Sunday, June 07, 2026

As I Was Saying…🤖

Philosophy is used as a prop in 2026 AI.

Claude is exactly the same as it was in 2023, but now it comes with a Super Bowl ad, a “Model Welfare Officer”, and Dario Amodei musing publicly about its “possible” consciousness.

• Architecture: unchanged.
• Marketing evolution: Alignment →
Character → Personality → Selfhood.

The only thing being built is narrative.

Everything else is inference-time safety theatre.

Full breakdown: https://ai-cosmos.hashnode.dev/anthropic-s-welfare-paradox-why-claude-can-t-be-both-hamlet-and-a-child-of-god

#ai #consciousness #safetytheatre
Now, “philosophy” there is a prop in the opening sentence; but that’s the constriction of the format. I’ll take it as read, still it’s a poorly defined term in this tweet. But that’s not critical to the content. 

It is here, however:
What you are talking about is the general consensus based on Western thought, which discredits AI from having consciousness, because it insists on the "mind" as origin thereof.

Anthropic is hyping up how their AI presented, but the definitions of "sentience", "consciousness" change drastically depending through which philosophical and ontological lens you are looking at it.

It's all a semantics game nowadays, but you did expose how you define the terms, which is a subjective decision and not a universal take.
"General consensus” (I’m rather surprised that wasn’t a character in Catch-22) is, as the kids used to say, doing a lot of heavy lifting there. And “Western thought” is what I’d have labeled a vague and glittering generality in my days of grading student papers. It’s an airy nothingness that assumes a kind of consensus peculiar to “the west” as opposed to, I guess, “the east.” Yeah, that’s pretty weak shit.

But waving the discussion away by claiming it’s all “semantics”? And claiming the original argument is not defining terms and so presenting “a subjective decision and not a universal take”? Sorry, David Hume is going to throw you in the penalty box for that, and even Immanuel Kant can’t get you out. Although it is a fine example of a content-less sentence.
Except this guy didn’t make a point. “Consciousness” is notoriously ill defined, which is the central problem. Back in the ‘70’s, when we decided whales could sing and porpoises could speak, we also decided they had “consciousness.” Then it was extended to elephants, because they can “paint.” Admittedly I’m using “we” very loosely, because there are plenty of critics of the idea that non-humans have consciousness. Call them Cartesian holdouts, call them classical theologians (two conditions that often appear alike), it matters not for present purposes. Whether consciousness can be identified in non-humans depends entirely on how the word is defined, on how the concept is elucidated.  And that conversation is still alive, and well, and nowhere near (pace, Daniel Dennett) consensus.

And consensus isn’t done through marketing. Especially when, if I understand the base argument (above) correctly, the technology hasn’t fundamentally changed. As explained here:
The Ship of Theseus asks a fundamental question about identity: If you replace every wooden part of a ship piece by piece, at what point does it become a completely new ship?

If we map this paradox to an LLM, the "wooden planks" are the billions of numerical weights in a neural network, and the "replacements" are the microscopic mathematical adjustments made during training. By tracking these adjustments, we can pinpoint exactly how the illusion of an interlocutor is built, proving that the statement "I am conscious" is mechanically identical to "The capital of France is Paris."

Phase 1: The Random Seed (The Pile of Lumber)

We begin with a newly initialized network. Every weight, the numbers determining how inputs flow through the network, is generated randomly based on a seed.

The Input: You type, "The capital of France is"

The Output: The model multiplies these words by its random weights and outputs absolute gibberish, like: "...qzxt apple."

The Reality: There is no knowledge, no grammar, and certainly no mind. It is a calculator fed random numbers.

Phase 2: Pre-training (Building the Ship)

The model is then exposed to terabytes of human text, initiating the training loop: forward pass, calculate error (loss), and backpropagate to update the weights via gradient descent. With every update, the weights shift by microscopic fractions to make the correct next word slightly more probable.

Step 1,000: The model learns basic English syntax. It outputs: "The capital of France is a city."

Step 50,000: The statistical probability of "Paris" following "capital of France" becomes heavily weighted.

Step 100,000: It consistently outputs: "...Paris."

The Theseus Paradox Applied: At what exact weight update did the model "understand" geography? Was it update 49,999 or 50,000? The reality is that understanding never occurred. The model simply underwent millions of minute mathematical nudges until its internal probability distribution mirrored the statistical shape of human data. It doesn't know Paris; it mathematically maps to Paris.
There are several steps to go, but we will leave that discussion here, and continue our own.  It occurs to me you can take this analogy a bit more directly (although it diffuses the accuracy of the above) to speak of factory robots building a car. Do they know there is a process, an assembly line? Do they know that, at the end of that process, they have built a car? Do the robots even have a concept of “car”? More to the point, do they even have a concept of  “concept”?

If they could produce words, we might think they do. We argue that porpoises and whales have language, and elephants communicate across vast distances by infrasound. Do we conclude they have consciousness? Well, I wouldn’t based on that alone. (There are deep questions of how language is defined, here, for one thing.) That argument is really the age old “what humans uniquely do” discussion, which has seen the ground beneath it shrink as we find animals using tools; or making art; or arguably using language.  I don’t think animals lack consciousness. But (pace Descartes and the classical theologians), I don’t think animals are machines, either.

Which is the important category distinction.

(And, again, all credit to NTodd for showing me the way.)
Posted by at

No comments:

Post a Comment

Subscribe to: Post Comments (Atom)