Description
When are you truly creative?
With insight from Eric Weinstein and the visionary movie Blade Runner, we explore the realm of ai and creativity, the secret sauce of large language models, how smart we really are and what jobs AI can and cannot replace.
Script & Sources
Eric Weinstein: “And then what we find is that every sentence that we speak can be tokenized, and then it’s just an input output machine, where you say, if this is the input, what has been the output. And you let the thing read some giant corpus of text.
And so it turns out that that’s pretty good to fool us into thinking that we were intelligent all along and the big discovery is: mostly we’re not.”
AI has surprised us with managing language and creativity more so than logical tasks
In 2023, the AI revolution came into full swing, with ChatGPT and StableDiffusion type large language models, continuously pushing the boundaries of what was deemed feasible.
From a fake historical photograph winning a prize, to the pope in a Balenciaga jacket, artificially generated images have fooled us into believing they are real, even just in this first year of the revolution.
It is clear that AI here to stay. Moreover, before Midjourney and ChatGPT, the one thing we believed for the longest time would be the most difficult for it, to be creative and to speak like a human, was actually a rather easy problem to solve.
Do we work similarly to LLMs?
However, very few people have started asking the question that this really throws up:
If large language models are able to emulate human speech by remixing parts of catalogued statements, are we maybe doing something similar ourselves?
Among all species on this planet, speech is unique to humans, at least at this level of complexity.
This is the first time in human history that we have a non-human conversation partner. A new ‘other’ that can respond, without any pre-written scripts. It also marks the first time that we are confronted with the reality of the question if speech equals actual intelligence, or just a complex understanding of syntax and relevance.
But maybe we got the question the wrong way around.
Ignorant of being a machine
In the movie Blade Runner, Deckard is tasked to hunt down and kill rogue androids, or replicants as they are called.
As replicants look exactly like humans, essentially hiding in plain sight, the blade runners use the Voight-Kampff test to be able to tell man from machine.
This empathy test measures the contractions of the iris muscle, respiration and heart rate in response to emotionally provocative questions, Using this test, Deckard finds out that Rachel is a replicant, who is not aware of herself being a machine.
“You think I’m a replicant’, don’t you.”
Even though she was never told about this, she suspects Deckard’s judgement and decides to confront him with photos of her childhood.
Deckard in turn, tells her about some of the most intimate memories that Rachel has, demonstrating to her that she only remembers implanted memories from someone else.
While it is never directly mentioned, that Deckard himself might be a replicant, there are many little telltale signs that give us a sense that Deckard may not be entirely what he appears.
In the world of Blade Runner, real animals are extremely rare and expensive, and many people have acquired fake ‘robot animals’ instead, as it has become a symbol of status.
One of these animals is seen when Deckard first meets Rachel at Tyrell Corporation. And one of its tell-tale signs are the reflecting eyes.
We can see these tell-tale replicant eyes in all the replicants we know, Zora, Leon, Roy, Pris, even Rachel. And in one particular scene, for the slightest of moments, we can see the same reflections in Deckard’s eyes.
Are we but simple machines?
Are we, like Deckard, unaware that we are but fancy machines? Are we like Rachel, in denial about an important insight that can help us grow?
Have we maybe fooled ourselves into believing that what we do day in and day out is intelligent? To answer this question, let’s find out what mechanism made AI and ChatGPT go leaps and bounds in only a few years, and understand how it is that the transformers can create such sophisticated sentences, that they even passed the Turing test.
Attention is all you need
AI is in many ways still in its infancy, but the outputs we have seen over 2023 are often stunning and mind-blowing.
However, if we look at what the mechanism of the transformer, as it was laid out in the 2017 paper Attention is all you Need is, the basic parts of the algorithm are actually quite simple.
In essence, the algorithm indexes the meaning of each word in a sentence according to their meaning, and also catalogues the order and relation of each word to each other in a matrix.
With this indexing, run across the meaning of words, the grammar, and any conjugation or other level of the language, the algorithm is able to understand what pronouns are referring to, and many other contextual relationships. This is something that was nearly impossible for older natural language processing algorithms.
Provided the model can receive lots of training, the networked logic between the queries, keys and values becomes richer and richer, and in a sense the model becomes better at understanding the associative ‘density’ between different words, and also knows how to form sentences with correct syntax.
The surprising element about these large language models is that once they have received enough training, all they do is look at the query, and then assemble a reply from words entirely based on simply ‘guessing’ the most likely next word. Given a large enough database of possible strings, scraped from the internet, you essentially get the models best guess at what you’re looking for, scrapped together from bits and pieces it indexed from the internet.
So, can we really be surprised that when Google AI researcher Blake Lemoine asked Google’s AI LaMDA in 2022 if it was sentient, it answered that “it is, in fact, a person.”? The algorithm was asked a question about AI and sentience, and the most relevant sources to draw from are the countless science-fiction books in which robots did become alive, and imitate the content from there. After all, relevancy of the terms of the query, is all that matters to the generative algorithm of ChatGPT.
Surpassing Limitations
In Blade Runner, it is ironically the replicants that end up surpassing their own limitations. Both Roy and Rachel come to their own way of accepting who they are, and exactly this trait allows them to grow beyond their programming.
In 2024, it is up to us now to understand what truly makes man different from machine.
Eric Weinstein:
“most of what we do is not generally intelligent, it's just repetitive. That's where computers spend almost all their time, in repetitive behavior.
And so, excellence, and what we do in school in general, trains you to do something that a computer will do better than you will.”
If our thought and speech patterns look so similar to ChatGPT, which is simply training and guessing the next best response, what is it that makes humans truly unique?
Eric Weinstein:
And so think very carefully about what you want to train your children to do. Think about how much of your day is actually spent in non intelligence.
Every act of creation is a violent act, which is difficult to do. So think about what portion of your day you're actually doing something that's never been done, even if it feels minor to you, and take your own genius seriously