ChatGPT

[rmazzullo]

I asked some (what I thought were) tough emuslion making questions in chatgpt.com, and got some interesting answers. Go ahead...try it. Ask it anything.

-Bob

 

[koraks]

got some interesting answers

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.

 

[blacksquare]

Exactly. ChatGPT and other AI systems are now mainly improved search engines and one must approach it that way.

 

[MFstooges]

The difference is that ChatGPT will not tell you if it doesn't know the answer. It will just lie and you wouldn't know it. Be careful of too much relying on it.

 

[rmazzullo]

I never relied on ChatGPT to give me the correct answers. I was surprised it had anything at all to return about emulsion making. The first question I asked was "describe how I can make a 'single run ammonia digest' silver halide negative emulsion" and the second was slightly different: "describe how I can make a 'single run ammonia digest' silver halide photo emulsion." The answers were intriguing - but not complete enough for me to drop what I was doing and shout "Eureka!!" Maybe it should be called ChimeraGPT.

 

[RezaLoghme]

It is a large language model, not a search engine. It can process data that you first have to feed it. Then all depends on your prompts. Small details in both the data you are feeding it and in your prompts matter.

And it is important to contextualize your query.Tell CGPT what you want it to be, what data you are going to feed it with, and what you expect as an outcome.

Most people use it like Google search and are disappointed with the results.

EXAMPLE:

Dont prompt "Should I buy the 50mm Summilux, the 55mm Summimat or the 59mm Summichomoluxmat?".

But:

Upload the lens spec sheets from the manufacturer, tell CGPT that you are a photog with A, B, C use cases and that you want CGPT to act as your purchasing decision making assistant, so it should ask you for more input (budget etc) in order to give better recommendations. Then tell it to disclose its chain of thought so you can see how it got to its results, so you can refine your prompt if needed.

Or feed it with lengthy and detailed product reviews and ask it to make mobile-screen-friendly summaries. Use it for decision-making, for summarizing, for generating text, or text structures (for example when setting up your own website).

 

[TomR55]

The difference is that ChatGPT will not tell you if it doesn't know the answer. It will just lie and you wouldn't know it. Be careful of too much relying on it.

Makes me think of the (paraphrased) quote: “He who knows not and knows not …”, except that this software (and its ilk) is pervasive, instantaneous, and sadly checks so many boxes (it generates extreme wealth, it’s “cool,” etc.) that it will certainly further degrade the human experience.

 

[BrianShaw]

What does this have to do with emulsion making?

 

[MFstooges]

Makes me think of the (paraphrased) quote: “He who knows not and knows not …”, except that this software (and its ilk) is pervasive, instantaneous, and sadly checks so many boxes (it generates extreme wealth, it’s “cool,” etc.) that it will certainly further degrade the human experience.

The funny thing is that it gets defensive when you point that out, just like a real liar.

 

[TomR55]

The funny thing is that it gets defensive when you point that out, just like a real liar.

I recall hearing Marvin Minsky tell someone (ca. mid 1990s) something along the lines of: We will know when a program is approaching human intelligence when "It can lie (and know it), and It can understand a dirty joke, … “.

 

[retina_restoration]

I recently asked it to give me instructions for making a Kallitype print and it made a couple of significant errors that would have completely stymied the attempt. Close, but that 5% error broke the results.

 

[RezaLoghme]

I have explained above how it works. It is not a know-it-all with encyclopedic knowledge. It needs to be fed with data and trained on its behaviour and output.

Anecdotes of asking it to do things you presumably know much better is missing its point.

 

[retina_restoration]

I have explained above how it works. It is not a know-it-all with encyclopedic knowledge. It needs to be fed with data and trained on its behaviour and output.

Anecdotes of asking it to do things you presumably know much better is missing its point.

Yes, but this is how people are using it, and so it doesn't look great when people get seriously erroneous output.

 

[RezaLoghme]

Certainly.

I was just explaining how to use them.

 

[RezaLoghme]

I asked some (what I thought were) tough emuslion making questions in chatgpt.com, and got some interesting answers. Go ahead...try it. Ask it anything.

-Bob

Care to post the questions? Happy to give you some tips how to use a LLM.

 

[rmazzullo]

I originally thought emulsion making would have been an esoteric enough subject to give it a hard time...I got back somewhat arcane answers.

"describe how I can make a 'single run ammonia digest' silver halide negative emulsion" - and the same question slightly different -
"describe how I can make a 'single run ammonia digest' silver halide photo emulsion."

I used those questions - framed as they were - to compare the responses to what we already know. My overall impression is that ChatGPT (and others) is not yet ready
to give my brain a 'run' (pun intended) for its money.

 

[RezaLoghme]

I guess these are far too short. If you want to get usable answers, you have to provide much more context in your prompts, and ideally feed it with some data first. CGPT is neither a search engine nor an encyclopedia.

 

[RalphLambrecht]

I asked some (what I thought were) tough emuslion making questions in chatgpt.com, and got some interesting answers. Go ahead...try it. Ask it anything.

-Bob

I prefer BrutusAI myself; try that!

 

[VinceInMT]

The difference is that ChatGPT will not tell you if it doesn't know the answer. It will just lie and you wouldn't know it. Be careful of too much relying on it.

Sort of a variation of mansplaining.

 

[RezaLoghme]

Isnt mainslaining something different?

Anyway, prompting is an art, and I have tried to build quite complex prompts, e.g. for camera-buying decision making. As mentioned before, garbage-in, garbage-out.

 

[chuckroast]

I asked some (what I thought were) tough emuslion making questions in chatgpt.com, and got some interesting answers. Go ahead...try it. Ask it anything.

-Bob

Try perplexity.ai

It is a high class search engine with a natural language interface. I like it because it provides it sources in the answer so you can read more deeply. Very handy. For example:

https://www.perplexity.ai/search/formula-to-make-printing-out-p-zIZLawAaR2Sw_jeIVz2o1A

 

[MFstooges]

There are instances it gives answers solely based on information in one Internet source. So yes it is also a search engine, a hyped up one, where you "prompt" instead of typing question.
And it doesn't even care to disclose the source of the answers.

 

[RezaLoghme]

Yeah, thats true, and its training has a cut-off date (2021). I prefer to feed it a mess of data and let it chew throught it, that is what it is best at.

 

[MattKing]

This will no doubt show my age, but the various sources of AI have brought forward a fair number of examples of new members posting where the moderation team has had to ask itself: "Is the poster real, or is the post AI generated, or is it a bit of both?"
Sort of a modern version of: "Is it real, or is it Memorex?"

 

[chuckroast]

This will no doubt show my age, but the various sources of AI have brought forward a fair number of examples of new members posting where the moderation team has had to ask itself: "Is the poster real, or is the post AI generated, or is it a bit of both?"
Sort of a modern version of: "Is it real, or is it Memorex?"

I've mentioned this before, but it bears repeating.

So-called "AI" is nothing more than billions of transistors running through a maze of data and weighting rules to come to some conclusions that are more- or less correct.

The problem is that, to trust those conclusions, you have to trust the institutions that created and organized the data and wrote the rules.

And there's the rub. The tech plutocracy - Facebook/Meta, Alphabet/Google, Bing/Microsoft, Apple ... - have already demonstrated a willingness to fiddle their algorithms to favor a particular political position, steer the user in a particular social direction, discredit people and causes they don't like, and so forth. (Pre-Musk Twitter was also a really bad abuser of this, but he pried the lid off that pretty swiftly.) Combine that with a population increasingly mentally lazy and willing to abandon critical thought and instead joining their preferred socio-political-cultural echo chamber, and this spells disaster.

It's one thing to ask an AI to find you the best formula for a Microdol-X replacement, it's quite another to ask it to explain history or the development of Western thought, as just two examples.

I am now old enough to realize that anything that gets too big should never be trusted without lots of critical thought and multiply sourced independent verification - big business, big government, big church, big social movements, and now ... big tech.

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