The Garbage In Garbage Out Nature of Machine Learning and AI

[npetrele]

ChatGPT AI may be groundbreaking, but for the meantime, AI itself will rarely if ever make any groundbreaking scientific discoveries. That's because AI like ChatGPT's "machine learning" curates existing information based on statistics. Show it enough pictures of bicycles and it will "machine learn" to recognize a bicycle in a picture based on what it determined was statistically relevant. It doesn't tell you how a bicycle works based on the images it sees. It has to get that information elsewhere. What it will do is scrape descriptions of how bicycles work and through the magic of statistics, curate a reasonable response.

I can't believe it's not butter

Let's apply that method to something like medical knowledge to identify the limitation. AI won't "machine learn" a medical fact. It will "machine learn" the current popular perspective on scientific data and repeat it back to you. 

Consider the question, "Is butter bad for you?" Today, ChatGPT answered:

"Saturated Fat: Butter is high in saturated fat, which has been associated with an increased risk of heart disease. High saturated fat intake can raise LDL (bad) cholesterol levels in some individuals. However, recent research has challenged the link between saturated fat and heart disease, and it is now recognized that the relationship is more complex than previously thought."

Here's a portion of the answer to "Is margarine bad for you?":

"Traditionally, margarine was made using partially hydrogenated oils, which contain trans fats. Trans fats are known to increase the risk of heart disease, raise bad cholesterol (LDL), and lower good cholesterol (HDL). Therefore, consuming margarine with trans fats can be detrimental to your health. However, in recent years, many margarine manufacturers have reformulated their products to reduce or eliminate trans fats."

If ChatGPT was around a couple decades ago, it would have warned you about the high cholesterol in butter and recommended margarine as an alternative. In other words, it would have given you bad advice, the same bad advice doctors and food companies gave at the time. Just 20 or so years ago the popular trend was that margarine was the heart healthy alternative to butter. Garbage in, garbage out. 

Nowadays, ChatGPT has been updated to reflect that margarine wasn't good for you at all back then, but it has gotten better since. Incidentally, margarine in the USA can no longer include trans fats, a detail missing from the ChatGPT response. Regardless, this is progress, but the progress was made outside of AI. The AI simply absorbed it and repeated it back. There's no guarantee that the information you get from ChatGPT is accurate or beneficial. It will simply follow the popular opinion. And if that popular opinion is wrong, garbage in, garbage out.

Sunny Side Up

Let's tackle a more controversial subject. 

Question: Is the sun a solid body made up of condensed matter?

"No, the Sun is not a solid body made up of condensed matter. It is primarily composed of hot, ionized gas, in a state known as plasma. The Sun is a massive, luminous sphere of plasma held together by its own gravity. It consists mainly of hydrogen (about 74% by mass) and helium (about 24% by mass), with small amounts of other elements such as oxygen, carbon, neon, and iron."

Question: Are black holes real?

"Black holes form from the remnants of massive stars that have undergone gravitational collapse. When a massive star exhausts its nuclear fuel, it can no longer withstand the inward pull of gravity. The core of the star collapses under its own gravity, and if the mass is sufficiently concentrated, it can form a black hole."

Watch a dozen or so videos from Sky Scholar, a youtube channel run mainly by Pierre-Marie Robitaille. Mr. Robitaille makes (IMO) very compelling arguments that dispute the standard solar model. He instead provides ample evidence that the sun is indeed made up of condensed matter, not plasma. He presents spectral analysis as evidence. He points out the errors in the mathematics, such as where a Newtonian concept of gravity is applied where only one body exists which violates the basis for Newtonian physics. And he deals with thermodynamics at length, demonstrating the error of mixing intensive and extensive properties in current formulas. I'll let you watch his videos and decide for yourself how much of what he says is right or wrong. 

But for the purpose of argument, let's assume he's right. The consequences of him being right are massive. If he's right that stars are made up of condensed matter, then there is no basis for believing black holes exist. If there's no such thing as a black hole, then a huge amount of astrophysics must be thrown out and new ideas brought in. 

Maybe someday we can teach AI to explore these topics and propose alternate hypotheses. But right now, all ChatGPT does is scrape the popular opinions and regurgitate them. So if Pierre-Marie Robitaille and his colleagues are correct, ChatGPT by definition is a prime example of garbage in, garbage out. 

(Coincidentally, Alexander Unzicker just released a book, "The Liquid Sun: A Coming Revolution in Astrophysics", which I'm told supports Pierre-Marie Robitaille's views. I ordered the book but haven't yet received it. If it catches on, perhaps ChatGPT will give us different answers sometime in the future.)

[npetrele]

Here's a paper I find fascinating, related to cosmology and the big bang theory.

https://www.researchgate.net/publication/343707905_The_myth_of_the_Big_Bang_and_the_redshift_controversy

[npetrele]

I recently watched a YT video that asked a great question. If there is ever a super-AI that surpasses human intelligence, will we ever find out? It could be smart enough to play dumb so that we won't pull the plug. 

[npetrele]

By the way, I still don't use margarine. I went back to butter some 30 or more years ago before the "consensus" about butter and margarine caught on. I didn't have the data back then, but I had a healthy distrust of what TV ads told me. 

[npetrele]

.I received my book Alexander Unzicker's, "The Liquid Sun: A Coming Revolution in Astrophysics". So far its a page turner, VERY interesting and not too complicated for the average reader. I highly recommend it. You can get it on Amazon as a Kindle or paperback.

[Joseph W. Doherty]

Came across your post, just being recently posted.

Not too long ago (last several weeks) someone brought up AI, obviously very impressed by something like ChatGPT.  I mentioned I obtained an account and took ChatGPT for a short "test drive".  I noted, I found the parser (of my input) and data information very, very impressive.  But as to "intelligence", I found the machine equivalent of a "savant".  I.e. I did not see any evidence of "thinking" and/or "reasoning".

[npetrele]

Funny you should mention that. I happen to be an idiot savant. Okay, well, that's actually only half right.

[Joseph W. Doherty]

Okay by me if you're only half right.  In my case, I once thought I made a mistake, but I was mistaken,  ; )

Also, BTW, "idiot savant", supposedly, is not in favor because not every savant is actually an idiot.  The preferred words now seem to be "savant syndrome" (at least according to Wiki).

BTW, I haven't heard the expression "Garbage In Garbage Out" since the last century.  Perhaps because "I went back to butter some 30 or more years ago", means your not a 20 something kid (laugh).  Or, perhaps it's still heard in the programming world (of which I was part of, also decades ago), but not so much in the networking world.  (In the networking world, we worry more about how fast we can deliver garbage.)

Much I didn't touch upon in my prior reply.  I too well remember the butter vs. margarine "flip flops" over the years.  I also remember, the "coming ice age"; and many other "flip flops".  (Those under 40 wonder why those over 40 tend to be "cynical" - ah, possibly something related to real life experience?)

I too agree much AI relies on what it has been fed.  At the time I tried ChatGPT, I had the passing thought, if its references declared the moon was swiss cheese, that's what it would "regurgitate".

I've often suggested, AI isn't new, we've probably had it since time immortal, only the "computer" version of it is relatively recent.

Perfect example, decade ago, a company brought me in for a day to review a possible programming issue.  The director asked what I might need to insure I might resolve the issue while I was there?  I said, well, it's "inconvenient" that to obtain 6 CPU seconds to compile a program, generally takes overnight.  Director said, would top priority job submissions help?  I said, sure.  Director asked, how many do you need?  I said, 15 should be more than enough.  So the director gave me a job priority form for 15 job submissions.  I took it down to the operations desk, and they said, you cannot have more than 5.  Having experience with AI, I asked, what if I provided you 3 forms, same job name, each for 5?  Operations said, that would be okay.  So, I went back to the director, got 3 forms, for 5 each, and went back to operations.  Since the job name was the same on all 3 forms, I watched as this information was entered into the system; i.e. approval for 15 job submissions.

[npetrele]

I just visited spaceweather.com and saw this:

In case you don't see the screen cap, it says:

"This is an AI Free Zone! Text created by ChatGPT and other Large Language Models is spreading rapidly across the Internet. It's well-written, artificial, frequently inaccurate. If you find a mistake on Spaceweather.com, rest assured it was made by a real human being."

[Joseph W. Doherty]

Earlier today, I was reading an article linked off msn.com about AI.

Article was touting all the amazing thing ChatGPT has done, and might do.  All the white collar jobs it might replace.  However, ChatGPT was also noted, it occasionally gets things wrong!  The latter doesn't appear to get as much press.

On the other hand, as I wrote in an earlier post, we already have much AI in the world, so its certainly possible, machine AI will replace much human AI too, much as the industrial revolution replaced much human labor.  Likely, machine AI will be less expensive.

Where machine AI might do might better than human AI, when I was introduced to ISO 9000 certification, I asked about actual improvements to a process, such as elimination of errors, and was told, that's unimportant.  The certification was for consistency.  I asked, even if consistently wrong?  I was told, that's correct.  IMO, machines generally are more consistent, so I can see machine AI being better for such certifications.  ; )

[npetrele]

Still reading Alexander Unzicker's, "The Liquid Sun: A Coming Revolution in Astrophysics".  I can't recommend this book enough. If you have any interest at all in cosmology and astrophysics, this is a must read. I can almost guarantee you'll leave the book convinced that stars are made up of condensed matter, most likely liquid metallic hydrogen. He also demolishes (disproves) Kirchhoff's law. Unzicker also talks a lot about Pierre-Marie Robitaille other contributions to science, including creating an MRI technique that all his colleagues said would be impossible. He abandoned that lucrative career in MRI in pursuit of correcting the junk science that passes for cosmology these days, which I find personally admirable. 

[npetrele]

Great Max Planck quote from the book: "Science advances one funeral at a time".  If the standard solar model isn't verifiable, then maybe it's time to kill it and move on. 

[Joseph W. Doherty]

"Science advances one funeral at a time".

Unfamiliar with Max Planck's quotes, but I wonder, whether he meant, as I believe, one of the biggest impediments to scientific advancement is the "experts" in any field of study.  Often, it seems, those who spent a lifetime becoming an "expert" are the most closed minded to anything that contradicts (the then current scientific) orthodoxly.  It seems, such experts need to die off, i.e. their funerals, and then follow on generations seriously consider ideas not part of the current scientific orthodoxly.

"If the standard solar model isn't verifiable, then maybe it's time to kill it and move on."

Well, if something cannot be verified, but also it cannot be proven to be wrong, i.e. it's possible, I wouldn't say you need to "kill it", but conversely, no reason you cannot continue to find a "better" possible explanation, ideally one that seems verifiable.

For example, somewhere along the line in my education of physics, I recall (?) I was told, Newton was "wrong", but his equations are still a great approximation in many, many cases.

[npetrele]

I don't know if that's what he meant, but your interpretation is the one I favor. In fact, the book I mentioned deals exactly with that subject. People (experts) get educated one way, and pride prevents them from considering alternatives. 

I think Mr. Robitaille has sufficiently falsified the standard solar model. His alternative, a liquid metallic hydrogen sun, still needs further investigation before it can be considered a sufficient replacement, but at least it makes more sense. It is supported by light spectrum analysis, surface features of the sun, and laws of thermodynamics. That isn't proof, but it's a great explanation that deserves more study.