According to Sam Altman, an AI answering a question vs a human answering it are the same. This reveals an engineering mindset with an efficiency-driven worldview. He treats both humans and AI as information-processing systems having inputs and outputs. His comment comparing human and AI training suggests that he views humans as just inefficient “training runs” and therefore if AI can produce the same output cheaper, it will be better. He seems to ignore the fact that intelligence also includes understanding the social and moral context not just symbol manipulation.
He has shown a tendency to devalue what it means to be human on other occasions too. In a podcast, he said: “It feels to me like we just stumbled on a new fact of nature or science or whatever you want to call it, which is, like, we can create, you can — I don’t believe this literally but it’s like a spiritual point — intelligence is just this emergent property of matter and that’s like a rule of physics or something."
In December 2022, in response to criticisms of the newly released ChatGPT, he tweeted, “i am a stochastic parrot, and so r u.” (He writes tweets all in lower case.) This was a dig at a paper titled “On the Dangers of Stochastic Parrots: Can Large Language Models Be Too Big? 🦜” This paper shows how language models only manipulate the form of language, with neither understanding nor communicative intent. With his tweet, Altman suggests that humans too are little more than machines that manipulate strings of letters.
Many people think that an AI would work something like this: We give a problem to the AI. The AI computes a solution to that problem and then it gives us that answer. That is not what's going on. Despite our use of the term “learning” to describe the process that an AI uses, it actually bears very little resemblance to how humans learn.
For creating an AI, you first need to design an algorithm that contains steps for calculating an “objective function,” namely, a mathematical description of the task. Then, “train” that algorithm on a data set relevant to the task. The result is an abstract mathematical model that represents patterns in the data and an approximate path to the solution. Next, repeat the training process with the right kind of feedback to tune the results. Over many training runs, the algorithm updates the “weights” in its initial mathematical model to get closer to, and eventually converge upon, an optimal solution to the problem.
Yet it seems to come out with impressive outputs. I once asked ChatGPT: "Write a conversation between Bertie Wooster and Jeeves about a patient with locked-in syndrome watching an ant coming towards him." (I have used the free version of ChatGPT. I hear that the paid version is much better.) Within a second, it gave this output:
"I had a rather disturbing thought this morning, Jeeves.”
“Indeed, sir?”
“You know how one occasionally sees ants marching about the place? Tiny black chappies with the air of junior solicitors?”
“Yes, sir.”
“Well, suppose one were lying absolutely still — unable to move a muscle — and an ant were coming toward one across the bedsheet.”
“A most unpleasant hypothetical situation, sir.”
“Not hypothetical, Jeeves. I was reading about this locked-in syndrome. Fellow’s mind perfectly alert, but the body refuses to rally round. Can’t move, can’t speak — only perhaps blink a bit.”
“A grave affliction, sir.”
“And there he lies while this ant approaches. Nearer and nearer. The thing simply keeps coming.”
“Yes, sir.”
I shuddered lightly into the teacup.
“It’s the calm determination of ants that gets me, Jeeves. They don’t hurry. They don’t pause. They advance with the air of creatures who have all eternity before them.”
“Ants are known for persistence, sir.”
“Persistence! That’s the word. If a bulldog advances, one understands the position. There’s drama. But an ant merely toddles onward in a quiet, administrative way.”
“Very true, sir.”
“And the poor chap can’t even twitch his nose. I ask you, Jeeves, what becomes of morale?”
“One imagines it would be severely tested, sir.”
“It made me realise how much of life depends on the ability to perform the occasional meaningless movement. Scratching the ear. Adjusting the sock. Flicking away insects. Civilization rests on these things.”
“A compelling theory, sir.”
“I mean, if an aunt approaches, one can generally escape.”
“An ant, sir.”
“That’s what I said.”
“You said ‘aunt,’ sir.”
“Did I? Freudian slip. Aunts can also produce paralysis of a kind.”
“Undoubtedly, sir.”
I brooded for a moment.
“Do you suppose, Jeeves, that the ant knows?”
“The ant, sir?”
“That the fellow is helpless.”
“I would not credit the insect with such psychological sophistication, sir.”
“No, perhaps not. Still, there’s something rather cosmic about it all. Man trapped in his own body. Ant advancing across the sheet. Silence. Mortality. The whole bag of tricks.”
“Yes, sir.”
“One could almost write poetry about it.”
“Possibly not before breakfast, sir.”
“Quite right. The mind should not grapple with the infinite on an empty stomach.”
It is easy to get amazed by getting such an output in no time. Large language models (LLMs) like ChatGPT are best understood as machines that learn structural patterns that let them produce coherent text. These systems can perform computations on the world’s data far faster than we can. They can create convincing stories because meaningful language has deep structural regularities. They use mathematical data structures to mimic the outputs of human intelligence — our acts of reasoning, speech, movement, sensing, and so on but they can’t understand it. The data storage facility has more in common with a file cabinet than with a living, feeling body.
It can’t solve a problem unless the solution is somehow embedded in the language data it is trained on. AI tools today lack any idea of what their data represent. It is bits stored on the server. These ones and zeros don’t organize themselves into the intelligent awareness of a complex, ever changing world. A data set, even the entire corpus of the Internet, is not a world. It’s just a digital record of measurements that humans have taken of the world at some point in the past.
No comments:
Post a Comment