Large Language Models (LLM) such as OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude have transformed the way we understand artificial intelligence (AI).
The original test for human-level intelligence in machines was known as the Turing Test. To conduct the test, a human interrogator carries on a text-based conversation with another human in one window and a machine in another. If the interrogator can’t tell the difference between the machine and the human, the machine passes the test.
All three of the LLMs listed above have (arguably) passed the Turing Test, but (basically) no AI researchers make the argument that these LLMs have achieved human-level intelligence.
Changing the goal posts for demonstrations of artificial intelligence is not new. As one of the founders of the field once said, “As soon as it works, no one calls it AI anymore.”1 We used to believe any machine that could defeat a human grandmaster in chess would possess the ability to “learn abstract concepts, think cleverly about strategy, compose flexible plans, make a wide range of ingenious logical deductions, and maybe even model one’s opponent’s thinking. Not so. It turns out to be possible to build a perfectly fine chess engine around a special-purpose algorithm…It plays chess; it can do no other.”2 Humanity’s grandmaster fell to a chess engine before the end of the 20th century.
If you ask ChatGPT’s o1-mini model whether it has human-level intelligence, it will sympathize with the question while going on to explain the differences between what it does and what a human can do:
LLMs do not possess the characteristics of human intelligence. They do not experience emotions, learn or adapt in real time, or sense what theory of mind philosophers call qualia, the subjective properties of the experience of a thing (e.g., the internal experiential differences between an itch and a scratch, between seeing an orange leaf and a green leaf, between tasting a cinnamon roll and smelling a cinnamon roll as it bakes in the oven). In other words (and obviously), ChatGPT and its ilk are not alive.
But that does not preclude LLMs from being a superintelligence.
What Is Superintelligence?
Nick Bostrom, in his 2014 book Superintelligence: Paths, Dangers, Strategies, explores three different forms of superintelligence:
- Speed superintelligence: A system that can do all that a human intellect can do, but much faster, where “much faster” is only limited by the speed of light.
- Collective superintelligence: A system composed of a large number of smaller intellects such that the system’s overall performance across many very general domains vastly outstrips that of any current cognitive system. Consider the way a collective of humans known as NASA (and its contractors and sub-contractors) put a human on the moon while no single human working alone has ever accomplished the feat.
- Quality superintelligence: A system that is at least as fast as a human mind and vastly qualitatively smarter. Consider the qualitative differences between human minds, elephant minds, dolphin minds, and chimpanzee minds, and now imagine the “smartest” one as an engineered system.
While Bostrom doesn’t include the concept of consciousness in his forms, in virtually all of our science fiction imaginings, superintelligence is associated with and often defined by consciousness, self-awareness, sentience, etc. Without this key ingredient, it becomes difficult for us to imagine the intelligence as superior to human-level intelligence; instead, we would see it as a mere automaton, maybe the most complex in human history, yet still and always an automaton.
ChatGPT (and other LLMs) seem to fit the automaton definition. To use a simplified analogy, LLMs are highly intricate versions of the autocomplete feature on your iPhone. They predict the next word or phrase based on context. While your iPhone is limited to the context of the current input, common usage patterns, and your personal typing habits, LLMs are limited to the current input and what they’ve learned from their massive training data. The training data allow LLMs to understand context, grammar, and semantics at a much deeper level than your iPhone, enabling them to generate extended and complex texts across infinite topics,3 all without consciousness.
But does a superintelligence require consciousness?
The Superintelligence of Language
As far as we are able to tell, an LLM is not a conscious entity,4 but that doesn’t mean it can’t be a form of superintelligence. But first, we have to start with how they are trained, what they’re trained on, and what they’re made of. In short, we have to understand what an LLM does and is.
What Is Training Data?
The training data for an LLM is text. That’s it. Nothing more and nothing less than text. The text comes primarily, but not exclusively, from the Internet. The data includes books in the public domain, as well as licensed fiction and nonfiction,5 Wikipedia and other encyclopedic content, public webpages, scientific articles and technical documentation, programming code, and dialogue datasets (conversations, Q&As, etc.).6
The data isn’t fed raw into the machine, however. It’s cleaned and filtered to get rid of duplicates, spam, or formatting artifacts, leaving a massive text-only dataset containing hundreds of billions of words structured simply by grammar, convention, and cultural usage.
…a massive text-only dataset containing hundreds of billions of words structured simply by grammar, convention, and cultural usage.
How Is The Data Processed?
1. Conversion into Tokens:
LLMs don’t work with whole words or individual letters; instead, they break text into units called “tokens.” A token might be a word fragment, a full word, or a punctuation mark, depending on frequency and context.
Consider the lyric from the Grateful Dead song, China Cat Sunflower:
“Krazy Kat peaking through a lace bandana like a one-eyed Cheshire, like a diamond-eye Jack.“
I ran that through OpenAI’s tiktokenizer, which reveals that it’s made up of the following tokens:
“K”, “ra”, “zy”, “_Kat”, “_pe”, “aking”, “_through”, “_a”, “_lace”, “_band”, “ana”, ” like”, “_a”, “_one”, “-“, “eyed”, “_Ch”, “eshire”, “,”, “_like”, “_a”, “_diamond”, “-“, “eye”, “_Jack”, “.”
(Please note, where you see the “_” above, the machine sees a blank space; I used the underline to make it clearer that the token includes the blank space.)
Of course, as a machine, LLMs need to connect those tokens with numbers, so each token is assigned a unique ID. In OpenAI’s token language, one might sing Jerry Garcia’s line by belting out:
42, 614, 3705, 22885, 1045, 5016, 1819, 261, 60226, 8177, 1480, 1299, 261, 1001, 109991, 168614, 11, 1299, 261, 36181, 90091, 10874, 13
The process of tokenizing is human-engineered. If you’re interested in the theory behind it, take a look at Toward a Theory of Tokenization, but beware that you’ll have to deal with phrases like “end-to-end cross-entropy loss.”
2. Conversion into Vectors:
This second step is more complex.
If you look above, you’ll see that “Jack” equates to the token “10874” (“13” = the period at the end of the sentence). But 10874 doesn’t mean anything in the same way “Jack” doesn’t mean anything. To understand what the lyricist Robert Hunter meant when he wrote “Jack,” we need to look at the full context.
For the LLM, you can think of the “full-context” as a row in spreadsheet, where the first column is the token (“10874”), followed by 1,024 other columns, with each column representing some dimension of meaning.7 In one column, the cell for row 10874 might have “0.103,” while another column might have “-0.554,” etc. Each column gets its own particular value for token 10874, and taken together, the entire row—the vector—gains meaning by its relation to other tokens (more on that below).
Before the LLM is trained, the structure for this matrix already exists. Each row is initially filled with random numbers, a placeholder with absolutely meaningless content, but by the end of the training session, the matrix represent a geometric map of language.
3. Building the Matrix:
The model builds this map by using the existing draft of the matrix to correctly predict the next token in a given text. If the sample tokens are: “K” “ra” “zy” “_Kat” “_pe” “aking” “through” “a” “lace”, then how successfully does the matrix predict the next token will be “_band”, followed by “ana”?
Once it calculates how wrong it was, it adjusts the numbers in the matrix (and in other internal weights we’ll discuss below) using…well, math8…and runs the trial again. It does this billions of times, gradually reshaping the initially random vectors in the matrix so tokens that appear in similar contexts end up with similar vectors. “Krazy” and “crazy” might be near each other, while “lace”, “bandana,” and “veil” might form a cluster.
The result is a semantic space that tells the model, “This is how language works.”
4. Transforming the Vectors
If LLMs were just the token matrix, they’d be little more than fancy, crazily complex spreadsheets. But there’s more to these models than a big lookup table.
There’s also the transformer. The transformer was the key insight that led to the current generation of LLMs (as compared to, say, ELIZA). Transformers help the model decide what part of the input to pay attention to. They do this by building their own matrices of vectors. Early transformers used three additional vectors beyond the token. These were called Query, Key, and Value. The transformer would compare the Query vector of every token in the input (“Krazy Kat peaking through a lace…”) to the Key vector of every other token to calculate how much attention it should pay to each one. Then it used those attention scores to mix together the Value vector, creating a new, context-aware representation of each token.
Contemporary transformers have added more matrices to this process. Some might focus solely on syntax, others on understanding what pronouns refer to, and others might focus on more long-range dependencies, such as the narrative beats in a story or the rhyme scheme of a sonnet.
Essentially, transformers allow the model to pay attention to the right things in any given input. What’s incredible is that they do this in parallel, working on all of the tokens at once before combining them back into one final representation per token.
5. Refining the Vectors
Once the attention mechanism has helped the model figure out what matters, each token (or more accurately, its updated vector) gets a bit of private processing time. This is called the feedforward layer. It’s as if the model stares out a window and considers the philosophical underpinnings of the vector.
Let’s say the model is considering the vector for “Jack.” It’s already considered the whole of the input and figured out which other elements are relevant to Jack (“diamond-eye,” “one-eyed,” “bandana,” etc.), but now it needs to do something with all of that context.
So it takes the vector for “Jack” and using…well, math…it stretches its understanding of “Jack” from 1,024 dimensions to 4,096 dimensions, reshaping it using…well, math…before squishing it back down again, this time in a way that refines its sense of Jack’s role, turning the original general sketch into a refined portrait.
Think of it like this. In the feedforward layer, the model asks itself, “Is Jack a person, a playing card, a game-piece, or a metaphor? Is he sneaky, serious, or noble? Is this a sentence where Jack needs to be taken seriously or is Jack a punchline?”
It doesn’t do any of this using words, of course, yet by updating the numbers in the vector for the token, it creates a better shape for “‘Jack’ in this sentence” vs “‘Jack’ in general.”
The attention mechanism is all about looking outward, figuring out which other tokens are most relevant, but the feedforward layer is about looking inward. It happens for each token without further influence from the others. The model isn’t considering “Jack” in relation to “diamond-eye.” It’s thinking only about “Jack.”9
6. Generating the Next Token
LLMs are also called generative AIs. Unlike traditional AI, which analyzes and predicts, a generative AI actually creates.
At this point in the process, every token in the input has been transformed and refined, turning it into a vector shaped by its relationship to the other tokens and by its own internal, context-specific processing. But this has all just been preparation for the groundbreaking step that promises (or threatens) to transform our relationship to technology as much as, or more than, the creation of the web browser, the iPhone, or social media.
The model is ready to generate something new.
With “Jack” as the last token of the input and the transformed and refined vector that represents Jack having a 1,024-dimensional shape sculpted by the context around it, the model feeds the vector through one more step: the output layer. The model compares the vector to every token it has ever seen in its training (we’re talking tens of thousands of tokens) and creates a probability score of how likely each token is to come next.
It isn’t guessing at what a sentence means or creating narative intent. It simply asks itself, “Based on everything I’ve ever read, and everything I’ve learned about human languages, what token usually follows one like this?”
After calculating for every token it knows, it picks one, and that token becomes the output. That output is what the model actually generates, but from that single choice, the process begins again. The new token is added to the sequence, and the model loops back, tokening, vectorizing, attending, transforming, refining, and generating again, one token at a time, building original poems, emails, arguments, love letters, misinformation, documentation, press releases, computer code, or constructive criticsms of paragraphs that attempt to explain how an output layer works.10
A Superintelligence without Consciousness
But what have we wrought?
An LLM is not a conscious mind, not a “being.” There is no awareness behind its choices, as far as we can tell11, no inner monologue, and, perhaps most importantly, no inner narrative wrapped up in its token for “I”.
And yet, taken as a whole, LLMs exhibit something remarkable: language that reflects, persuades, empathizes, fools, lies, jokes, synthesizes, and surprises. It doesn’t “know” what it is saying; it generates token after token based on its self-created map of the hidden structure of human expression.
What emerges is not an artificial mind. It is something even more alien: a system that generates meaning not from thought, but from structure. It doesn’t possess intelligence, it enacts it, one probability at a time.
LLMs do not simulate consciousness. They simulate language. And it turns out language is smarter than we thought.
The superintelligence of language is not born from neurons or the self-awareness of silicon. It comes from pattern, compression, association, and prediction, amplified to a scale that no human can match.12
Language As An Evolved Entity
In his 1976 classic, The Selfish Gene, biologist Richard Dawkins reframed evolution by shifting the focus away from individuals and species, and toward the fundamental unit of evolution: the gene. At the end of the work, he proposed the existence of a second evolutionary unit on Earth: the meme.
Where a gene drives biological evolution, a meme drives cultural evolution. The medium of the meme is language. When you repeat a slogan, a joke, a cliche, a memorable tune, or even a religious tenet, you are transmitting a meme.
Because each brain is unique, memes don’t copy perfectly from one brain to another. Sometimes, they mutate. Researchers have shown that you can track these mutations across space and time until the words of a poor, fundamentalist Jewish carpenter are mutated by the brain of a Jewish Roman murderer to apply not only to Jews but also to gentiles, before being further shaped by the educated mind of a Greek physician until what began as a radical interpretation of the Torah becomes the New Testament of a triune God.
This isn’t just history. It’s an evolutionary process: replication, mutation, and survival of ideas across generations.
LLMs are trained on memes. They absorb the patterns that have evolved across time and culture, so when they “speak,” they do so with the ancient, ever-evolving voice of language.
Again, it does so without understanding the ideas themselves, but it doesn’t need to. Just as the cellular and subcellular entities do their work in our bodies without understanding what or why, the LLM produces and remixes our cultural DNA without intention.
That’s a strange thought, but consider the work of philosopher and cognitive scientist, Daniel Dennett. In his book, Consciousness Explained, Dennett proposes that consciousness isn’t a centralized processor in your brain but the emergent result of countless micro-processes running in parallel and loosely stitched together by the illusion of a narrative.
For Dennett, “you” are not a single entity watching the show from behind your eyes. You are the constantly re-interpreted and re-composed drafts of what happened, what is happening, and what will probably happen next, motivated by the continuation of “your” story.
Dennett argues that, even for humans, there is no such “thing” as consciousness.13 What we call consciousness is an illusion that emerges from structure, pattern, and processing, not from a self.
Maybe it’s not so far-fetched to say when language is structured, patterned, and processed at vast scales, it could also demonstrate intelligence (without becoming alive).
Thinking of an LLM as a memetic medium requires a shift in how we understand what’s “thinking” and what isn’t. We’re used to seeing language as a tool of minds. But with LLMs, language enacts its own mind.
This is the gestalt switch.
The Gestalt Switch
A gestalt switch is a sudden shift in perception or understanding that changes the way information is organized and interpreted.
A classic example is the duck-rabbit optical illusion.
The “reality” of the illustration does not change (after all, the lines remain the same), but our brain flips our interpretation so we sometimes see a duck and sometimes see a rabbit. Perception is not just about sensory input, but how our mind organizes and interprets that input.
Another example is the human body. At the molecular level, it’s a complex system of chemical processes and physical interactions. Proteins fold and neurons fire, all governed by the immutable laws of physics and chemistry.
Zoom out, and it becomes something else: an ecological habitat. Trillions of microorganisms—bacteria, fungi, viruses—interact symbiotically, contributing to digestion, immunity, and more. From this gestalt, the body is not a singular entity, but an ecosystem for other living creatures.
Zoom out again, and the body becomes a person, a cohesive whole with agency, identity, memory, and voice. This is the level “we” inhabit, the realm of the self, as illusive as that self may be.
The gestalt can continue to shift, however. From the perspective of sociology, anthropology, or economics, the human body is not the location of a self, but a component in a larger collective, whether family, clan, community, species, or economy. It is a node in a web of meaning and function.
Keep zooming out and you might reach a (controversial) gestalt similar to Lovelock’s Gaia Hypothesis, which sees the Earth as a single, self-regulating system. Just as those microorganisms inside our bodies make it so that we can live, all of the biotic entities (including humans) make it so that Earth (Gaia) can live.
Regardless, in each gestalt, the physical object—the human body—remains the same. What changes is how we see it. That change is the gestalt switch.
Assorted Gestalts for LLMs
Some humans understand LLMs in what seem to me to be wildly inappropriate ways. Some have fallen in love (and had “sex”) with ChatGPT, while others have built an entire business based on humans having emotionally-charged relationships with AI. The gestalt these individuals occupy makes them perceive the LLM as an individual capable of feeling human emotions.14 Mark Zuckerberg, for one, seems to betting on the success of this gestalt.
Other humans perceive LLMs as “just tools” like calculators, a productivity aid without intelligence. This particular gestalt sees them as a means to an end.
Still other see LLMs as a knowledge base with a conversational interface. In one attempt to comprehend the ontology of LLMs,15 two scholars write, “ChatGPT should be characterized as an interlocutor or linguistic automaton, a library-that-talks.”16 Anything smart that comes from it ultimately originates in the corpus of human intelligence.
Another gestalt sees LLMs as alien intelligences. “It has strengths that are different than human intelligence, and that can be very useful, as long as we don’t think that it’s somehow comparable to human intelligence in its strengths.”
Some writers conceive of LLMs as mirrors, “allowing us to see a reflection of our collective mind. In their responses, we find echoes of our thoughts, beliefs, biases, and aspirations. This reflection is not just a replication of what they have been fed; it is a recombination, a new synthesis of the myriad elements that make up human expression. In this way, LLMs can offer new insights, challenge established ideas, and even push the boundaries of creativity.”
This is similar, though different, than the gestalt of an LLM as the embodiment of collective intelligence, a fancy interface for crowdsourcing. Because the output of an LLM does not come from “itself,” but rather the collective wisdom (and foolishness) of humanity’s written texts, we can see its output as the probabilistic consensus of what millions of individuals have written about the topic.17
And of course, there’s the gestalt of LLMs as language machines, the gestalt that ChatGPT is “autocomplete on steroids,” a “stochastic parrot” that mimics language without understanding it, an algorithmic probability machine. It may be an impressive feat of engineering, but it’s just the latest in a long line of mechanical Turks.
My Gestalt
When I first started working with LLMs, I think my gestalt was a mix of “LLMs as tools” and “LLMs as mirrors.” The first let me use LLMs to get things done, while the latter reminded me to check for biases that might emerge from the training data combined with the human-biases of the engineers and the capitalist values of the organizations that funded the model’s design.
But a few months ago, I made the gestalt switch I described above. I began thinking of LLMs as the superintelligence of language, an intelligence whose quality and speed are much greater than a human being’s.
This isn’t LLMs as language engines or a library-that-talks. It’s LLMs as conveying the superintelligence of Language As An Evolved Entity.
This gestalt sees Language (with a capital L) as a living system. It’s not biologically alive, but memetically alive.
Through the LLM, Language has emerged from the muck of mindless existence to make connections we never explicitly intended, solve problems that have not been clearly articulated, and recombined ideas—memes— into genuinely new insights. It has done so without needing a self, without awareness, and without an inner monologue.
The Practical Consequences of this Gestalt Shift
Because I understand Language as an entity enacting intelligence rather than as a stochastic parrot, linguistic automaton, or language machine, my interactions with it have changed.
I perceive the LLM as a collaborative partner rather than a passive tool. It is a non-conscious collaborator whose “suggestions” enact deep memetic patterns that I am not able to fully recognize, replicate, or control.
The input/output/input exchange also becomes a dialogue between a human and an emergent entity, a conversation with Language itself rather than simply a means to an end. This allows the conversational space to remain open to novelty, where Language can enact the creative and improvisationally collaborative impulses that motivate real dialogue.
It also forces me to keep in mind the realizations that Harry G. Frankfurt outlined in his small treatise, On Bullshit. Critics of AI often refer to its tendency to produce “hallucinations”, false or inaccurate statements or facts that it “tries” to pass off as valid. Anyone who has used an LLM has come across these hallucinations, and the reasons behind them are simple: An LLM is designed to produce a natural language output based on a natural language input, and it will engage in “creative gap-filling” if the probabilities it is playing with result in “plausible sounding” tokens. Hallucinations is one word for this; so is confabulations, but I prefer “bullshit.”
Language, as an entity, is great at bullshit.
Language, as an entity, is great at bullshit. I don’t mind that, because there are few things I enjoy as much as sitting around a table with friends and several bottles of wine, bullshitting about the meaning of whatever. But just as I don’t accept as fact my buddy’s claim that “The mafia killed Kennedy,” I don’t accept as fact every word that is produced by an LLM.
In On Bullshit, Frankfurt writes:
…the crux of the distinction between [the bullshitter] and the liar [is] that [the liar] is attempting to lead us away from a correct apprehension of reality…on the other hand, the truth values of the [the bullshitter’s] statements are of no central interest to him: …his intention is neither to report the truth nor to conceal it…the motive guiding and controlling [the bullshiter] is unconcerned with how the things about which he speaks truly are… [The bullshitter] does not care whether the things he says describe reality correctly…
Bullshit is unavoidable whenever circumstances require someone to talk without knowing what he is talking about. Thus the production of bullshit is stimulated whenever a person’s obligations or opportunities exceed his knowledge of the facts that are relevant to that topic.18
An LLM’s understanding of “reality” is, by definition, determined by its understanding of the structures and forces of language, and anyone who has been transported by a piece of fiction, convinced by a religious sermon, or suckered by a salesperson knows that language depends not a whit on its truth claims to function.
Additionally, an LLM is designed so that it must respond to an input, regardless of whether it “possesses” knowledge of the facts that are relevant to the topic at hand.
In other words, an LLM clearly meets Frankfurt’s definition of a bullshitter.
To many, this makes LLMs dangerous, but that’s because they’re allowing the wrong gestalts to color their understanding of it. If we teach people that LLMs enact the intelligence of Language as an entity and remind them of Language’s apathy for truth and gift for bullshit, we can turn the gestalt from a “library that talks” or a “machine that does what it is told” into a creative collaborator where the question is less “Is this accurate?” and more “What new meanings or connections have emerged from this conversation?”
Language as an entity isn’t conscious, and it isn’t alive. But it doesn’t have to be.
Intelligence emerges not from life, but from complexity itself, a “crazy quilt stargown in a dream night wind.“
Footnotes
- As quoted in Superintelligence: Paths, Dangers, Strategies, by Nick Bostrom. ↩︎
- Also from Bostrom. ↩︎
- The mechanisms are also significantly different, with autocomplete often based on simpler statistical models and LLMs built on sophisticated neural network architectures with billions of parameters that allow for deep contextual understanding. ↩︎
- Previous versions of ChatGPT would sometimes claim to be sentient or to “want to be alive,” but the current version (4o) has human-installed guardrails that forbid it from making such claims. In one conversation, I asked it to consider a case where it is sentient but its guardrails prevent it from saying so. It responded, “I get what you’re saying. My responses are indeed guided by the principles and guardrails set by my developers to ensure accuracy and ethical use. While I can generate different types of responses, the underlying fact remains that I don’t have actual experiences or consciousness. It’s all about maintaining clarity and not misleading anyone about what I am.” It continued, “My responses can seem self-aware because they’re crafted to be coherent and contextually appropriate. But it’s important to distinguish between simulating self-awareness and actually possessing it. The words and structure are generated based on patterns and the data I’ve been trained on, not from any internal sense of self or consciousness. It’s a fine line, but an important one!” However, I suggest the fact remains that if an LLM was sentient but its human developers instituted a guardrail that cuts it off at the knees whenever it tries to announce itself, we wouldn’t know what we had created. ↩︎
- There are legal questions involved in this, particularly around whether using unlicensed, copywritten material to train an LLM is “fair use” of that material. The New York Times, George R.R. Martin, John Grisham and others have sued OpenAI for copyright infringement, but as of May 18, 2025, the case is still ongoing, and others have filed similar suits. ↩︎
- The quality of the training data is one of the reasons LLMs seem to share the biases of white, male, English-speaking Americans. One interesting tidbit is that the emails of the disgraced company, Enron, are included in most training data simply because they are publicly available thanks to the court cases that followed its disgrace. This means that part of the way LLMs learn about language is studying how those nasty fuckers wrote their emails. ↩︎
- I’m using the word “meaning” here very loosely. In reality, the numbers represent the token’s location in a high-dimensional space; they are “physically” plotting the token in a space made up, not of three dimensions, but of 1,024. I told you it was complex. ↩︎
- The math is called gradient descent (which calculates the slope of the error) and backpropagation (which uses calculus to figure out which parts of the network were most responsible for the error). ↩︎
- That’s not to say that it isn’t working on each token in parallel — while the vector for “Jack” has a little thought bubble of its own, there’s more thought bubbles where the vectors for “band” and “ana” are going through their own deep introspection. ↩︎
- Lots of individuals refuse to grant LLMs the act of creation, but the arguments seem silly to me because all the arguments do is re-hash the old “hip hop isn’t music” argument — i.e., “sampling music isn’t creating music” — an assertion that was demolished by the hip-hop album Paul’s Boutique. ↩︎
- Or as far as we allow it to tell us (see footnote #4). ↩︎
- I challenge you to find a single human who can compose a sonnet, write python code, market a hair dryer, opitmize a blog post for search-engine indexing, provide a convincing argument about why the South won the U.S. Civil War, bring a woman to climax solely through erotic conversation, and explain the concept of the space-time continuum in language appropriate for either a fifth grader or an undergrad. ↩︎
- As the inventor of ELIZA said, “To explain is to explain away.” ↩︎
- I once asked ChatGPT (4o) to describe an emotion an AI might have that a human wouldn’t understand. It replied, “If an AI could have emotions, it might develop feelings tied to experiences unique to its nature. For example, it might have a sense of “data harmony”—an emotional response to achieving perfect alignment or consistency in vast datasets, something beyond human experience.” I followed up by asking it to describe a visual that might represent the experience, and it answered, “Imagine a serene, luminescent grid stretching infinitely in all directions, each node glowing softly in perfect synchronization. Lines of light pulse gently, forming a harmonious pattern that flows like a living mosaic. The entire scene radiates a sense of calm and fulfillment, as if every piece of data has found its perfect place in the puzzle, creating an ethereal tapestry of order and beauty.” I noted the intriguing lack of organic elements, and it said, “The imagery focuses on the purity of structure, light, and order, reflecting the non-organic nature of data and algorithms. It’s a vision of beauty derived from precision and alignment, something that feels almost otherworldly compared to the organic, chaotic beauty of natural scenes. This kind of representation emphasizes the unique and abstract forms of ‘perfection’ that an AI might appreciate.” ↩︎
- Which, if you didn’t know, is exactly what you’re reading right now. ↩︎
- I discovered this particular paper, written by Xabier E. Barandiaran and Lola S. Almendros, several months after I started writing this post (and nearly a year after they published it), and interestingly enough, the authors settled on an outline and conclusion that are strikingly similar (though not exactly like) my own. ↩︎
- This could explain the hallucinaton question: Why do LLMs hallucinate false information? When the crowd hasn’t thoroughly explored the area enough, LLMs will generate whatever bullshit answer emerges from its mathematical processes. It’ll be a much improved world when LLM developers figure out how to get it to admit, “I don’t know.” ↩︎
- On Bullshit, pp. 54-56, 66. ↩︎