Saturday, April 29, 2023

Nothing is Real

Nothing is real. Perhaps it’s the only real thing. Except it’s no-thing.

 

Mindbending Zen koan? Or just physics? To really understand nothing, Amanda Gefter interviews the most famous theoretical physicists alive. The tale of her quest is told in Trespassing on Einstein’s Lawn, probably the best physics-related book for a popular audience I’ve ever read. That’s saying something. Or maybe that’s really nothing. At the end of her narrative, Gefter concludes: “We had found the universe’s secret: physics isn’t the machinery behind the workings of the world; physics is the machinery behind the illusion that there is a world.”

 


Gefter goes on a decade-long quest with many twists and turns. She begins early on by posing as a journalist to gatecrash a science conference to meet John Archibald Wheeler, and ask him about the secrets of the universe. Wheeler is cryptic, speaking in riddles or Zen koans. But that makes Gefter and her co-conspirator even more hungry to know what physics says about the nature of reality. And this drives Gefter further into her “fake it to make it” charade as a journalist writing about physics. Her co-conspirator? Her non-physicist father who loves reading and thinking about physics for fun. Her book is a wonderful autobiography of their joint quest to unlock the secrets of the universe.

 

Why is there something rather than nothing? That was the question Jim Holt asked in his book (which I read just before reading Gefter’s). Where Holt focuses on philosophy, which didn’t seem to get anywhere in my opinion, Gefter focuses on physics and shares some profound things to her non-physicist readers. It’s all about nothing! That’s where everything comes from. She begins with the premise that if something is invariant, then it is real – it is something, not nothing. She makes a list of things you and I would think are invariant. And as she talks to physicist after physicist, the list gets whittled down. Until you’re left with nothing.

 

The iron-cladness of conservation laws start to melt in the withering attack of relativity. Not just Einstein’s special relativity. Not just Einstein’s general relativity. Rather all seems relative. All is observer-dependent. There’s no god’s-eye view, and trying to take one leads to all the spookiness of physics. Gefter discusses light cones, boundaries, black-hole horizon complementarity, holographic spacetime, Rovelli’s relational quantum mechanics, Witten’s M-Theory, Maldacena’s anti-deSitter space, quantum loop gravity, and the strangeness of what we think of as constants: cosmological, Planck, or c. It makes my head spin. It throws me for a loop. Except that I might be creating all of my reality in a loop by being its observer and by thinking about it.

 

Nothing is invariant. Nothing is symmetric. But introduce a boundary or an asymmetry into the system, and voila, information appears. And that’s something. It from Bit. Wheeler’s hypothesis. In the Bible book of Genesis, God creates by introducing boundaries. Separation ensues. Light and Dark. Heavens and Earth. Sea and Land. Life and non-life? The origin of life (my area of research) must have come from asymmetry. But what is that fundamental asymmetry? The right hand no longer knows what the left hand is doing? It has its own frame of reference. Or maybe we’re imposing Boolean logic because it’s easier for us to think in those terms instead of wrestling with the strangeness of quantum mechanics.

 

I admire Gefter’s dogged persistence in pursuing her question about nothing. And the tale she weaves is really something. Trespassing on Einstein’s Lawn is superb and I highly recommend it.

Friday, April 21, 2023

ChatGPT answers my Prompts

When we went remote for the pandemic, to stimulate more discussion, I came up with a Discussion Board assignment where students answered weekly prompts. The prompts were loosely connected to what students were learning in class, and I was aiming for them to be somewhat broad and open-ended. I also hoped they would elicit thoughtful answers. I chose not to use prompts this year because I’m working on new things like my study guides. It’s a coincidence that I stopped using them before ChatGPT was released because I suspect it would be well-suited to answering my prompts, given that it is a stochastic parrot.

 

Let’s put it to the test. I’ve picked a number of prompts from this list where you can see what I thought about the student responses. In my classes students essentially got free participation points by meeting the minimum requirements and posting something. I preface my question to ChatGPT with: “Please write a one-paragraph (120-200) word answer to the following prompt from the perspective of a chemistry college student:” before pasting in the prompt. Here are the results.

 

Prompt: The process of breaking something down into its smaller parts is called Analysis. We've discussed atoms as being the fundamental parts that make up molecules and matter. Can matter be fully described by "the sum of its parts"? That is by breaking something down into its fundamental parts, can you understand what it is (or how it works), or might you lose some understanding in the process?

 

ChatGPT: I would say that matter cannot be fully described by "the sum of its parts." While atoms are indeed the fundamental parts that make up molecules and matter, breaking something down into its constituent parts may not necessarily provide a complete understanding of the material. For instance, certain properties of matter, such as its color, texture, and reactivity, may not be immediately apparent from a purely atomistic view. Additionally, the behavior of complex systems may not be predictable solely from knowledge of their individual components, as emergent properties may arise from interactions between these parts. Therefore, while analysis can be a useful tool for understanding the composition of matter, it is not sufficient for a complete understanding of the material's properties and behavior.

 

A bit bland, but I like that ChatGPT inserted a discussion about complex systems. Not as impressed about reactivity not being from an atomistic view because that’s the whole point of chemistry. But hey, if it’s G-Chem 1, maybe the students haven’t internalized this yet.

 

Prompt: On the first day of class, we talked about the simplicity and versatility of Aristotle's Four Elements (Earth, Water, Air, Fire), but modern chemistry has chosen a different (possibly messier and less versatile) way to classify elements. Today, scientists have "identified" close to 120 elements. And although isotopes exist, we have chosen not to label them as different elements. What do you think about the way elements are classified? Could different choices have been made? Feel free to speculate on creative possibilities.

 

ChatGPT: I believe that the modern classification of elements is more accurate and reflective of our understanding of the nature of matter than Aristotle's Four Elements. While Aristotle's approach was intuitive and elegant, it did not take into account the complexity and diversity of chemical substances. Modern chemistry, on the other hand, is based on the periodic table, which arranges elements based on their atomic structure and chemical properties. This classification allows for a more nuanced understanding of the behavior of different elements and their interactions with each other. While it is possible that different choices could have been made in the classification of elements, it is unlikely that any alternative system would be as comprehensive or useful as the periodic table. However, there is always room for innovation and creativity in science, and new ways of classifying elements could be developed as our understanding of the nature of matter continues to evolve.

 

More bland and robotic sounding. No comment.

 

Prompt: If life was found on other worlds, would organisms have evolved eyes? (Why or why not?) If an organism did evolve an apparatus for vision, what might their world look like to them and how might that view be different from alien human Earthling visitors to their world (if interstellar tourism was safe and possible)?

 

ChatGPT: Whether or not organisms on other worlds would have evolved eyes depends on the unique environmental and evolutionary pressures of their respective worlds. If the environment of a planet is consistently dark or lacks any sources of light, then the development of eyes may not have provided an evolutionary advantage to organisms. Alternatively, if a planet has a consistently bright environment, then organisms may have evolved adaptations such as protective membranes to shield their eyes from the intense radiation. If an organism on another world did evolve an apparatus for vision, their world would look drastically different to them compared to human Earthlings. Depending on the wavelengths of light their eyes are able to detect, colors and shapes may appear completely different than what we see on Earth. Additionally, the visual acuity and perception of depth may be different as well. An interstellar tourism experience would need to take into account the potential differences in sensory perception and environmental adaptations in order to provide a safe and informative experience for both the aliens and the Earthling visitors

 

A bit more interesting and resembles what some of my students would write. Not bad.

 

Prompt: Chemistry is all about the behavior of electrons in atoms. While we know the mass and charge of an electron, it turns out there are many other things we don't know, for example its size, or where it is located in an atom (Bohr was wrong about the orbits). It's not just that we don't know, but we fundamentally CANNOT know (Heisenberg's Uncertainty Principle). We can only describe electrons in probabilistic terms. What do you think about this "fuzziness" in our knowledge of electron behavior? Is it strange that we know the mass and charge but can't know other things?

 

ChatGPT: I find the "fuzziness" in our knowledge of electron behavior to be fascinating and essential to the nature of chemistry as a field. While it may seem strange that we can know the mass and charge of an electron, but not its precise location or other characteristics, the probabilistic nature of electron behavior is what allows us to understand the chemical properties and interactions of matter. The Heisenberg Uncertainty Principle tells us that there is a fundamental limit to how much we can know about the properties of particles at the quantum level, and this applies not just to electrons but to all matter. Instead of being limited by our inability to know certain properties of electrons, we can embrace the probabilistic nature of quantum mechanics and use it to predict and understand the behavior of atoms and molecules. In fact, the "fuzziness" in our knowledge of electron behavior has led to many breakthroughs in fields such as materials science and nanotechnology, where the unique properties of matter at the quantum level can be exploited for specific applications.

 

Mostly repeating what I said and adding applications vaguely at the end. Meh.

 

Prompt: Last week we discussed quantum numbers, their rules, and their restrictions and how those might have led to the arrangement of the periodic table (that's not what happened historically but it's interesting to think about). If one of the quantum number rules were different, how might that change the arrangement of the periodic table? Would there be shorter or longer rows? Might the shape change? Suggest a speculation!

 

ChatGPT: If one of the quantum number rules were different, it could have a significant impact on the arrangement of the periodic table. For example, if the restriction on the maximum number of electrons in each orbital were lifted, the shape of the periodic table could change from its current block layout to a more radial or spherical arrangement. This is because the current block arrangement is based on the maximum number of electrons that can occupy each subshell, which is determined by the Pauli Exclusion Principle and the Aufbau Principle. However, if these principles were relaxed, there would be no restriction on the number of electrons that could occupy a particular subshell, leading to a more complex and potentially less orderly arrangement of the elements. Additionally, the properties and behaviors of the elements could be affected in unpredictable ways, as the electronic configurations of the atoms would be different from what we observe in our current understanding of quantum mechanics.

 

Vague and general-ish, but that’s what I got from my students. ChatGPT mimics them well in this instance.

 

Prompt: While living organisms utilize metals, the vast majority of compounds are covalent, formed from non-metals and predominantly C, H, O, N (with smaller amounts of S and P). Carbon, in particular, forms the backbone of almost all these compounds. Life predominantly uses only a few elements in the periodic table. Any thoughts or speculations as to why this might be? Is there something special about covalent compounds, or about carbon? Is there something puzzling about this?

 

ChatGPT: The reason that living organisms utilize covalent compounds formed from non-metals, with a heavy emphasis on carbon, is likely due to the unique properties of carbon. Carbon is able to form strong covalent bonds with a wide variety of other elements, allowing for the formation of complex and diverse organic molecules. Additionally, the ability of carbon to form stable double and triple bonds allows for even greater complexity in molecular structures. Furthermore, the strength and stability of covalent bonds allows for the storage and transmission of genetic information, which is essential for life. While other elements are also important for life, the versatility of carbon in forming a wide variety of compounds is likely a key factor in why it is so heavily utilized. Overall, while it is interesting to ponder why life primarily uses a few select elements, it is clear that the properties of covalent compounds, and carbon in particular, play an essential role in the chemistry of life.

 

Not bad. Bringing up bond strength, and also double and triple bonds of carbon and the diversity of molecules is what I would have expected my students to say.

 

Prompt: Write one balanced chemical reaction (that involves bonds being made and broken). For each substance in the chemical equation, state (1) whether it is covalent, ionic, or metallic, and (2) whether it is a solid, liquid, gas, or in aqueous solution, under the conditions of the reaction. Then in a couple of sentences, tell us why you think this reaction is interesting -- presumably why you chose it.

 

ChatGPT: The balanced chemical reaction I have chosen is the combustion of propane, which can be written as: C3H8(g) + 5O2(g) -> 3CO2(g) + 4H2O(g). In this reaction, propane (C3H8) is a covalent gas, while oxygen (O2) is a covalent gas. The products, carbon dioxide (CO2) and water (H2O), are both covalent gases as well. This reaction is interesting because it is an example of a combustion reaction, where a hydrocarbon fuel (propane) reacts with oxygen to produce carbon dioxide and water. Combustion reactions are essential for energy production in many industries, and they also have important environmental implications due to the release of carbon dioxide, a greenhouse gas, into the atmosphere. Additionally, balancing the equation requires careful consideration of the stoichiometry of the reaction, which is an important skill in chemistry.

 

Okay, so my question wasn’t that interesting. ChatGPT does a solid job here in answering it.

 

Why this blog post? I’ve been wondering if ChatGPT could take over writing my blog or at least providing significant scaffolding. I think the answer might be yes. Most of today’s blog was cut-and-paste. I’d like to think that I have a certain quirky writing style and off-the-wall ideas that ChatGPT is unlikely to emulate. Then again, I could feed it a bunch of my data and ask it to write the way I do, and maybe it will. Ah, well, should I keep on writing? I suppose so, since the point of my blog isn’t so much to communicate my thoughts to others as providing an extended storage of my thoughts that is easily searchable. I also use it to try and improve my writing by just doing it more! But I can thank ChatGPT for providing some fodder for today’s post. I cranked it out in 30 minutes, much faster than average. Lazy me. Not sure whether I’ll assign prompts on the LMS Discussion Board again.

Thursday, April 20, 2023

Messy Miller Experiments

Research into the chemistry of the origin of life got a kickstart with the 1953 publication by Stanley Miller reporting on his experiments that produced amino acids, the building blocks of proteins, from simple substances (H2, CH4, NH3, H2O). The energy source used was a spark discharge (emulating lightning) so these are sometimes referred to as spark discharge experiments. They are also known as Urey-Miller or Miller-Urey experiments, attaching the name of Miller’s supervisor, Harold Urey. While Urey had initially tried to discourage Miller from origin-of-life chemistry, thinking it would not yield anything useful, he was very supportive by giving Miller full credit as sole author on the now-famous 1953 paper.

 

Seventy years later, such experiments continue. We’ve learned a lot since then. One of the important lessons is that to get a wealth of organic compounds in your experiments, you have to embrace the messy! A recent publication by Root-Bernstein and colleagues is the latest iteration. Here’s a picture of the abstract. (The article is open-access so you can read it in full; DOI: 10.3390/life13020265)

 


As the title suggests, the experiments produce a rich mess: not just amino acids but also sugars, nucleobases, lipids, and even oligomers – the linking of building blocks to make larger molecules that life utilizes. To the solution they add a particular sea salt from the Mediterranean along with hydroxyapatite (a phosphate salt) and magnesium sulfate (Epsom salt), common salts you might expect to be present in prebiotic chemistry. They ran pre-tests to make sure no ATP or other “biological” molecules were present. Other than that, they tried to stick to Miller’s original initial conditions. One crucial way in which the protocol differed from the original, is that instead of a one-pass reaction (Miller’s was seven days), the researchers regularly reintroduced the “atmospheric gases” (H2, CH4, NH3) every seven days and ran the experiment for multiple weeks. This allows fresh “nutrients” to enter the system so that the chemistry can progress much further – and get messier!

 

At regular intervals, they remove some sample solution to figure out what’s in it. As expected, within the first week, amino acids show up similar to the ones Miller found. Sugars also show up early. This is expected since the Strecker reaction for producing amino acids relies on the building block formaldehyde, which is also the building block for producing sugars in the formose reaction. By the end of the second week, nucleotides are found, i.e., nucleobases were synthesized. Again, not surprising, since precursors such as HCN and formamide are expected to play a role. The next few weeks bring fatty acids and peptides (oligomers of amino acids) into the mix. There are a couple of dicarboxylic acids (malonic and succinic) and even a few sterols. It’s a very impressive mix – you’ve got stuff for membranes, proteins, sugars, nucleic acids, and metabolic diacids.

 

The main question that arises in this embarrassment of riches is whether there might be contamination, especially when “regassing” or when removing samples for testing. The authors have expected this question and discussing how they have been careful in their protocols, and how they have tested for errant contamination. While contamination cannot be completely ruled out, it seems to me that they have done due diligence and there don’t seem to be any telltale biomarkers present that you’d find in contaminated samples. The other thing that jumped out at me was the presence of the (sulfur-containing) amino acid cysteine. Cysteine is sort of a magic amino acid that catalyzes all manner of reactions. You can produce it by adding H2S to a spark discharge experiment, but in this instance, H2S was probably produced in situ by reduction of sulfate. This means that thioesters could come into play, although no mention is made of them in the paper. I didn’t expect thioesters to be directly observed because they would hydrolyze easily under these experimental conditions.

 

The authors rightly emphasize the messiness and the advantages of what they call “dirty” experiments. I’m in agreement. The analysis is painful, and teasing apart what’s going on in such reaction mixtures will continue to be challenging, but I think this is the right approach to make progress in the field. Hurrah for the mess!

Saturday, April 15, 2023

Magic of Babel

The biblical tower of Babel fell into ruin because of confusion among people who spoke different languages. Presumably it was a severe breakdown in communication, although the Bible doesn’t provide details. Building on this idea, the author R. F. Kuang has written a novel, aptly named Babel, with a magic system that exploits the differences of meaning between different languages. What is lost in translation becomes a source of power. It’s a difference that makes the difference!

 


While Babel has been compared to Jonathan Strange & Mr Norrell, other than the setting of the early nineteenth century and the common theme of magic, the two novels are very different. Babel is another iteration of the coming-of-age story whereby a young protagonist is initiated into the study of magic while making new friends – we’ve seen this before many times, it’s a very effective trope. While this is interesting in the first half of the book, the second half focuses on race, colonialism, power, economics, and the lengths people will go through to maintain the hegemony of the status quo. As someone from a Commonwealth country, I’m familiar with the historical setting of Pax Britannica, and I also had the experience as a multilingual international student of being in a foreign place with no other countrymen (at my college). That being said, I feel that Babel was overall depressing, and I think the sociological arc of the story is weaker – it feels like the author is trying too hard to push a particular point of view.

 

The tower of Babel is the centerpiece of Oxford because that’s where the magic happens. The ascendance and economic might of the British Empire and its East India Company maintain hegemony because they fashion magical technologies by etching words in different languages on silver. When someone fluent in those languages speaks the words, it imbues the silver object with magic that enhances a particular technology. For example, it can make cannons fire more truly, carriages travel faster or carry heavy loads, and caravels to more ably weather storms at sea. Metal-etching and language as the foundation of magic isn’t novel – the recent Foundryside series employs a related system – but the system in Babel is more interesting. Kuang’s background in multiple languages enhances her novel, and in my opinion, these are the strongest and most interesting parts.

 

When our main protagonist Robin and his fellow first-year students get their first tour of Babel, a professor says: “Now that you’re part of the tower… the tower knows you.” They’ve just given their blood so that they will be allowed safe entry and exit into the hallowed halls of Babel. Gotta keep the riffraff out. (Foundryside has a similar situation.) Then the professor intones: “You’re in the place where magic is made. It’s got all the trappings of a modern university, but at its heart, Babel isn’t so different from the alchemists’ lairs of old. But unlike the alchemists, we’ve actually figured out the key to the transformation of a thing. It’s not in the material substance. It’s in the name.”

 

I like that Kuang makes the connection to alchemy and the transformation of substances (transmutation), but then emphasizes that it’s not so much its ‘elementary’ material as its name that makes the difference. The power of names and naming is spread across myriad ancient cultures, and still holds significance today. It’s unfortunate that our hypermodern world has too a large extent lost this significance. I also think the idea that information underlies matter, rather than the other way around, is particularly intriguing – and there might be evidence deep in fundamental physics – but that’s a story for another time.

 

In the introductory class to Translation Theory, the professor says: “The first lesson any good translator internalizes is that there exists no one-to-one correlation between words or even concepts from one language to another… Language does not exist as a nomenclature for a set of universal concepts… How do we render the French esprit into English?” This then leads to the “three principles of translation”: A student recites these: “First, that the translation conveys a complete and accurate idea of the original. Second, that the translation mirrors the style and manner of writing of the original. And third, that the translation should read with all the ease of the original composition.”

 

The emphasis of really understanding more than one language deeply and fluidly moving between one language and another is magical, even in our non-magical world. I have first-hand experience trying to learn a new language that I did not grow up speaking. It’s hard work! I haven’t achieved depth or fluency, but I know it’s possible. And the Babel students work insanely hard at translation and getting to the tangled evolutionary roots of words in different languages. The professor also hints at a primordial ‘Adamic’ language that may have been lost while making hints to biblical Babel. A similar concept in Foundryside takes centre stage, while it’s merely a curiosity in Babel.

 

An example of how the magic works by exploiting the differences of word meanings in languages is provided by the professor: “The Greek word idiotes can mean a fool, as our idiot implies. But it also carries the definition of one who is private, unengaged with worldly affairs – his idiocy is derived not from lack of natural faculties, but from ignorance and lack of education. When we translate idiotes to idiot, it has the effect of removing knowledge. This [silver] bar, then, can make you forget, quite abruptly things you thought you’d learned. Very nice when you’re trying to get enemy spies to forget what they’ve seen.”

 

But there are problems. Silver tarnishes over time and must continue to be ‘maintained’ by cleaning the dross and re-inscribing. The professor explains that depending on the desired effect, you might need larger quantities of silver or higher purity. Most of the bars are alloys, and the continual need for repair and maintenance keeps the scholars of Babel rolling in riches. They have a monopoly. The other interesting aspect comes up in a conversation among students. As languages evolve and start to share more similarities, the erasure of differences causes the silver bars to lose power. This anchors an aspect of the story where Babel is interested in Oriental languages and the Far East. The larger differences in translation result in more ‘raw’ magical power to be exploited. This gives the book an international cast and emphasizes a variety of languages. The interplay between language and magic in Babel is deeper and more interesting than in Foundryside, and in my opinion, it’s the most magical part of the novel.

Thursday, April 13, 2023

Reminiscing Catan

Klaus Teuber passed away last week. For those who don’t recognize his name, Teuber designed the boardgame Settlers of Catan. It was one of two games that I discovered in the ‘90s that led to my renewed interest in boardgames. (The other was Richard Garfield’s RoboRally). Settlers of Catan led to a renaissance in boardgames that hit a Goldilocks sweet spot. Not too long and complicated. Not too short and trivial. It has a nice balance of luck and strategy, and there was both a competitive and cooperative aspect to winning gameplay.

 

In those early years, I played hundreds of games of Settlers, as we called it then. (Now the game is just called Catan.) I still have my first (English) edition. Although the box cover is slightly beat-up, the components are still in good condition. But with the arrival of newer games, Settlers began to lose its shine. I will still happily play it with the right crowd, i.e., one that doesn’t devolve into analysis-paralysis or hyper-competitiveness. For me, the sweet spot is a four-player game that takes 45-60 minutes. It’s much better as a relaxing, low-key, game of building and trading. And occasionally frustrating your opponents.

 

The first edition does not mesh with any of the expansions, so I’ve never bought them. Also, the expansions add to playing time. I like the Seafarers expansion. It adds 15-30 minutes, depending on the scenario, but provides an element of discovery (which is fun!) and also mitigates the sheep problem. (If you’ve played Settlers a lot, you know what I mean.) On the other hand, the Cities and Knights expansion easily adds an hour or more, and results in more hindrance strategies. I tired of it after my first game, although I did give it a chance by playing it more than once. There are now many themed versions of Settlers of Catan. I only own one: Settlers of the Stone Age because I like the history-evolution theme to it. The gameplay is, overall, slightly inferior to the original, but I like the theme enough to play it sporadically.

 

I looked up the list of games designed by Teuber to see how many I’ve played. Elasund and Entdecker are games I used to own but gave away to friends as I accumulated too many games and they were hardly played by me. Teuber hasn’t actually designed that many games outside of the Catan spinoffs and expansions. Two famous ones that I haven’t played are Barbarossa and Hoity-Toity. Looking at my own collection, the designer whose games I have collected the most is Reiner Knizia, a prolific game designer who has elegant designs likely influenced by his background in mathematics.

 

Back in the day, these “famous” German game designers worked on games as a side-project. They had full-time jobs. Teuber worked as a dental technician. Knizia worked in finance. Both finally quit those jobs to become full-time game designers in the late ‘90s. What is it about the Germans and their side hobbies? One of the most influential theories of the origin of life came from a patent lawyer, Gunter Wachtershauser, who quietly toiled away from years before unveiling his metabolism-based iron-sulfur world. It shook up the origin-of-life community and opened new lines of research.

 

Nowadays, my game-playing has ebbed to a low level. The pandemic didn’t help – although there was a resurgence in my playing the boardgame Pandemic during that time. I read more now; I suppose that’s my new hobby. But I’m always up for a good game. As long as it doesn’t take too long, and has an appropriate mix of luck and strategy commensurate with the time I invest in playing it. Settlers of Catan paved the way for the myriad choices available today, thanks to Klaus Teuber.

Wednesday, April 5, 2023

Stochastic Parrots

Digging into ChatGPT and Large Language Models (LLMs) is leading me down a rabbit-hole. In my last post, I looked at a paper discussing whether such models ‘understand’ language, and if so, how different might it be from how humans learn. The crux was whether statistical correlations are sufficient mimics that can, for practical considerations, substitute for knowing causal mechanisms. We don’t really understand what these LLMs are doing when they seemingly come up with novel and surprising responses, which they have not been ‘trained’ for (kinda, sorta). It’s a black box even to those who developed such AIs.

 

Today, I’m looking at a different paper with a catchy title: “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” The article is open-access at the Association for Computing Machinery digital library linked here. It’s a little more technical, and assumes some familiarity with LLMs. Since this article is from 2021, it does not include the latest version of ChatGPT or some of the newer-released chatbots. But they have more than enough data to make the key arguments in their paper. The background information provided in the paper was very useful for a novice like me to get a brief history of developing such models and a brief overview of the different LLMs out there. There’s a bunch of them!

 

Whenever you query ChatGPT, it takes up energy and compute time. How much energy? There are various estimates on the internet and the kWh per query number may look small, but a quick calculation will show that it is 2-3 orders of magnitude (i.e., 100-1000 times) more than the human brain. LLMs are energy guzzlers compared to us. This doesn’t account for the amount of energy used to train the LLMs in the first place to get them to the user-friendly stage. The bigger the LLM, the more energy it guzzles.

 

Reading about the training data was an eye-opener. I’m a computational chemist so I’m familiar with using data to train computational models. An important thing to keep in mind is GIGO: Garbage In, Garbage Out. How well your program will perform its specific task depends on the quality of your training data. Where does the training data come from for something like ChatGPT? It’s not just the things you might expect like Wikipedia, digitized books, and vetted data repositories. Turns out there are very large datasets known as Common Crawl. Sources include “scraping outbound links from Reddit” and other user-generated content such as Twitter. There are biases in those data sets in terms of who the users are and what they might discuss on the internet. We shouldn’t be surprised at the misogyny that shows up from chatbots. The article authors provide numerous examples of why these issues crop up. I recommend reading their article in full.

 

Given that the amount of garbage on the internet is increasing exponentially faster than anything else, the worse things get. Over the years, I’ve noticed more wrong things (related to chemical data) show up in the top few hits on a standard Google search. Search has in fact become more tedious when I’m looking for good data on specific things I care about. For those researchers who work in Natural Language Processing and are trying to move into Natural Language Understanding, the approach of just throwing more scraped data is likely going to make things harder especially when resources are being thrown at these LLM approaches. Is Big Data the answer? The authors make the point that “[language] coherence is in the eye of the beholder… human communication relies on the interpretation of implicit meaning conveyed between individuals… [it] is a jointly constructed activity… even when we don’t know the person who generated the language we are interpreting, we… [intuit] what common ground we think they share with us, and use this in interpreting their words.”

 

The authors refer to LLMs as Stochastic Parrots. This is an apt name given what these models are doing by applying statistics and Bayesian probabilities to constructing plausible-sounding text. We should think carefully about taking advice from these parrots, and certainly double-check what they’re telling us. But we’re lazy. Why do the hard work that you can outsource to a machine? Isn’t it just a useful tool? Like a calculator? Well, partly yes, but every year I see students punch their calculators and write nonsense answers. You could call it user error, but it’s an error often associated with not understanding how the calculator parses the input you provide to generate the output. I could make an analogy to the workings of a chatbot. Except that things are even fuzzier. There’s now a stochastic wrench thrown into the works. But all that being said, I’m still finding the stochastic parrot interesting to interact with – it gives me a sense of what an alien intelligence might be like.

Tuesday, April 4, 2023

Alien Understanding

As I’ve been playing with ChatGPT, I’ve also been reading about Large Language Model AIs. It’s clear to me that ChatGPT doesn’t understand chemistry in the same way humans do, nor should I expect it to given what it is programmed to output; but does it understand language? I suppose that depends on what one means by “understand” and what one means by “language”. The distinction that comes to my mind is between syntax and semantics. My opinion is that ChatGPT mimics a human speaker because it is well-trained on syntax. We humans feel that it is an effective mimic because in a conversation, we have evolved to impute semantic information to any conversation partners, human, animal, or machine.

 

A useful primer that’s accessible to the non-expert is a recent short four-page article by Melanie Mitchell and David Krakauer. (“The debate over understanding in AI’s large language models”, DOI: 10.1073/pnas.2215907120.) Here are some of their statements that I felt clarified the key questions or highlighted the key issues.

 

“… the current debate suggests a fascinating divergence in how to think about understanding in intelligent systems, in particular the contrast between mental models that rely on statistical correlations and those that rely on causal mechanisms.”

 

“[In the past], the oft-noted brittleness of these AI systems – their unpredictable errors and lack of robust generalization abilities – are key indicators of their lack of understanding… However, [present LLMs]… can produce astonishingly humanlike text, conversation, and in some cases, what seems like human reasoning abilities, even though the models were not explicitly trained to reason.”

 

“… such networks improve significantly as their number of parameters and size of training corpora are scaled up… [some claim] will lead to human-level intelligence and understanding, given sufficiently large networks and training datasets.”

 

“…[others argue that LLMs] cannot possess understanding because they have no experience of mental models of the world; their training in predicting words in vast collections of text has taught them the form of language but not the meaning.”

 

“… the Eliza effect named after the 1960s chatbot… refers to our human tendency to attribute understanding and agency to machines with even the faintest hint of humanlike language or behavior.”

 

The crux is that we don’t exactly know, down to a nuts-and-bolts level (or should that be bits-and-bytes), what human understanding is. We have mental models – conceptual, abstract, and tied together by particular notions of causality – and we perhaps assume that such models aren’t solely statistical correlations. In this model, we have a self. And this self needs to function in the real world of blood-and-guts and nuts-and-bolts. As we interact with our environment, which may include other entities like ourselves, we use what we have learned to make predictions of what might happen next. We have expectations based on our conceptual models, and we are surprised when reality doesn’t meet expectation in a particular functional test. Then we update our model.

 

One argument against LLMs having human-like understanding is that they “exhibit extraordinary formal linguistic competence… they still lack the conceptual understanding needed for humanlike functional language… and use language in the real world. An interesting parallel can be made between this kind of functional understanding and the success of formal mathematical techniques applied to physical theories. For example, a long-standing criticism of quantum mechanics is that it provides an effective means of calculation without providing conceptual understanding.” As someone who teaches quantum mechanics, this analogy strikes home for me. But maybe it’s our present lack of understanding of our own understanding that’s the roadblock. And if we can learn how LLMs “understand” in their own way, that might shed more light on what human understanding means. It’s unclear that the lack of embodiment in present LLMs constitutes a significant hurdle, but one could conceive outfitting a physically mobile robot with an LLM. Such an integration could prove interesting. And scary.

 

If future improved LLMs cannot create something that resembles a rich-enough mental model that seems akin to humans, can they at least do what is functionally equivalent? Mitchell and Krakauer suggest that when we see “large systems of statistical correlations produce abilities that are functionally equivalent to human understanding… [or] enable new forms of higher-order logic that humans are incapable of accessing… will it still make sense to call such correlations ‘spurious’ or the resulting solutions ‘shortcuts’? And would it make sense to see the systems’ behavior not as ‘competence without comprehension’ but as a new, nonhuman form of understanding?” They also caution us not to jump to conclusions when trying to interpret how LLMs perform on our “tests” of comprehension based on our own assumptions of how human understanding works – something we don’t really understand at a deep level. It’s perhaps an alien understanding. We might anthropomorphize such alien understanding in sci-fi literature and media, but perhaps that’s the only way we as humans can try to understanding something beyond our ken. If anything, all this reminds me that we need to remain humble about what we think we know. We’re limited by our own human embodied understanding.