of training data is deliberately tailored by the agents/institution you mention towards specific preferences. This is likely not a yes/no answer, but a massive grey area, and some biases may turn out not to be all that deliberate, but simply the result of opportunistic selection of training data or the same kind of biases and oversight every human is liable to.

Having said this, from the perspective of a moderator I need to remark that exploring the nature of these AI/language models is OK, but hypothesizing (let alone anything that even remotely approaches proselytizing) about the political force of tech companies is likely to be qualified as 'political' and we may/will cut such debate short for this reason.

Understood and I wasn't trying to go down that rathole. I was only trying to point out that potential for silent, intentional bias exists and has been demonstrated.

 

[chuckroast]

This assumption is quite widespread and the role of online datasets is often exaggerated. The OpenAI team has always been very much aware of the "online garbage problem" and the datasets used for training are highly curated. I am not too close to them, but I am fairly sure that online forum data wasn't used for training.

But what's interesting is that LLMs aren't good at admitting and clearly saying: I don't know. Instead, they proceed full speed to advice giving. In that sense they behave exactly like a typical photrio or photo.net regular.

More broadly, AIs won't seem truly human until they exhibit ALL of our vices by being biased, stubborn, short sighted, ill tempered, intolerant, and generally mean spirited. Only then will they actually pass the Turing Test

 

[chuckroast]

Speaking of which:

Mark V. Shaney - Wikipedia

en.wikipedia.org

 

[koraks]

Understood and I wasn't trying to go down that rathole. I was only trying to point out that potential for silent, intentional bias exists and has been demonstrated.

No worries - and yes, agreed, there are problems with bias and by extension, ethics. Heck, it's a massive expanse of quicksand!

More broadly, AIs won't seem truly human until they exhibit ALL of our vices by being biased, stubborn, short sighted, ill tempered, intolerant, and generally mean spirited. Only then will they actually pass the Turing Test

So far, the big ones are just too polite to qualify!

 

[chuckroast]

No worries - and yes, agreed, there are problems with bias and by extension, ethics. Heck, it's a massive expanse of quicksand!



So far, the big ones are just too polite to qualify!

How much you want to wager that someone is working on the Drunken Uncle personality on top of OpenAI ...

 

[koraks]

How much you want to wager that someone is working on the Drunken Uncle personality on top of OpenAI ...

Not much; it's a bet I'd be bound to lose on. I'm sure our guy Murphy is hard at work on this one, too.

 

[RezaLoghme]

Lets get back to the OP's question. How to use CGTP to actually solve that emulsion questino.

 

[MattKing]

Lets get back to the OP's question. How to use CGTP to actually solve that emulsion questino.

I don't think that the OP asked that.
Instead, the OP related some experience with ChatGPT, and invited comment.

 

[RezaLoghme]

I don't think that the OP asked that.
Instead, the OP related some experience with ChatGPT, and invited comment.

Thank you for keeping an eye on that. Lets see what he says.

 

[snusmumriken]

I thought AI did an excellent job of the Tour de France

 

[loccdor]

I was asking ChatGPT a lot of questions about the history of films a while back. It was getting almost every answer wrong.

Unfortunately people are using it for fake reviews, fake websites now, which is going to make it even harder to find the signal in the noise.

 

[BrianShaw]

ChatGPT (AI in general) is just like some real humans - striving to be a know-it-all dilettante. The goal isn't to be always right but, rather,to always have something to say. The three most feared words: I Don't Know.

 

[MFstooges]

But it does pretty good job on language. If you're multilingual you can converse with it and feel like chatting with a polyglot. I use it for writing macros and arduino program and it never makes syntax error, it actually makes me lazy. No surprise there since bunch of code nerds must have been feeding it. I read that AI can produce code for its own sake just like SkyNet in Arnie's Terminator movie.
I expect it to perform better in science than art. So you probably can pick couple of your favorite films and feed it with developer combo, dev time, temperatur, shooting data and dev results. It will have more comprehensive data than Master Dev Chart. My question is with the free version you have everyone feeding all kind of information including garbage info so how does it filter out the misinformation? Sort of like Wikipedia where everyone can edit the 'history'. Will it grow up as an AI adult who still believes in Easter Bunny?

 

[RezaLoghme]

It is important to understand how it works if you expect it to perform tasks for you. First of all, detailed prompting. Then, understanding that it only remembers what is in the chat, and even that is limited to certain volume. Then you need to give context, what you expect it to behave like, who are you, what is the kind of output and what is the expected format. It cannot read your mind.

You can even copy/paste from Excel but need to clarify what is coming and how the columns are organized. Then it can be very powerful.

 

[RezaLoghme]

I was asking ChatGPT a lot of questions about the history of films a while back. It was getting almost every answer wrong.

Unfortunately people are using it for fake reviews, fake websites now, which is going to make it even harder to find the signal in the noise.

The history of films is probably described elsewhere in detail, e.g. Wikipedia, or some blogpost. Take that, and then play with it, making it summarize or interpret the content, or use it to answer specific questions you might be too lazy to pull out of a lenghty text. That is what works best. Its training had a cut-off date (ca 2021) so never expect it to be able to refer to current events. Also, the more specific your request, the better it is to feed with complicated details and give it the task to filte out what you want from it.

Otherwise its replies are superficial, what is to be expected.

 

[MFstooges]

You need to use it more. Just recently the free version had boosted its memory so it is expanded beyond a chat.
And having something that is supposed to be intelligent but ended up spending your time to understand how to instruct (or to use the AI snobs terminology : "prompting") is not what people expect.

 

[RezaLoghme]

"Prompting" is indeed the right terminology, nothing to do with "snobs". And prompts can be long and complex. You can even let CGPT produce HTML or formatted output.

 

[koraks]

I use it for writing macros and arduino program and it never makes syntax error

Does it produce usable C++ code at this point, at least some of the time?
I regularly see "Arduino sketches" appear that are ChatGPT generated and they are invariably poor in quality. The only good thing I can say about them is that they're at least commented - but the comments are like much of what comes out of CGPT: verbose, but not particularly rich in information. The generated code is poorly structured and usually suffers from fatal flaws. Such as code generated for a particular Arduino board that turned out to heavily rely on direct register access for a totally different microcontroller. Evidently, the person who 'commissioned' the code from CGPT couldn't put heads or tails to it, and indeed, the fastest way to solve the problem was to just start over from scratch.

 

[RezaLoghme]

I dont think so. For Python maybe.

 

[koraks]

I dont think so. For Python maybe.

Yeah, maybe. HTML and PHP may also be feasible at this point.
Mind you, I'm sure things will improve rapidly. The fact that it's capable of producing halfway readable code to begin with is nothing short of a miracle, if you think about it.