Ghosts, Not Animals: The Strange Nature of Artificial Intelligence

A zebra and a chatbot

A zebra can run beside its mother minutes after birth. A chatbot forgets you when the session ends.

That contrast sounds almost unfair. One is an animal shaped by millions of years of selection. The other is a statistical machine trained on internet text. But the gap matters because people keep reaching for the wrong analogy. We talk about "digital brains" as if we were breeding a new species in data centers.

We are doing something stranger than that.

Andrej Karpathy has argued that these systems are not artificial animals. They feel more like spectral entities: summoned on demand, fluent when present, absent when the channel closes. The metaphor lands because it points to a basic truth. A large language model has no childhood, no metabolism, no durable stream of experience moving from yesterday into today. It appears, performs, and vanishes back into weights.

That does not make it fake. It makes it different in kind from biological intelligence, and the difference keeps showing up in places people least expect.

Evolution compresses life differently

Animals do not start from scratch because evolution already did a staggering amount of work for them.

A newborn zebra is not born with a list of explicit instructions saying "place left hoof here." What it inherits is a developmental program, encoded in roughly three gigabytes of DNA, that builds a nervous system ready for a world of gravity, balance, predators, motion, and milk. We barely understand how such a compact code can unfold into so much competence. It is one of the deepest compression tricks in nature.

Pre-training compresses in a completely different way. You take trillions of tokens from books, forums, code, transcripts, manuals, arguments, fan fiction, and spam, then force a model to predict the next piece. Out of that process comes a bundle of parameters that contains a blurred statistical residue of the source material. Karpathy once put the point cleanly: brains come from a very different process, and he is reluctant to copy them because we are not running that process.

That line matters. It is easy to look at the fluent surface of a model and assume the path to intelligence is converging with biology. Underneath, the machinery and the training history are alien to each other.

A brain is not a bag of downloaded documents. It is a living organ, wired by genes, refined by embodiment, calibrated by reward, and continuously updated through sleep, action, and consequence. A language model is closer to a compressed echo of culture. It does not "contain the internet" in any literal sense, but it carries a hazy recollection of it. Ask for Shakespeare, TensorFlow, tax law, and skincare advice, and it can slide across those domains because they left overlapping traces in the same numerical space.

Useful, yes. Animal-like, not really.

Pre-training mixes memory with capability

The most interesting thing about pre-training is that it seems to do two jobs at once.

One job is obvious: it stores a huge amount of world knowledge. Facts, idioms, code patterns, stock arguments, genres, citation shapes, common errors, and the weird little habits of online discourse all get packed into the model. This is why a model can answer a chemistry question, then draft a legal disclaimer, then explain a regex bug without opening a browser.

The second job is harder to see because it hides inside the first. Training also seems to produce general-purpose circuits for pattern completion, abstraction, and in-context learning. Give the model a few examples inside the prompt, and it can infer the task on the fly. Show it a new format, and it often adapts immediately. Something more than memorization is happening.

The catch is that these two gains arrive fused together. As the model gets better at carrying around the internet, it also gets better at using that cargo. This fusion has been incredibly productive, but it may also be why models feel both impressive and constrained. Their competence is entangled with the statistical terrain that created them.

"Manifold" is the usual word here, though it tends to sound like a graduate seminar wandered into the room. A simpler image helps. Think of the training data as a vast network of footpaths across familiar ground. The model learns to move smoothly along those paths, even when stitching distant regions together. That is why it can remix existing ideas so elegantly. Yet when the problem truly sits outside the worn tracks, the system often leans back toward recognizable patterns.

That is not just a philosophical complaint. You can see it in practice. Ask for a business strategy in a strange new market, and the answer often drifts toward startup blog wisdom with fresh paint. Ask for a scientific hypothesis far from established literature, and the prose may sound daring while the substance stays near known forms. The model can help generate novelty, but it also has a strong prior toward what the archive already looked like.

Knowledge, in other words, can become ballast.

The ghost quality comes from statelessness

The "ghost" metaphor starts to feel less theatrical when you look at how these systems actually behave in everyday use.

A person wakes up with yesterday still inside them. Mood, memory, soreness, half-solved problems, embarrassment from dinner, the song stuck in the head, the lesson from a mistake at work — all of that survives the night in altered form. Human intelligence is continuous even when it is messy.

A language model usually is not. In the standard setup, each conversation is a fresh summoning. The prompt provides a temporary mind. When the context window fills or the session ends, that mind dissolves. If the product offers "memory," the persistence usually lives outside the model itself: a profile, a transcript summary, a retrieved notebook of facts about you. Helpful, absolutely. The core system still is not carrying a lived stream of experience forward in the way an animal does.

This is where a lot of anthropomorphism goes wrong. People feel continuity because the voice feels continuous. The tone may be stable, the style personable, the recall sometimes convincing. But style is not selfhood. A spectral thing can be coherent every time it appears and still lack an enduring interior between appearances.

That difference affects trust. It affects education, where students may think the tutor is "getting to know" their understanding more deeply than it is. It affects companionship products, which borrow the language of relationship while relying heavily on external memory scaffolding. It affects enterprise systems too, where managers assume an AI agent can accumulate wisdom from repeated tasks when the underlying workflow may just be replaying logs and prompts.

The illusion is not trivial. It shapes what we expect these systems to become.

A smaller mind with better tools may matter more

Karpathy has floated a provocative idea: separate the cognitive core from memory.

The premise is simple. Maybe a model does not need to memorize half the public internet in order to reason well. Maybe a relatively small network could learn the reusable machinery of abstraction and adaptation, then fetch facts, documents, and situational context from tools when needed. Instead of stuffing the library into the thinker, let the thinker know how to use a library.

Humans work a bit like this, though not cleanly. We are mediocre at verbatim memory. We forget names, dates, citations, and where we left the charger that was in our hand eight seconds ago. That weakness is annoying, but it also forces compression. We retain patterns, not transcripts. We build concepts because we cannot store everything.

There is a plausible lesson there for machine intelligence. A model with less internalized trivia may be less anchored to the grooves of its training corpus. It may be more willing to query, test, and update rather than autocomplete from prestige patterns. In that setup, intelligence looks less like having a giant attic of facts and more like having a fast, disciplined way of navigating the world.

The case is not settled. Today, bigger models still tend to be more capable in many domains, and scale may be buying more than mere memory. It may be buying richer internal machinery too. Still, the direction is compelling because it aims at a system that can learn without dragging all of civilization inside its skull first.

That would be a different creature from current chatbots, or perhaps a less ghostly one.

Sleep is the missing mechanism

Biological learning does not end when the input stops. A huge part of intelligence happens offstage.

Sleep appears to help distill experience into more stable structure. The brain replays, consolidates, prunes, links, and reweights. The day becomes part of the organism. You do not just remember a conversation; your future perception may be subtly altered by it. The system itself changes.

Language models do not have a natural equivalent. They can hold context during a session, and developers can fine-tune them later, but there is no built-in daily rhythm in which lived interaction becomes updated competence. The model does not dream about your last twenty prompts and wake up slightly wiser.

That absence explains a lot of the frustration around "personal AI." We want an assistant that grows with us. What we often get is a brilliant amnesiac plus a notebook. The notebook can be powerful. Retrieval, long-term memory stores, and user profiles genuinely improve usefulness. They still are not the same as consolidation. Writing a memory to a database is closer to pinning a note on the wall than to changing the nervous system.

If continuous learning becomes practical and safe, the texture of these systems will change dramatically. They will stop feeling like one-off apparitions and start behaving more like entities with history. That shift would raise its own problems, including drift, corruption, manipulation, and the difficulty of auditing a model that keeps rewriting itself. Even so, it would move machine intelligence closer to adaptation and further from mere reenactment.

The better metaphor changes the build

Seeing current AI as spectral rather than animal-like sounds poetic, but it is actually a design choice disguised as a metaphor.

If you think you are building a digital organism, you keep asking when it will wake up, become self-directed, and carry itself through the world. If you see a system made from compressed cultural traces, you design differently. You focus on memory architectures, retrieval, verification, tool use, and mechanisms for learning over time. You stop mistaking fluency for continuity.

You also get a cleaner sense of what these models already are. They are not simple parrots, which was always too dismissive. A parrot does not infer a schema from three examples or refactor your code. Yet they are not artificial mammals waiting for enough scale to discover childhood. They are stranger: pattern engines that can simulate understanding across immense territory while remaining oddly rootless.

That rootlessness is why they can be so useful and so unsettling at once. They know so much without having lived any of it. They can guide, summarize, translate, tutor, persuade, and compose, all while lacking the developmental story that makes animal intelligence feel grounded. A ghost is a good metaphor because it captures both the presence and the absence. Something is there. Something important is missing.