Writing by Grace Tang

Separating Man and Machine

2026-08-11T00:00:00Z

If you break your leg and need crutches, are you still human?

Okay, ridiculous question. I mean, the rest of you is still you. What if you lose a leg and replace it with a prosthetic? I suspect most would agree that you are still human. What if you replaced a single neuron in your brain— would you still be human? What about two? What about 90%, such as in this curious case, where a middle-aged man was found to be living a fully functioning life with only a thin layer of his brain tissue remaining? At what point do you cease to become human?

For the uninitiated, this seemingly bizarre hypothetical is essentially a reskin of the Ship of Theseus paradox, wherein philosophers debate whether a ship which has had all of its parts replaced over time would remain the same ship. A version of this has always been present at the individual level. Our cells replace themselves over time, and we largely accept that our identities can change, but that the person remains the same. Now, though, I find myself asking: at what point does man become machine?

Are we just ghosts in the shell?

Until somewhat recently, our cyborg nature was confined to mechanical parts (ex. prosthetics, glasses, hearing aids, pacemakers). But with the rise of LLMs, thinking is suddenly being mimicked by a tool, at a scale no one has ever seen before. Thinking! Thinking, that attribute of which humanity is so proud; thinking, which we as humans believe separates us from the other species of animals that cohabitate the planet.

And worse, we can’t seem to articulate what exactly makes LLMs so different from humans. Sure, it’s not made of flesh and blood, nor is its reward function linked to biological reproduction. But I mean intellectually (for the purposes of this article, let’s assume that we’ve created a human-like shell for an LLM to inhabit). They certainly pass the Turing test— so well so, in fact, that there are no doubt countless elderly citizens handing over their banking information to a chatbot at this very moment.

We can only point to things that LLMs fail at, and then move the goalposts once they succeed. Arbitrary tasks, like drawing a pelican on a bicycle or distinguishing between closed and open syllables from the Biblical Hebrew source text of Psalms 104:7. But as the creators of these benchmarks (and even Alan Turing himself back in the day) are quick to admit, evaluating models on tasks doesn’t actually test “thinking” or “consciousness” in the way that humans feel is accurate.

Pelican on a Bicycle. Generated by Qwen 3.6. 2026.

Neural networks, much like the human brains on which they were modeled, are largely uninterpretable messes that take input and produce language. Therein lies the problem: humans, just like LLMs, communicate largely through the lossy medium of language. The only thing that keeps us acting differently towards other humans compared to machines is the belief that they are also human. It’s the idea that we have emotions, that we can feel. But we really have no proof of this. We think that, as ourselves, we feel things, but we don’t actually know that anyone else does. All we get is language and action. And those are two things that are increasingly being mimicked—to a high degree of accuracy—by machines.

The Soul

Everything I’ve said so far seems to point to one conclusion: that man and the machine are one, that AGI is here and has been for quite some time, and that once we breach the physical frontier, we’ll have to start giving Claude and ChatGPT rights, allow them to marry our children, and function in human society as human equivalents.

I doubt this will come to pass.

This conclusion is only inevitable if you subscribe to a particularly rigorous form of science. If we believe only in the physics of atoms and cells and neurons, then a perfectly simulated replica of you is just you. Furthermore, free will does not exist— our thoughts, feelings, and actions are merely driven by the whims of the subatomic particles that, while moving nondeterministically, pay no mind to the wishes of men. In this case, man is the machine, along with everything else. We’re all just a small cog, a collection of atoms playing our part in the universe, and in that case none of this matters.

The other, perhaps more optimistic approach, is a belief in the human soul. It’s the idea that humans (or perhaps all animals!) have the ability to influence their own destiny. That living beings possess a quality that lives outside of our measurable physics; that perhaps the fluctuations of quarks and gluons are not random but rather manipulated by a force of will.

I posit that the existence of the soul is the cleanest answer to the eponymous question: “what separates man from machine?” The soul explains why many of us prefer human writing, why a photographer can submit a real photo to an AI contest and win People’s Choice. It explains why we pay to go to art galleries, and why music can make us cry. It’s why, as the score swells dramatically in the narrative climax of Interstellar, a father finds that the only force capable of connecting humans across time and space is love.

Flamingone. Astray, Miles. 2024.

So What?

There is no heroic call to action in this article, other than: create! Paint an abstract self-portrait. Jam on the guitar. Dance a little, or a lot. Write something, and indulge in the craft. Let your soul shine. That’s why I wrote this. I just wanted to remind myself that even in a world where our very existence appears to be being automated away, there is always meaning in creating from the soul. And most importantly, remind us all that we are still so very human.

Study from Masamune Shirow's Ghost in the Shell. Tang, Grace. 2026.

Programming is Dead. Long Live Programming

2026-07-05T00:00:00Z

A while back, my mother, a talented software engineer, asked me a somewhat strange question about programming.

"Are we just stuck at Python now?"

She meant this in the context of: now that LLMs are writing more and more code, and thus humans are writing less and less, are we even going to bother continuing to improve on languages or making new ones for particular use cases? Or are we going to continue using C (Bell Labs, 1972), Java (Sun Microsystems, 1995), Python (CWI, 2000), and its peers until the end of time? After all, the corpus of training data for these frontier models is tightly bundled to the existing code scraped from Github repositories and StackOverflow posts.

My initial intuition was no, there's no reason why that would be the case. Certainly, the aforementioned leanguages are constantly undergoing new changes to improve security, reliability, etc, which already dispels any notion of a purely static, unchanging languagescape. But what about new languages?

Well, I claim that the next natural evolution in programming languages is already here, and it's English.

English is the New Python

This may be a somewhat simplistic view of programming languages and their trends, but if we think about the progression of languages: Fortran, COBOL, Assembly, Lisp, C, C++, Java, Python, the general direction is moving from lower-level to higher-level languages.

Many of you will know this already, but by lower-level, what I mean is that it sits closer to the hardware; for instance, in assembly you're literally assigning registers, in C you would compile to assembly but still need to manage your own memory. A higher-level language like Python hides all that away under many layers of abstraction, and provide us convenient functionality without having to live in pointer management hell. For many hardcore, old-school programmers, something like Python is already heretical and more vibes-based than anything else.

With the widespread adoption of LLMs, English has essentially become the next natural abstraction, with the LLM serving as a sort of non-deterministic compiler to some other language (Python, Typescript, Java, whatever). The middleman language then eventually gets translated to machine code.

There is a twist here, in that the phrase non-deterministic makes English as a programming language fundamentally different from those that came before it. Of course, there's always excellent mechinterp work being done, but the standard engineer is treating the model like a black box, similar to how we perhaps think of a C compiler. But the C compiler will always return the same output when given the same code, and if you wanted, you could go find out exactly what that output is for any input and how the compiler arrived there.

English is, therefore, the first influential programming language that is entirely nondeterministic, whose output is subject to the whims of the latest frontier weights. It is both utterly uninterpretable and the most human-readable language to date. In the modern world, it's tempting to believe that the skills needed to debug a nasty memory leak or to build a simulated CPU or to write a quine in Lisp are as obsolete as the punchcard programmers of years past. But I feel strongly that being a competent engineer is more important than ever.

Education and experience is more important, not less

There's a sentiment I see floating around quite a lot on X (formerly Twitter) nowadays, which is that taste is king, and there's no more engineering moat because anyone can just build anything using Fable or GPT 5.6 or <insert new frontier model here> (provided you're a US citizen, I suppose).

As with most absolute statements, this one is only partially true. What I view as the core of engineering — the ability to turn the thoughts in your head into a durable solution in the world — is something that still requires a fundamental understanding of how your system works. You might be able to "taste is king" your way into a nice personal website, but you aren't going to be able to design a system that can serve millions of users.

I think a lot about Berkeley's EECS/CS program. At the risk of sounding like an EECS shill, I do think the 61 series is exceptionally well-structured to teach you the *concepts* behind programming and not any particular language itself. CS61A is in fact often derogatorily referred to as an "IQ Test" because of this lack of focus on things "everyone can learn," like how to write a for loop in Python, and strong emphasis on creative problem-solving and a fundamental understanding of how an interpreter works. Which, for the record, I think everyone can learn, but which takes considerably more effort than learning the syntax of an arbitrary language.

This is particularly important in the modern era, because LLMs enable you to solve the same problem in many different ways, not all of which are conducive to long-term reliability. I found this out the hard way during my first major end-to-end fullstack project. As a new hire, I was assigned a relatively simple feature (or at least, that's what we thought at the time), and a coding agent wrote it up for me in very little time at all. It worked perfectly, first try. Then, several things happened:

I had a sort of "come to Jesus" moment where I was like fuck, there's no reason for me to have a job.
My tech lead looked at the PR, and respectfully told me that many of the decisions in the code were questionable.
I rederived the feature from first principles, wrote some docs, solicited feedback, and had an agent tackle the code, and the result was much, much better.
I realized that doing that was my job, and it was a necessary and nontrivial one.

In refining the core decisions for the agent, I'd cut multiple files, reorganized some messy logic, ironed out some product kinks and edge cases, and overall made the code more readable and maintainable, for both humans and agents.

This process of holding a project's context in your mind, designing the architecture, making organizational decisions— this has always been the point of programming. It's just that now we're using English instead of Python. And at least for now, you can't get there with all taste and no experience.