I recently met myself. Or something that was supposed to be me. Delphi had been trained on my content. ElevenLabs had been trained on my voice. By any reasonable technical standard, the result was impressive. The cadence was right. The vocabulary was plausible. And yet the moment I heard it, I knew something was off. Not broken. Off. The simulation hadn’t failed. It had succeeded at the wrong thing entirely.

This is the uncanny valley nobody talks about. Not the one where a robot’s face almost looks human, but the one where your own words come back to you, emptied out. You recognize the shape of what you said, but the thing that made you say it is gone. It’s like hearing a cover of a song you wrote, performed by someone who learned every note but never knew what the song was about. The technical execution is there. The reason it existed in the first place is not.

What’s missing isn’t data. It’s not a fidelity problem. It’s a category error. The assumption that capturing enough of someone’s output captures something essential about them.

That replicating the signal with sufficient precision yields the source. It doesn’t.

I keep coming back to the distinction between creativity and production. These aren’t two points on a spectrum. They’re two different things.

Creativity is pre-linguistic. It happens before the sentence, before the image, before the sound. It’s the moment something that didn’t exist begins to take shape. Not as an artifact. As a direction. An instinct. A commitment to this and not that, before you can fully explain why.

Everything after that moment is production. The writing, the editing, the recording, the designing. That’s the craft of turning the creative impulse into something others can encounter. And production is where AI lives. Entirely. It operates on the artifact layer. It can recombine, extend, polish, and mimic what’s already been produced. What it cannot do is reach back into the moment before production began.

This matters more than people think. Creativity can’t be scraped. It can’t be trained on. It precedes the artifact, which means no amount of ingesting artifacts will reverse-engineer the thing that generated them. You can feed a model every essay I’ve ever written, and it will learn my patterns. It will not learn why I chose them, or what I was reaching for when I did.

In this context, aesthetics takes a very different meaning than we normally give it. People treat it like it means style. A visual preference. A tone. A vibe. But aesthetics is what makes direction possible, what makes things different.

When you have real aesthetic judgment, you can walk into a room full of options and know which one to pursue. You can look at a draft and feel one word louder than the other. You can make the call that pushes a project, or let it go. Not because you ran the numbers, but because something in your body told you this is right, or this is wrong, and you’ve been right enough times to trust the signal. Aesthetics is what lets you move. It’s the layer underneath every meaningful decision about what to make and how.

And you only develop it by living with the output of your choices. In practice, over time, through the accumulation of choices that mattered. It’s embodied. It requires a presence that can be affected by what it creates.

Utility, by contrast, is disembodied by definition. Its function is stripped of uniqueness. A useful output doesn’t need to be yours. It doesn’t carry your intuition. It just needs to work.

This is exactly where Delphi’s replication breaks down. It can produce sentences that sound like mine, but it will never feel the weight of choosing one over another. It doesn’t care if it’s wrong, because it can’t mean any of it.

AI is worth using when it moves in the direction of something you’ve already decided to write. Use it to clean up data you’ve already decided matter. Use it to accelerate production on a project whose creative direction you own completely.

But the moment you use AI to figure out what to say, what to care about, what direction to go, you’ve crossed a line. You’ve handed over the one thing that makes the work yours.

I see this constantly now. Someone uses AI to generate a brand strategy, a content calendar, and a positioning document. They present it with confidence. But push on any single point. Why this word and not that one? Why this audience? Why this tone? There’s nothing underneath. No rationale that comes from lived experience. No moment where they chose a direction and felt the value of it. They show up with a disembodied artifact that cannot grow with their thinking.

AI has no meaning, so its output can never stand scrutiny.

The difference between using AI well is about whether you had something to say before you opened the program. If you did, AI accelerates you. If you didn’t, AI gives you the illusion that you did, and that illusion has a half-life. It shows. Maybe not immediately. But eventually, and always.

The real literacy of our moment isn’t knowing how to use AI. It’s knowing what layer you’re operating on.

When someone shows you a beautifully produced piece of content, the question worth asking isn’t “Is this good?” It’s “what layer did this come from?” Was there a genuine creative act underneath? A person who took the time to actively pursue and articulate meaning and created a place for that to meet others.

The risk now isn’t that AI will produce bad work. It’s that AI will produce work good enough to fool us into thinking the creative layer is there when it isn’t.

All technology exists to mediate human communication; that connection is always at the level of meaning.

Organizations will ship strategies nobody actually believes in. Writers will publish essays nobody actually wrote. It will be editorial communication all the way down, with no generative value (value that stays after you leave the room).

The cost of not being able to tell the difference is enormous. It’s not just a quality problem. It’s a trust problem. When you can’t distinguish between a genuine voice and a well-produced simulation, you lose the ability to know who actually stands behind what they’re saying.

This is the new literacy: not prompt engineering, not tool fluency, but the ability to look at a piece of work and articulate its both structured and open-ended value.

In the early 2000s, researchers studying epilepsy patients identified a single neuron in one patient’s brain that fired only when she encountered Jennifer Aniston. Her photograph, her name in text, even a stylized cartoon drawing — the same cell, every time. That is how human cognition organizes meaning: symbolically.

Large language models work dif…

Read more

Read more

Most conversations about AI start with what it can do. This one started with what it actually is.

At the first AI Literacy Salon, hosted by In Process Coaching and joined by guests Noa Dolberg (Google’s Material Design), Cherae Robinson (Head of Platform, Flybridge), and (Founder, Deliberate Works), we spe…

Read more

In a recent AI Literacy workshop, I asked the prompt: *What do you wish an algorithm could predict?* As the answers moved from the practical towards bigger, harder questions, I arrived at a realization I haven’t been able to shake since.

Algorithms cannot predict anything subjective.

That sounds obvious when you say it like that. But the implication…

Read more

The first AI Literacy Salon is here — an in-person gathering that extends the weekly workshop.

Join us on March 26th at Index in Greenpoint.

AI Literacy is a space to consider where AI meets human process — to decipher how these tools can shape our attention, our communication, and the ways we work with computers and each other.

Most conversations ab…

Read more

Marvin Minsky (1927–2016) and John McCarthy (1927–2011) founded MIT’s CSAIL Lab1 and are responsible for many of the field’s foundations, including LISP2 and early attempts at …

Read more

Read more

The opportunity to select a product that promises…

Read more

Trying to compute AGI (artificial general Intelligence) is not new.

How far back depends on whom you ask, for this write-up, I will focus on Minsky, McCarthy, and MIT’s AI Lab as the starting point (but we could easily go back to Turing and beyond).

When Minsky and McCarthy started the lab in 1959, they were very much set on computing general intelligence.

Machines that think, paired with consciousness, and capable of learning.

“There are 2 ways of looking at computing artificial intelligence, you can look at it from the point of view of biology or point of view of computer science You could imitative the nervous system as far as you understand the nervous system or you can immediate human psychology as far as you understand human psychology”

To annotate: “looking at it from the point of view of biology…” means the kind of “wet” programming the brain does, neurons, axons, and so forth. Simulating such process is likely to be a reference to neural networks, which are (broadly speaking) sets of networks modeled around the way the brain works: namely, taking a stab at clustering logic in proximity and generating hierarchies of information. “Looking at it from the point of psychology” is equally as provoking, as it refers to human intelligence as buckets of knowledge.

To push that slightly further before annotating, here is a quote by Marvin Minsky, part of a monologue in Machine Dreams.

“In order (for a machine) to be intelligent we have to give it several different kinds of thinking, when it switches from one of those to another we will say that it is changing emotions” “Emotion (in itself) is not a very profound thing, it’s just a switch between different modes of operation”

This is the other way of computing a brain. Clusters of knowledge and sets of actions are compartmentalized in different vertical buckets and switched by emotions.

Beyond being offensive to humanity, this argument is also straightforward to debunk using a quick thought experiment.

Imagine an intellectual person sitting in a chair, doing nothing at all, staring into thin air. That person is conscious and intelligent. Her lack of action does nothing to rob her of the consciousness and intelligent title.

In other words, intelligence is not conditioned by action. It need not be modeled around goals, nor operational switches.

This is a critical point to stay on. This idea that we can compute human intelligence (we can’t), or that the brain is a computer (it’s not), is the underpinning belief that has been fueling generations of researchers, keen and persistent in their pursuit to compute a brain. Allan Watts refers to life as an ongoing oscillation—a continuous frequency of the brain.

You can imagine your intelligence (or soul, or consciousness) as a vibrating frequency. Like the trajectory of an analog sound wave as it travels through space. It is continuous, and unlike a digital factor, it is not broken into a set of high-fidelity instances.

A digital sound wave — say an MP3 file — is economical in storage because it can slice the edges of the frequency range, and then create a high-res sound–alike composition.

Let’s say that I could somehow peer into your brain and start creating a map of your intelligence. It may take me up to a few decades, but eventually, I will succeed in computationally solving your intelligence.

The problem is that I only solved the computation of your brain, and only at one point in time.

In other words, I only solved one instance of intelligence.

Anchored to a point in time and a subject. The brain, as is consciousness and intelligence, is an ongoing analog frequency. Not instance-based digital permutations.

For a much broader palette of opinions on the prospect of AGI (artificial general intelligence), I highly recommend What To Think of Machines That Think, by John Brockman.

If you’re still reading, it is safe for me to assume that you’re on board with my pro–human view. That is the view that the brain is not a computer problem, and hence we should abort the idea of AGI. Narrow AI, on the other hand, is alive and strong.

Once we accept the premise that AGI is of no use, we can start identifying the opportunities in narrow AI. Think of your calculator as a narrow expert in calculation, think of a medical journal scanner as the best tool we have to consume terabytes of medical journals, and machine-vision algorithms as the best face recognizer.

We can do the same tasks the machine is doing: calculate on a piece of paper, skim journals, and classify images, but only to a certain extent. Let’s draw this.

Intelligence on a narrow plain

On the left, there is ‘0’, no intelligence is being used, and nothing is being done. We can learn to perform a task, and as a next step, we might relay this knowledge to a machine so that it could hyper-mathematize it.

We can foresee improvements in computing power, access to data, and other technologies pushing the machine’s ability into infinity and the unknown. ON A NARROW DOMAIN. A singular plain. Life, as a system, contains many of these trajectories: flying, writing, internet browsing, coding, dancing, and driving. We learn new tasks, and some of them can graduate to automation.

New domains are added, while others become obsolete. For example, a self–driving car technician, or a horseshoe maker, respectively. We can imagine the system of life — i.e., humanities — as made of an infinite amount of these single line trajectories.

What would AGI need to achieve

Under this lens, we can position AGI in the bottom right, as holding infinite ability in infinite domains. But we established the fallacies of that view, so that leaves us with a much more current and useful alternative.

In the absence of AGI, is it up to us to navigate this bend? Crossing and linking disparate skills and disciplines. We hold an intellectual monopoly in that regard.

Machines are competent in their unique domains, but are blind to anything else; they only compute and extend the steps we relayed to the domain.

Another way of thinking about it is when we come up with a new skill or a technology, we might slowly improve it, maturing the domain for other humans to participate in it. As part of that onboarding, sets of instructions need to be written. At which point are the domains ready for a machine to excel in them?

That machine is domain-specific and holds no intelligence. Its algorithms could do a lot of things that we could never do. For example, untangle messy data or make assumptions about the future.

But the bend is uniquely creative and can’t be mechanically produced.

This is the core of it. If your product or service is narrow by design, then you’re open for a machine to excel, or replace you.

This is not limited to unskilled jobs, as the machine holds no interest in your prestige. It holds no interest at all. The system can perform tasks efficiently. And if you’re job is on a narrow domain, and can be broken into steps, then you’re not using your human advantage properly.

Our unique, innate ability as creative humans is a crucial aspect of our identity. It will reveal currents in automation, AI, and innovation opportunities.

Nitzan Hermon is a designer and researcher of AI, human-machine augmentation, and language. Through his writing, academic and industry work, he is writing a new, sober narrative in the collaboration between humans and machines.

Welcome to AI literacy, a new section of Being in Space (updated breakdown of sections).

It builds on other writing and spaces (On Creativity, What is Work, Everything Will Happen) that were looking to understand (1) what all of this (tech in this case) means, (2) for you.

You can expect links and essays, original writing, and other material that will h…

Read more