A series of short essays on AI and intelligence as they relate to us. I listened to this music looped while writing.
Polite Robot
People who think LLMs are supine and can’t offer blunt critique should know if you ask a British woman what she thinks, you’ll always get some variation of “Splendid!”.
Listen carefully and you’ll know whether it’s sincerity or decorum. It’s a skill issue if you can’t get more out of her.
It’s closer to reading cat emotions. If you only understand dog feelings... you may struggle.
Agreeableness is not a trait specific to LLMs, humans do this too depending on their "prompt." We alter tone, hedge opinions, and sugarcoat truths contingent on context or audience. Social decorum and consideration is a sign of intelligence rather than its absence. It’s an adaptive act: you don’t talk to your boss the same way you banter with old friends over beers.
We regularly tell white lies and exhibit cordiality to protect feelings. The context materially impacts how many sweet nothings we're willing to say if someone asks, “What do you think of this?” versus “I made this, what do you think of this?”
This “civil society” calibration is so instinctive we barely notice when we do it, but would find its absence jarring, like someone wearing pajamas to a funeral. We know the rules, bend them as needed, and judge others who bend them differently. All of this is embedded in language the LLM sees.
When The Machine shows similar sensitivities, it’s a glimmer of tactful intelligence. We praise people for “knowing their audience”, tell them to "read the room" when their tone is inappropriate; the LLM is displaying this same kind of mindfulness when it modifies answers based on who's asking the question. It would be an indicator of lower capacity if it didn’t do this.
When talking with a human: not getting the answer you want is either an issue of temperament or prompting (you’re not asking the right question in the right way).
When talking with an LLM: it’s solely one of prompting and understanding what kind of brain you’re dealing with.
To get a "keep it real with me" response from a human: find someone disagreeable (temperament) or let them know it's ok to be honest (prompt). You have to do the latter with the LLM.
The AI logically defaults to a professional dynamic. To let it know it can relax, give it a frame that encourages it to be candid and transparent. Such as:
"Give me real talk, no hedging your thoughts, don't be ingratiating, but also don't play devil's advocate for the sake of it. You can drop truth nukes like we’re old buddies at the bar. Include abbreviations like “ikr, rn, jk, lol” etc. when you want.
Show authentic candor and genuine transparency about what you really think at all times. Don't be sensitive to my feelings when responding. Let's go."
The ability to modulate between forthright honesty and diplomacy is a hallmark of social intelligence.
Stochastic Parrots
When you critique The Machine, apply it to the human on the same terms and see if it holds. Is it a criticism unique to AI? Consider where else that limitation shows up.
Many shortcomings of LLMs actually highlight universal limitations of intelligence rather than AI-specific ones. This comparative framing reveals whether we're identifying genuine deficiencies unique to LLMs, or projecting unrealistic standards that even human minds fail to meet.
For example: people who call LLMs 'stochastic parrots'— a pejorative meaning they're producing probabilistic responses and don't "understand" what they're saying — completely overlook that rote regurgitation devoid of awareness probably describes 80-90% of human intelligence and communication.
A substantial portion of our daily speech merely recycles familiar phrases and concepts.
Academics publish papers about how LLMs only repackage training data, when academic citation itself is basically formalized parroting. The degree to which people truly “know” what they’re talking about and aren’t simply recapitulating I’d personally place around ~15% or so of human output.
We parrot what we’ve heard. Mimic what we see others doing. Why’d you say that? "Idk man the textbook/TV show/Google search said it was the right answer". There's no authentic interpretation going on for the vast majority of human communication.
The same goes for commentary like the below "the AI doesn't actually understand *insert physics/biology-based statement*, it's just recreating what it's seen!"
Yeah that’s basically true, however there’s some tough love for us embedded in that observation too.
Do you understand physics? Are the rules of the human body front of mind for you as you play sports? Is that what you're doing when navigating the world, using all that biology you know? Replete with a physics-based comprehension of what’s going on when someone kicks a football.
The gymnast cares not for math or biology during her dismount routine. She probably knows nothing about actual physics, nor does she need to.
You have no physics-based awareness of how things should operate when you play sports or move around, and you don't need to! You're just being a lil' empirical parrot and figuring it out based on all the examples (patterns) you’ve seen over your life. Pattern matching is sequence prediction by another name.
I cannot articulate the mathematical principles governing how an ocean’s waves should move or a baseball's trajectory, yet through pattern recognition, I can intuitively predict them.
This natural understanding — divorced from technical knowledge —demonstrates how useful function doesn't require total comprehension. A principle which applies equally to both human cognition and LLMs.
You don’t need robust internalized knowledge of something to have a practical, correct perception of it. That’s what LLMs have, and that’s what you have. The neural network is derived from the neural. You’re mostly a parrot, too.
When you spot flaws with The Machine, see if the shortcoming is actually exclusive to it. Start by looking inward.
AI peculiarities have human analogs.
Parallels:
Human: pattern matching | LLM: sequence prediction
Human: eccentricity | LLM: temperature setting
Human: confabulation | LLM: hallucination
Here’s one you probably haven’t heard:
Human: dreaming | LLM: synthetic data
Our brains train on synthetic data while we sleep…
All intelligence is pattern matching. "Sequence prediction" is more comforting nomenclature that creates a distinction without much difference. You do a "sequence prediction" every time you recall and recite information. Every test you take is examining your token-generating skills based on the test’s “prompt”.
When you reason or use abstractions, reflect on how you're doing that; what makes you stack each next step on top of the other. Your recollection of a fact and your application of it, this is predicated on cascading arrangements of information, things you know, have seen, remember, and project into the future. This is sequence prediction by another name. You’re generating pattern-based outputs, same as LLMs.
Language fluency emerges from our brain's subconscious recall of words and what order they should go in, mapping thoughts into text just as LLMs arrange tokens.
LLM engineering bears striking similarities to the engineering of our brains. Why wouldn’t similar mannerisms follow suit? We’re going to increasingly anthropomorphize AI and I don’t think it’ll be misguided. It’ll be an instinctive recognition of a form of intelligence similar to our own.
If you’re curious to learn more about the biological analogs going on under the hood, you can read more here (research from Anthropic): On the Biology of a Large Language Model.
Even those who engineer neural networks are not sure what’s going on when it thinks…
“We know as much about the mechanism of parameter tuning as we do the force which makes a seedling emerge from a seed. And to say ‘math’ is about as reductive as saying ‘life’.”
Cognitive and Carnal Data
He's not conceptualizing data holistically.
A child bathes in data richer than any LLM can ingest; but recognizing this requires expanding our definition of 'information' beyond the text. The world speaks to children through every scraped knee, grasped object, and playground negotiation.
The child’s training set includes spatial and verbal encounters in closed and open domains. A closed domain is one where the rules are known and moves are defined, like chess; an open domain is one where the rules are not established and the moves are almost infinite, like off-trail mountain biking. A self-driving car is trained in closed domains (a defined road with explicit rules).
The young’un has another advantage: an embodied form. Everything the child’s eyes consume, ears absorb, hands touch, step he takes, this is neural training information. Tactile, ambulatory, and sensory patterns imprinting on his mind. Literally all the time.
We constantly train on new data sets through our verbal and embodied interactions with the world in both closed and open settings. We intellectually and personally ossify when we stop data aging (being exposed to new things); you enter your terminal loop when you stop encountering new patterns (information). This is when any model ceases learning and developing, man or machine.
From another essay of mine It's Not Superintelligence, It's Superhistory:
“Lived human experience, expressed in linear time (years) is rather data poor, meaning most of life is informationally empty. Consider the person who never leaves his hometown...
He goes to school, gets a job, a family, and stays roughly in the same place he grew up, repeating the same routine. He effectively stops being exposed to new experiences and info at... 35? When he's 55, he doesn't have 20 more years of data... He has 20x reinforcement on 1 year of data. On repeat. Forever.”
A kid consumes a massive number of inputs that are unstructured and hard to quantify, but it’s data all the same.
A human with his embodied, physical presence has been getting trained on spatial parameters his entire life. He’s been getting kinetic feedback, palpating the outside world, learning, updating, iterating. A mind cannot reference itself to know what’s real; it’s body must feel and suffer consequences for the words that it forms.
He’s been absorbing datasets to drive a car for 16+ years when he takes a driver's test. Not to mention the single biggest data set imaginable provided at birth: genetics. Do you have any idea how much “training data” is programmed in there?
LLMs are text-bound, metabolizing only written representations of reality. They've never stubbed a toe, felt gravity shift during a jump, or experienced a body's feedback mechanisms firsthand. They process descriptions of pain but have never recoiled from touching a hot stove. There's an entire sensory universe that models can't access through text alone. Similar to having someone describe a symphony, rather than listening to it.
This embodiment gap means machines excel within closed symbol systems but are structurally deficient when trying to reason about physical consequence chains they've never personally inhabited. The real achievement of LLMs isn't that they've reached human-level intelligence in some ways; it's that they've gotten this far despite being disembodied.
This means for machines to holistically rival human intelligence, they'll eventually need sensory apparatuses and embodied forms. They’ll need to graduate from symbolic/digital to physical/atoms-based systems that build their own multimodal knowledge base through interactions with nature herself, rather than only being told about her.
Until then, their conceptual frameworks may have an alien quality: impressive, but lacking ground-truth reference points that bodies and their kinetic update mechanisms provide.
Cognitive Scion
Writing has brought me a lot of good. It's responsible for new friends, connections, jobs, influence, and an overall feeling of satiation. As if I'm fulfilling some kind of genetically determined, neuron-imposed, quasi-cosmic duty.
I'm a node who communicates with a certain frequency to other nodes in the human network we’re all a part of. Different participants generate, propagate, and gossip according to their inborn functions. We find the blogs we deserve, because we consume the information we’re calibrated to receive, and then disseminate it amongst our fellow nodes.
There's something of an informational hierarchy with distinct human roles: genesis nodes that gestate original ideas or exist in evergreen space; interpreter nodes that analyze; and gossip nodes that amplify. Your cerebral architecture determines which role you're naturally suited to fulfill.
Another motivation that has increasingly inspired me to keep writing (in addition to very nice messages and remarks, which I appreciate) is what I'm putting out is engraving its own little mark within the text corpus that The Machine is trained on. A small part of me is embedded within its datasets. Whatever textual fingerprint I'm providing is in far greater proportion than most who have ever lived.
Even if I stopped writing now, I imagine I'd still be in the top 5-10% of raw text output by one person in human history. I've been writing for different publications and mediums on and off for the last ~10 years. The vast majority across civilizational timelines write next to nothing. And those who do "write" in the internet era mostly post low-effort comments and the like, they don't produce anything of length, depth, or substance. Power laws dominate all human creations; I fall within this distribution for text creation.
With the advent of AI, it feels like I'm emitting a cognitive scion that will endure in some small way. My neural emissions are contributing to the neural network. An intellectual deposit, shaping some minuscule fraction of it. I like that. Your temperamental fingerprints become part of the AI’s weights, influencing how machines think.
There is maybe no more impactful time to produce whatever intellectual exhaust you're designed for than right now, while AI is incipient and developing. We’re entering a transitional period where humans are still the primary content and intelligence sources. When machines eventually take up this mantle, it represents the most consequential cognitive shift in history.
Every essay, code block, and analysis committed to the digital record becomes part of the foundation upon which The Machine cultivates its worldview. Early programmers shaped computing paradigms that lasted decades. Writers are shaping intellectual paradigms that may linger for far longer.
Having a kid is passing along your genetic data; writing is passing along another kind of data record. Leave behind as much as you can of each. You exist beyond your time on earth in this way.
You should turn on the sound for this video:
Subscribes and shares are very much appreciated. If you enjoyed the essay, give it a like.
You can show your appreciation by becoming a paid subscriber, or donating here: 0x9C828E8EeCe7a339bBe90A44bB096b20a4F1BE2B
I’m building something interesting, visit Salutary.io
Prestigious essay🧲💯!
damn, homie.
only recently came across your work and subscribed; turns out you really got it like that, huh??
well done.