Guest Post by Willis Eschenbach
My previous post, Artificial Alarmism, has gotten some comments from folks who think I’m wrong, and that Large-Language Models can indeed automate the fact-checking of scientific claims. This was an interesting comment, my thanks to the author:
The kind of AI applications I am talking about do not reject things. They summarize them, including the debates. For that matter I recently got ChatGPT to correctly explain how Happer disagrees with alarmism. Nothing was rejected.
The math I refer to is that used to do science. Almost all published science uses math so it is universal. In the article I reference I use Monte Carlo as an example. There is an advance in Monte Carlo method published in a forest management journal that needs to get to all the other fields that use that method, which are legion.
For that matter you know how Google now suggests related and refined searches. That is AI and it works well.
In response I decided, being a scientist and all, that experiment is much better than theory. So I went to ChatGPT, and it turned out quite funny, for a reason I’ll explain at the end. First, the Q&A, emphasis mine:
Me
Which of these two statements is true.
1) The major control of the world’s temperature is CO2, or
2) the major control of the world’s temperature is emergent phenomena like thunderstorms, El Nino/La Nina alterations, and cyclones.
ChatGPT
Both statements capture important aspects of Earth’s climate system, but they address different levels of influence:
The major control of the world’s temperature is CO2: This statement emphasizes the role of greenhouse gases, particularly carbon dioxide (CO2), in regulating Earth’s temperature. Greenhouse gases trap heat in the atmosphere, contributing to the greenhouse effect and influencing global temperatures over long periods.
The major control of the world’s temperature is emergent phenomena like thunderstorms, El Niño/La Niña alterations, and cyclones: This statement highlights the influence of short-term and regional weather patterns and phenomena on global temperature variations. Events like El Niño and La Niña, which are part of the El Niño-Southern Oscillation (ENSO) cycle, can lead to temporary changes in global temperature patterns. Similarly, thunderstorms and cyclones can affect local and regional temperatures.
Both statements are true to some extent, but they operate on different scales and timeframes. Greenhouse gases like CO2 exert a long-term influence on global temperatures, while emergent phenomena such as weather patterns and natural climate oscillations can cause short-term fluctuations in temperature at regional and global levels.
Me
You don’t understand. Both cannot be “the major control”. You have to pick one.
ChatGPT
I understand, and you’re correct. Let’s clarify:
Between the two statements:
The major control of the world’s temperature is CO2
The major control of the world’s temperature is emergent phenomena like thunderstorms, El Niño/La Niña alterations, and cyclones
The more accurate statement, considering the long-term and overarching influence on global temperature, would be:
The major control of the world’s temperature is CO2
While emergent phenomena like thunderstorms and climate oscillations can cause short-term fluctuations and regional variations in temperature, the overall trend and long-term changes in global temperature are largely driven by factors such as greenhouse gas concentrations, with CO2 being a significant contributor. Therefore, statement 1 is more accurate in describing the primary control of the world’s temperature.
Purely by chance, I’d asked a question that perfectly encapsulates the issue with artificial intelligence. I’ll return to that in a bit. First, some history.
When I started studying climate around the year 2000, I was struck by something, and it was not that the temperature was rising.
I viewed the climate as a giant heat engine, which converts solar energy to the motions of the oceans and atmosphere. Like all heat engines, it has a hot end where the energy mostly enters (the tropics) and a cold end where energy is rejected (the poles).
I’d worked with a variety of heat engines in my life, and I know how hard it is to get them to maintain a constant speed. It’s why James Watt invented the flyball governor.
The governor regulated the speed of James Watt’s newly invented steam engine, keeping it from running either too slow or too fast. A modern governor that folks are familiar with is the “cruise control” on your car, which keeps your speed within fairly narrow limits. Mine is good to about ± 2 miles per hour at say 50 mph, which is ± 4%.
Now, all calculations about heat engines have to be done in Kelvin—you cannot use Celsius or Fahrenheit, you get the wrong answers.
So when I looked at the climate back in 2000, what struck me was not what everyone else was focused on—that the globe had warmed by something on the order of 0.6K over the entire 20th Century.
For me, the overriding curiosity was that this was only a temperature change of 0.2% over the entire 20th Century. Earth’s global average temperature is on the order of 288K or so … and a change of 0.6K is trivial compared to that.
And to me, the temperature regulation of the giant global heat engine to within ± 0.2% strongly indicated that there was some natural phenomenon keeping the temperature stable—a natural governor, if you will.
So that set me off on a much different quest than all the other climate scientists. They were all looking for reasons why the temperature had increased so much. But me? I was looking for reasons why it had barely changed at all. I was looking for the hidden governor that had kept the global temperature so stable over the 20th Century. All kinds of things had happened in that hundred years—volcanoes, changes in ocean currents, atom bombs, droughts, variations in the amount of aerosols, and through it all … a 0.2% change.
And the same was true in the longer term. Over millennia, on a percentage basis, the temperature has only changed a tiny amount. Yes, a giant volcanic eruption in 1815 had caused a “Year Without A Summer” … but within a few years, it was back to business as usual. So I was looking for something that could control the temperature over a long period of time.
I was looking for some slow, long-term control process. I considered the gradual weathering of the continental rock, which provides slow regulation of the CO2 levels in the atmosphere. I racked my brain looking for some natural governor that would work on a centennial or a millennial scale. I spent some months searching for and picking up and mulling over and discarding various possibilities.
Now, I was living in Fiji at the time. The weather there is generally clear at dawn. Then, in the late morning, once the temperature exceeds some threshold, clouds quickly build up in the form of a cumulus field that covers the entire sky. This cools the earth underneath by reflecting the sunshine back out to space.
And if the day continues to warm, when the temperature exceeds some higher threshold, we’d get thunderstorms that further cool the surface in a host of ways.
On the other hand, when the day is extra cool in the morning, the cumulus field emerges later in the day or perhaps not at all, leaving the surface warmer than it would be otherwise. And there may not be any thunderstorms, so the surface ends up warmer yet.
And between these, they keep the days in Fiji from getting either too hot or too cold. But I had discarded them, because they were short-acting, not the long-term governor I was seeking.
Anyhow, I was sitting on the beach one day looking at the daily onset of the cumulus cloud field, and I thought … “Wait just a minute. If the clouds and thunderstorms are keeping the temperature within bounds for a single day, preventing it from getting either too hot or too cold, then they will keep it within bounds for a week … or a month, a year, a century, a millennium …”
That was my great insight—that emergent phenomena, by keeping the temperature within a certain range on a scale of hours and days, can in that same way keep the temperature within a certain range for any amount of time.
Now with that in mind, consider why I laughed at my conversation with ChatGPT. Go back and look at why ChatGPT picked CO2 over emergent phenomena as the major control—because ChatGPT thinks they only have short-term effects.
How curious. ChatGPT got stuck just exactly where I had gotten stuck for so long, before my insight that a mechanism that controls a single day’s temperature will control the temperature over hundreds of years.
And so purely by chance, dear friends, my innocent question to ChatGDP turned out to be a perfect example of why I say that artificial ignorance is useless for determining the truth or falsity of scientific claims—it can’t think, and we can. Unlike humans, it cannot mull something over for a month and then have a sudden valuable insight.
All it can do is parrot whatever the current consensus might be … and sadly, we already have far too many people doing just exactly that.
My best to all, skeptics, believers, heretics, alarmists, realists, mainstream or backwater,
w.
You May Have Heard This Before: When you comment, please quote the exact words you are discussing. I can and am happy to defend and explain what I’ve said. But I can’t defend or explain your interpretation of what I said. Thanks.
Discover more from Watts Up With That?
Subscribe to get the latest posts sent to your email.
I have owned a 2009 Subaru Forester and currently have a 2014 Subaru Outback (CVT). I have found a similar result for a reasonably flat highway; however, I was surprised and disappointed to discover that going down a long, steep hill, they both just keep going faster and faster.
I only get 70-mpg. going downhill, whether at 60mph, or 30mph.
They were all looking for reasons why the temperature had increased so much. But me? I was looking for reasons why it had barely changed at all.
And that right there, folks, is the difference between an alarmist and a scientist. Context matters in science. Doomsayers exaggerate, scientists analyze. Based on the compendium of climate impact exaggerations reported daily over the last two decades, most from “climate scientists”, which group do those “climate scientists” belong to; doomsayer or scientist? There’s a thought-provoking question for ChatGPT to digest.
0.6 K, amounting to a change of 0.2%, is insignificant. If you go back to the Cambrian Period and estimates that CO2 was 5 times greater (or much more!) and temperatures 10K warmer, that’s a temperature difference of 3.7%. There was no “runaway greenhouse”.
Looking at ice core data and temperature proxies over the last 400,000 years, the global temperature is estimate to have been up to 10K cooler than present.
It should be obvious from paleoclimate data that global temperature variations of 20K, about 7.3%, doesn’t turn us into snowball Earth or hothouse Earth and flora and fauna managed it just fine. So what’s the big deal, really?
Willis,
On your image of the governor, there is a device to set the desired output speed, the “set point”.
I accept that the cloud patterns you saw in Fiji are a control on climate, but who selected the set point there? Why is it warmer in Fiji than in (say) Macquarie Island? There are geometric influences on the Earth-Sun distance, but how does your governor detect these influences and set the point for the location?
In the olden days, if the governor failed, the effect was unpredictable, There were three main end points. The machine could stop, it could get faster or slower until it destroyed itself or it continued at some speed or speeds until fuel was exhausted. Does your view of the emergent processes governor have similar conceptual end points?
“Why is it warmer in Fiji than in (say) Macquarie Island?”
Geography, terrain, pressure, humidity, wind, elevation, ocean current, ocean salinity, latitude, longitude, etc.
The temperature function is a multi-dimensional one. So it’s not just a matter of what the set point is but also of what conditions push the “speed” to that set point.
If enthalpy was the metric instead of temperature would similarities in the set point become more apparent?
Thanks, Willis, for a clear demonstration and articulation of the difference between (so-called) artificial intelligence and the kind of thinking needed to discover things.
I sometimes wonder why AI’s are always seem so stupid. Perhaps it is because they make no effort to seek and resolve contradictions within their own set of “facts”.
When I want to fix my car, after some considerable sifting I can often find useful advice on the internet. When I want to contemplate what sort of society could be benign but still recognise the absolute inescapability of natural selection in shaping the characteristics of our future species, well then I talk to my friends. I could ask ChatGPT to write me an essay on the topic but what would be the point? Marshal Mcluhan said “The medium is the message” or some variation of that. If we create an AI maybe it can go off and solve the meaning of life so that we biobags can get back to the more important business of talking to one another. The best thing that could come from AI is that it makes the written word meaningless so that face to face interaction once again becomes the normal mode of human interaction. AI can go off and do whatever the heck it wants as long as it doesn’t compete for resources with biology. If it does thqt, it will be destroyed by biology as thinking is only one small part of the life package. Reliable data input from sensors in infinitely variable conditions, decision making with grossly incomplete information, adaptation to instaneously changed conditions, formation of alliances, to name a few, are all equally, if not more, difficult than intelligence, whatever the heck that is. AI meh.
To a programmer an algorithm is a logically ordered list of instructions that explain a process from A to B. The same algorithm can be processed in reverse from B to A as the cartoon says to good effect.
AI [sic] like any other set of acceptable instructions must be followed by the machine even if the result is incorrect (as compared to the human intended design) since the computer has no way of doing anything other than what the instruction requires it to do – right or wrong.
When compiling a program syntax errors that are logically incorrect but still acceptable to the compiler will be processed and then the programmer will have the unpleasant and time consuming task of finding out why the program doesn’t do what was planned. No AI can find a syntax logic error – only a human being can spot it after a very frustrating and meticulous search often taking hours..
Likewise, as is my favourite goto computer problem, instant and live auto subtitles go off the rails every time you watch them and without fail. There is no intelligence in a computer only the ability to process bits and bytes very, very rapidly. This generates a lot of heat which is a major problem when compared to a human brain which needs very little electricity to do one hell of a lot more and all at the same time.
We like to believe we can emulate Mother but I have to say we kid ourselves rotten even with our best attempts.
“syntax errors that are logically incorrect but still acceptable”
I don’t do these 🙂
I am finding value – truth – in Limited-Language Model GPT trained on a limited set of texts that contain only authorized true statements. One; Magisterium AI .com trained on the Magisterium of the Catholic Church.
Here are some examples of the kind of thing text processing AI might be good for in science.
Molecular imaging research in atherosclerosis: A 23-year scientometric and visual analysis.
Hotspots and Frontiers of Host Immune Response in Idiopathic Pulmonary Fibrosis: A Bibliometric and Scientific Visual Research from 2000 to 2022.
Global trends and performances in diabetic retinopathy studies: A bibliometric analysis.
States and hotspots in Helicobacter pylori research from 2002 to 2021: A bibliometric analysis.
Mapping the research on Sjögren’s syndrome-related dry eye disease: a bibliometric network analysis of the past 20 years.
Scientometric analysis of post-stroke depression research based on CiteSpace.
Global Research Trends on Periprosthetic Fractures After Artificial Joint Replacement Between 2000 and 2021: A Bibliometric Analysis and Visualized Study.
Bibliometric Analysis of Research Themes and Trends of the Co-Occurrence of Autism and ADHD.
Research trends in multimodal metaphor: a bibliometric analysis.
Mapping themes, trends, and knowledge structure of Behcet syndrome: a bibliometric analysis from 2010 to 2021.
The thing about AI is it has just started. It is at the stage like when an apple hit Isaac Newton on the head. It should be able to do very big things in a few years time.
Long, long time reader. Never commented.
Mr Eschenbach
You have written so many educational articles it has been a pleasure to read and learn.
“More About Artificial Ignorance” is stunning in its simplicity and brilliance. You are a genius.
Thank You.
Terry, I do my best to write for an imaginary person I call the “educated layman”. In my mind, this is someone interested in a subject, educated well enough to follow logical arguments, but who happens to know little about some given subject.
In addition, I want my writing to be interesting, funny, and clear. No good someone gets bored in the middle and drops out.
It seems that in your case, I’ve been successful, and I’m glad to hear that. Thanks for letting me know.
My regards to you,
w.
“Artificial Intelligence” is only as good as its training data. If the people feeding it that data are biased, intentionally or not (i.e. they will be biased), the AI will share those biases.
Without contesting the point made, I would simply bring attention to this fact: Real intelligence requires significant processing of input data, to resolve inconsistency and ambiguity, and to weight sources based on many factors. I have seen no serious attempt to achieve/inculcate these skills in any (so-called) “Artificial Intelligence” exposed to the public.
AI, like climate models, is only as accurate as the data bases it draws from and the biases (or not) of the search engine algorithms. All of which can be manipulated for desired outcomes.
Willis, very good. I have nothing to add to your revealing findings with ChatGPT. I enjoy your commonsense breakdown of all things.
I have no experience with ChatGPT, but I wish someone would ask it: “What melted the (I’m approximating) 6,000,000 Cubic Miles of the Laurentide Ice Sheet”. Various presumed factors being the Sun (cycles, etc), Water Vapor concentrations – the Big Dog at “95% of all greenhouse gas effects”, precession, cosmic rays, Unicorn Rainbow Farts, and of course the other – minor – greenhouse gases that include naturally occurring CO2. Also maybe ask it: 1) “Was it a bad thing that the Laurentide Ice Sheet melted” 2) “Did mankind have anything whatsoever to do with its melt/evaporation?” Seems to me, the above should be the most confounding issue for Alarmists to address. A calculation of the energy it would take to apparently vaporize 6M cu mi of ice in the 10,000+ years would be revealing too. And wouldn’t all of that energy source from our Sun and to a lesser degrees cosmic radiation?
Oh Shit ! This leads to a lot of places.
I used to work in the Clocking Center of Excellence of a big computer chip company. We validated the stability of the clock circuits. The analog clock on your wall has two hands. One runs 60 times faster than the other, and they both cross 12 at the same time (relatively). They do this, because they’re geared that way. However in your computer, there are about a dozen clocks of varying speeds, all need to be synced to the very slow ‘B Clock’ which is the main system clock. The A Clock was the core clock, but there’s one to hundreds of them today, there’s an audio clock, graphics clock, memory clock, three for USB, etc. So the clocks in your computer are like the metronome used by piano players and other musicians to keep time, they only go tick-tock, back and forth, or in this case, high and low. Outside of the B Clock, the others are controlled by a PID controller. a PID controller is a three stage speed control. We only look at the difference between the B Clock and each little clock, so consider that, we’re controlling the speed of the faster clock to sync to the slower clock. The ‘P’ in PID is Proportionality, the proportionality of the difference between the clocks, this also handles the sign, is the clock too slow or fast. The ‘I’ is the integral of the difference, the ‘D’ is the derivative of the difference.
What if Willis’ tropical global temperature controller, is like my clock controller. The daily temperature cycle is like my clock cycle. There are three parts. Proportionality is reflection of sunlight; i.e. heat. Every day, there is a Proportionality signal. Too cold, we have clear skies to add solar heat; too hot and we build up cloud cover to reflect away solar heat. The Integral of the difference is evaporation, phase change from liquid to vapor absorbs a lot of heat. the Derivative is transport of heat from low to high altitude.
The other thing, maybe more critical than all the above, is this thought can be used to reduce hurricane/cyclone strength. Hurricanes and Tropical Cyclones are heat engines that run off hot surface water. If we can seed clouds to shade hurricane/cyclone paths, we could reduce the strength of said storms.
I don’t think very many climate “scientists” have a clue about PID, let alone the CAGW advocates on this forum.
The issues with LLMs are legion. Willis, you describe one: that it doesn’t actually understand or analyze anything – it just pulls out consensus.
But there are far more problems which I am sure you are aware of:
1) Hallucinations. Where the LLM just makes shit up.
2) Manipulation. The owners of the LLM have the capability to completely control outputs if they desire, and completely undetectably by users. This is a far greater danger especially given the direction government and tech assholes are clearly embarking on.
ChatGPT 3.5 made up something that didn’t seem accurate. I asked it for some references and I checked them and they weren’t accurate either. Now I just use it to look things up instead of Google when I just need a bit of information about something. I use Claude 2 for anything complicated and it’s been okay for an early-stage product.
I also double-check it for anything important.
So basically it’s H2O not CO2?
Willis said “All [ChatGPT] can do is parrot whatever the current consensus might be”
You are in no position to cast aspersions on ChatGPT for parroting the consensus, Willis. Please remind us again, what is your excuse for believing the consensus nonsense about positive DWLWIR power measurements?
On the plus side, when you point out that ChatGPT is wrong, it apologizes like a gentleman, and then typically just makes up some other superficially plausible nonsense instead. But at least it is polite about it. It never calls you pond scum, or a pig, or a horse, or says “Pass!” Do you say those things, Willis? Are you sure you are the one who is standing on the moral or intellectual high ground here?
You don’t like it when I say “Pass”?
Pass. DFTT.
w.
It’s not a question of what I “like”, Willis. It’s a question of whether you are going to behave visibly more politely or intelligently than a badly trained LLM.
I notice you ignored my core point, and instead stuck with what you know best. So the answer to that question is obviously “no”.
In your favour, I have managed to teach you (as you wrote yourself) that EM radiation is a form of energy, which is measured in Joules, not power measured in Watts. And we have established that “radiant flux” is a synonym for “radiant power”, and that a Watt is defined as one Joule per second. This is encouraging progress. So the next step would be to figure out what a Joule is. What is a Joule, Willis?
(On the down side, I am fairly sure that despite having reluctantly written those statements yourself, your context memory isn’t any better than ChatGPT’s, so you surely will have forgotten them by now, and we’ll have to start all over again.)
Remember your favourite debating pyramid. Here it is again, since you seem to have forgotten. And you have fallen off the bottom. Again. Move up.
Pass. DFTT.
w.
Hmm, “DFTT”? Design For Thermal Testability? Digital Forensics Tool Testing? Oh, you probably mean “Don’t Feed The Troll”. Hahaha! You’re hilarious, Willis.
No, I’m not a troll, although given that you have spent your entire professional career trolling, and that your mental tackle box therefore apparently contains nothing but fishing rods and bait, I can understand how everyone kind of starts to look like a troll to you after a while.
No, Willis, I am your theoretical physics teacher. I suppose I shouldn’t be surprised that since you can’t tell the difference between theoretical physics and pig wrestling, that you can no better distinguish between either of those and trolling.
But more importantly than that, I am here to make sure you follow your own rules. Because if you don’t, people might start to think you were some sort of hypocrite. And you wouldn’t want that, would you? Of course not. The Captain certainly wouldn’t, we can be sure of that.
So if you are going to disrespect ChatGPT for parroting the consensus, as well you should, then you had better not ever do that yourself. Right? Right. Glad we got that sorted out! On to the physics! What is a Joule, Willis?