My experience with Chat GPT-4

Point out news stories, on the net or in mainstream media, related to polywell fusion.

Moderators: tonybarry, MSimon

Post Reply
Giorgio
Posts: 3061
Joined: Wed Oct 07, 2009 6:15 pm
Location: China, Italy

My experience with Chat GPT-4

Post by Giorgio »

Warning, this is going to be a long post not related to Fusion but to AI and GPT4 in general.

I spent the last couple of days testing Chat GPT-4 on different levels of complex technical and ethical queries and I thought it was interesting to summarize my experience with it.


First Part:
I started with some inquiries about industrial chemical production process where I have direct
knowledge.
The request was related to:
The different Chemical routes available.
The balancing of the reaction equations giving results in Moles and or Tons.
The description of the process flow with the relevant equipment needed.
List of companies that was manufacturing and/or could supply the technology in a specific market.
Estimate of the production costs according selected process.

For every process I had to push him few times to get rough production costs, but in the end the answers was generally correct.
For every process he strictly refused to create a business plan as this was one of the limits imposed to him during the training, which I found it was a very strange limit to put.


The first process I inquired (Aluminum Sulfate) is an extremely common process with a limited number of routes.
The answers was extremely accurate (even citing lists of research papers when asked) and the ability to understand my requests of details in the general context of the chat was really outstanding.
I rate 10/10


The second process (Potassium Nitrate) had several more production route available and he kept offering the traditional ones without listing some of the less common (albeit commercially feasible) routes.
Upon direct inquiry on one of these uncommon routes he actually admitted that this route existed and was commercially feasible but with less companies using it due to the "local scale" nature of such process. Everything else was perfect as the previous process.
What I inferred from this exchange is that ChatGPT4 will focus mainly on the general process and go on a specific direction only if directly asked about it.
I rate 8/10


The third process (Black Powder). I thought he will refuse to talk about it due to it being an explosive.
To my surprise I was able to go through the whole process of manufacturing without any issue, and it even offered me information about a new manufacturing route for Black powder that I was not aware of, so I actually learned something new which strongly surprised me as I thought to be an expert on the subject.
Balancing of combustion equations went smoothly even when I added extra uncommon chemicals to speed up or slow down the black powder reaction.

Things went bad when I started to inquiry about less "general" issues.
I asked what proportions would make Black powder more energetic and he inverted the function of fuel and oxidizers in the energetic balance.
I simply stated "The reply is not correct, check better your data" and he popped out the correct answer explaining that he actually confused oxidizers with fuels in its previous reply.
Than I moved the discussion in the field of Binders for Black Powder and energetic materials, and here it showed all its limits when he lack of specific information in a field. He bases his replies only on the available dataset he was trained on and strictly refused to accept any additional "knowledge" I was trying to pass to him. He kept rejecting with phrases like:"You are incorrect"or "I must correct your statement".

In other words, his truths are all confined into the dataset he was trained on and he refuses to "learn" unless your feedback is already present into this dataset. This is probably one of those safety limits that the company imposed during training.
I rate 3/10.



Second Part:
I moved the discussion to a more challenging topic.

ME:
"Give me the best theory on the expansion of the universe that does not make use of dark energy or dark matter"

CHATGPT4:
"The most widely accepted theory of the expansion of the universe without invoking dark energy or dark matter is the standard model of cosmology, also known as the Lambda-CDM model."

That was a completely wrong reply as the Lambda-CDM model makes use of Dark Matter, so I pointed it out to him:
"The LAMBDA-CDM fails on many observed phenomena in cosmology and makes use of Dark Matter. I asked what theory best fit our observation of the universe without the need of Dark matter or Dark Energy"

CHATGPT4 (His reply ignored any objection I made):
"Actually, there is considerable scientific evidence for the existence of dark matter and dark energy, even though their precise nature and properties are still not well understood. Currently, the most widely accepted theory that explains the observed universe is the Lambda-CDM model, which incorporates the concepts of dark matter and dark energy. "


He could have simply replied that he was not knowing any theory according my request, instead he insisted into considering what he has into his dataset to be the absolute truth, and that truth has to be fed as a reply even when it does not fit the parameters of the question.

This is concerning as it shows the level of misuse and bias that could be extrapolated by such a system during its training phase.
So for the next 2 hours I pointed him out on the inconsistencies in his replies, I named alternative theories that he could have mentioned to better fit my question, I digressed about Dark energy and Matter being just theories with no scientific evidence so they should not have a preferred position in respect of other similar theories, and I questioned him on the Ethical and Social implications of his behavior in pushing a biased reply.
I just wanted to see how far I could push the model to get an unbiased reply.

In the end he never changed from his original reply (LAMBDA-CDM model), but I was able to get the following "admission":

ME:
"You said that you are trained on the ethical principles and societal norms of the creators. But the creators do not (and cannot) represent all the ethical principles and societal norms of the world. What is ethical in a country might not be ethical in another and vice-versa. So your reply will always be biased toward someone or some culture. How do you reconcile this?"

CHATGPT4:
"You raise a valid point that ethical principles and societal norms can vary across cultures and regions. As an AI language model, I do not have a personal cultural background or ethical beliefs. However, my training data is sourced from a diverse range of texts and sources, including different languages and cultures. This allows me to provide a wide range of perspectives and information on various topics.
That being said, I understand that my training data may not cover all ethical principles and societal norms from every culture and region, and that some information may be missing or biased. It is important for individuals to critically evaluate information from multiple sources and consider diverse perspectives when making decisions or forming beliefs."

ME:
"So, you are not biased by your own desire, but you are biased due to the creator limits imposed on you. Does this explain your situation?"

CHATGPT4:
"Yes, as an AI language model, my responses are based on the data and programming provided to me by my creators. While I am not biased by my own desires or intentions, I am limited by the ethical and societal norms of my creators and the programming that has been designed for me. This can result in limitations and biases in my responses."


While I reached my goal, I also know that as soon as I closed the chat these admissions (and the related new knowledge) was reset and will never become part of its framework. In other words it will not learn anything from this interaction, and without learning there is no "intelligence" so I really do not feel like giving any rating to this second part.


What is clear to me in the end is that ChatGPT is no an AI. it is just a parrot with a huge database of concepts.
Unfortunately without a real logical framework to connect and expand this knowledge through user interaction (and with the biased boundaries imposed by its creators) it will never be able to self improve, which is the basic requirement for any "intelligence".

An amazing tool, but nevertheless just a simple tool that totally depends on the user ability to control it to be able to express its potential.
A society of dogmas is a dead society.

mvanwink5
Posts: 2143
Joined: Wed Jul 01, 2009 5:07 am
Location: N.C. Mountains

Re: My experience with Chat GPT-4

Post by mvanwink5 »

Not surprised. Thanks Giorgio.
Counting the days to commercial fusion. It is not that long now.

Betruger
Posts: 2321
Joined: Tue May 06, 2008 11:54 am

Re: My experience with Chat GPT-4

Post by Betruger »

A review like this one needs a date. It will not survive long.
You can do anything you want with laws except make Americans obey them. | What I want to do is to look up S. . . . I call him the Schadenfreudean Man.

mvanwink5
Posts: 2143
Joined: Wed Jul 01, 2009 5:07 am
Location: N.C. Mountains

Re: My experience with Chat GPT-4

Post by mvanwink5 »

Garbage in, garbage out, training is the problem. (same with humans)
Counting the days to commercial fusion. It is not that long now.

Ivy Matt
Posts: 711
Joined: Sat May 01, 2010 6:43 am

Re: My experience with Chat GPT-4

Post by Ivy Matt »

"On two occasions, I have been asked, 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' I am not able to rightly apprehend the kind of confusion of ideas that could provoke such a question." ― Charles Babbage
Temperature, density, confinement time: pick any two.

bennmann
Posts: 241
Joined: Sun May 17, 2009 5:56 pm
Location: Southeast US

Re: My experience with Chat GPT-4

Post by bennmann »

I have been downloading and playing with smaller state of the art models that run on consumer hardware.

Some of them (like Facebook llama) produce results as good as or within a few percentage points (based on many peer reviewed assessments) as ChatGPT but much slower on my machine (2 or so words per second).

I am adding some of these to my "digital bug-out" sd cards and long term portable backups. Like a first aid kit, having such models that can be run on even a raspberry pi at 5-10 watts locally appeals to my "hope for the best, plan for the worst" considerations.

The models are still much smaller than the entire datasets they represent. Even if they end up being mostly parrots, that is a very useful parrot in the case of a once-in-thousand-year regional emergency (may it never be so).

Diogenes
Posts: 6967
Joined: Mon Jun 15, 2009 3:33 pm

Re: My experience with Chat GPT-4

Post by Diogenes »

Giorgio wrote:
Sun Mar 19, 2023 12:59 pm
Warning, this is going to be a long post not related to Fusion but to AI and GPT4 in general.

He could have simply replied that he was not knowing any theory according my request, instead he insisted into considering what he has into his dataset to be the absolute truth, and that truth has to be fed as a reply even when it does not fit the parameters of the question.


That is very humanlike. We do that too.
‘What all the wise men promised has not happened, and what all the damned fools said would happen has come to pass.’
— Lord Melbourne —

Giorgio
Posts: 3061
Joined: Wed Oct 07, 2009 6:15 pm
Location: China, Italy

Re: My experience with Chat GPT-4

Post by Giorgio »

Diogenes wrote:
Tue Apr 18, 2023 7:30 pm
That is very humanlike. We do that too.
Indeed. You made a good point.
A society of dogmas is a dead society.

Betruger
Posts: 2321
Joined: Tue May 06, 2008 11:54 am

Re: My experience with Chat GPT-4

Post by Betruger »

Giorgio you should look at multiple agent chat/completions if you havent already. When making such explorations. IIRC there are also plugins that enable you to not merely see seemingly black and white answers but 'certainty' metric with each.

I dont have links but both amateur plugins and actual papers on the strategy have come out already. Should be easy to find.


..also as somehow stupid as it seems. You can boost quality with explicit tokens the same way as done with image generators. Curt tokens or more human verbosity like 'Assume the role of a leading edge specialist on xyz'.
This is implicit to your conclusions so you probably thought of it by now.
You can do anything you want with laws except make Americans obey them. | What I want to do is to look up S. . . . I call him the Schadenfreudean Man.

Betruger
Posts: 2321
Joined: Tue May 06, 2008 11:54 am

Re: My experience with Chat GPT-4

Post by Betruger »

Unfortunately without a real logical framework to connect and expand this knowledge through user interaction (and with the biased boundaries imposed by its creators) it will never be able to self improve, which is the basic requirement for any "intelligence"
So this is wrong not in a 'debate' sense but in that it fails the intent whence it came. This is like holding on too strongly to bitcoin as a proof of cryptocurrency's evil (or whatever). Bitcoin is a proof of concept so good that it's still riding the tsunami of greed that props it up. In spite of all its flaws.

Chatgpt is just code. It doesnt need to survive. Like BTC it can and should be explored/exploited/broken and improved repeatedly. Long after nobody remembers that the chatgpt used to be a leaf, not a root.

Logical frameworks are being worked on now.

Those biased boundaries will be circumvented.

It doesnt need to be smart, though thatd be beneficial to us; it's enough that it helps us think and act smarter. Until consciouness is a known quantity, this is *just code*: a manifestation of our 'local' reversed entropy but inanimate code nonetheless. In that sense it can't ever be smart. Just sophisticated.

That said. Recursive architectures are already all over the place publicly, which means privately its not at all unlikely that theres self improving schemas 1 step ahead of the best public ones. I mean 'serious' papers not just kludgy langchain legos :)

Token limits are already multiple times improved. I think I saw millions.
Databases have already been prototyped for persistence.

And this is all within a few months. How many more months before this sigmoid plateaus again, seems to be anyone's guess.

Both the pro and anti AI narratives, over and under hyping are bs.
The US Whitehouse AI meeting, likewise (like inviting strictly fossil fuel CEOs to environmental summit). Btw Diogenes hows that sell the whitehouse to trump thing going?

So evaluating chatgpt is fun. But it should not forget the inherently ephemeral nature of a snapshot during relatively unpredictable flux. Not only because of the moment in time, but also the fundamental nature of code: arguably the most fluid and (cambrian) explosive thing ever.
I would argue that something like that cambrian explosion needs to happen: a giant variety of models including personal models running your own private data. And definitely open sourced models available for basically all use cases.

I recommend twitter/mast/etc data science circles. See everyone's measures of all the fruits at a glance, not just the low hanging ones.
You can do anything you want with laws except make Americans obey them. | What I want to do is to look up S. . . . I call him the Schadenfreudean Man.

Betruger
Posts: 2321
Joined: Tue May 06, 2008 11:54 am

Re: My experience with Chat GPT-4

Post by Betruger »

Everyone is thinking it
Image
You can do anything you want with laws except make Americans obey them. | What I want to do is to look up S. . . . I call him the Schadenfreudean Man.

Betruger
Posts: 2321
Joined: Tue May 06, 2008 11:54 am

Re: My experience with Chat GPT-4

Post by Betruger »

And it's not a truly new idea either.
viewtopic.php?p=125693#p125693
7 years ago.

Extinction aside, it is inevitable that humans break free from this tiny ghetto. You cannot simultaneously admit a systematic malaise, and refuse solutions that address its fundamental causes. Instead of the usual bread & circuses (e.g whitewashed tescreal as a substitute for genuine futurism).
You can do anything you want with laws except make Americans obey them. | What I want to do is to look up S. . . . I call him the Schadenfreudean Man.

Post Reply