At New Sapience we are developing digital entities called sapiens. Unlike chatbots, sapiens can learn through language comprehension; that is, acquire new knowledge by decoding human language. In turn, sapiens can express their own internal knowledge to humans via language.
Today, talking to a sapiens is a little like talking to a 4-year-old child: it makes mistakes in grammar, has limited vocabulary and has only a rudimentary knowledge of the commonsense world.
In comparison, OpenAI’s ChatGPT (Generative Pre-trained Transformer), a chatbot launched by OpenAI in November 2022, can on command, compose a perfectly intelligible essay on any desired topic worthy of at least a high school student[i] and even create poems, generate code, and translate between languages.
Its vocabulary, syntax and grammar are impeccable – better than some people – and we are starting to hear more and more about human-level “performance” and beyond.
Naturally, I am getting a lot of questions about how sapiens compare to ChatGPT. Has the race to human level comprehension of language already been won?
We don’t think so. In fact, we’re not even running the same race.
Chatbots are software applications designed to convincingly simulate the way a human would behave as a conversational partner. From the first chatbot, ELIZA, created in 1964 by Joseph Weizenbaum at the MIT Artificial Intelligence Laboratory, researchers have found people surprisingly easy to convince. ELIZA was based on a simple pattern matching and substitution methodology, but nonetheless many people attributed it with comprehension and intelligence.
Humans are not naturally equipped to recognize the illusion of communication. If a chatbot outputs intelligible text, a well-documented aspect of human psychology called theory of mind kicks-in. We feel there must be something there, a mind not unlike our own, communicating its thoughts and ideas[ii].
But chatbots have no more intelligence, knowledge, or comprehension of language than a toaster.[iii] No meaningful comparison of chatbots and sapiens can be made if this is not understood from the outset.
How then can chatbots like ChatGPT output text that can be mistaken for an essay written by a human when they do not, in fact, understand a single word in or out?
ChatGPT is in a class of applications called Large Language Models (LLMs). A language model is a probability distribution over sequences of words. LLMs are created by iteratively “training” large artificial neural networks to find patterns in “datasets” consisting of language written by and for people.
Once trained, an LLM, given a word or sequence of words, can output a statistically likely remix of vocabulary, grammar and syntax based on language in the training dataset. The fact that these outputs are perfectly intelligible by people (even though they might not make sense) is a stunning achievement in statistics and algorithm design.
But impressive as they are and despite being called AI, LLMs have no intelligence. They have more accurately been called “autocomplete on steroids” and “stochastic parrots.”
Nor is there any clear roadmap or compelling evidence that LLMs can ever be intelligent. ChatGPT is more impressive than GPT-3, but a better parrot is still a parrot.
Sapiens are the result of a completely different approach. New Sapience makes no attempt to simulate features of natural intelligence. We make no use of artificial neural networks, datasets, or statistical algorithms. We believe we have created something the rest of the world gave up on as impossible decades ago: a practical, scalable way to endow computers with common and expert knowledge of underlying reality.
Synthetic Knowledge is a model not of language, but of reality itself. Its development shatters long-standing assumptions about the relationship of language to knowledge and knowledge to reality. Its foundation rests upon a radical new epistemology and the discovery that knowledge itself has a unique structure composed of fundamental building blocks analogous to how the material world is composed of atoms.
Sapiens’ core synthetic knowledge already is giving them an unprecedented comprehension of human language, including contextual cognitive learning, the ability to distinguish subjective from objective concepts, and even sense from nonsense…capabilities we believe no chatbot has or ever can have.
It may be then that even while ChatGPT and other LLMs are having their moment in the limelight, the Natural Language Processing (NLP) market created by chatbots is already marked for disruption – by sapiens.
In the meantime, the hype surrounding LLMs has reached a fever pitch with ChatGPT. We are told it will make Google’s business model obsolete; it will replace Wikipedia, it will take the jobs of countless knowledge workers. Microsoft is contemplating investing an additional $10 billion in OpenAI, betting that it will give them an unassailable advantage in the big tech war for dominance.
Maybe it will, but AI hasn’t done any of these things yet. The field of AI is prone to such cycles of exuberance, followed by pessimism when the technology fails to deliver.
What exactly do we know ChatGPT is good for? It provides a stunning illusion of “knowing what it is talking about” and illusions are good for cheating and deception. On the same day the Microsoft investment was announced, the New York City public school system banned students and teachers from using ChatGPT[iv].
Even while the media is churning out article after article proclaiming that ChatGPT will change our world, others are vigorously pointing out its capacity for abuse, citing plagiarism, spreading misinformation, and reflection of the biases and prejudices of the humans who wrote the content on which it was trained.
If we strip away the hype, mystery, and allure surrounding the term AI in general, can we find an honest, useful application for this interesting technology? I think the answer is, Yes.
Imagine you are a writer inspired to create on some topic or theme and are staring at a blank piece of paper. A common starting point is to ask yourself, “What have others said about this?”
If you want to say something new and original (as good writers do), it is important to know what others have said on the topic to build on and go beyond. So instead of Googling individuals in the field or resorting to Wikipedia, give ChatGPT a sample of your thoughts and see what it returns.
But keep in mind, what you get back will be a statistical aggregate of what most people have said. The really interesting and creative thoughts of the people out there on the ends of the statistical bell-curve will most likely be filtered out.
Still, once you read and think about what the generated words mean to you, you may find value there. Possibly the generated words express what you wanted to say, but with better vocabulary and grammar than you could manage. Even better, autocomplete on steroids.
LLMs will find their place but generating text that is an illusion of language is not going to transform the world. Or if it does, the impact will be dystopian.
Some LLMs can generate images. Here too we see similar hype and similar uses and abuses. Art too is a form of communication, though its matter is of the heart, sentience, rather than of the intellect, sapience. The images generated should not be considered art in themselves, but they certainly can be a valid component of the work of a human artist who causes them to be generated and can honestly say, “This image expresses what I feel.”
Like other AI technologies before it, the development of LLMs has consumed vast resources and investment without getting us any closer to the real goal of AI: intelligent machines that comprehend reality as well or better than humans do.
At New Sapience we are giving machines something to think about, curating human knowledge to serve as the common educational core for all sapiens.
It only takes a small team when to teach one student is to teach them all, especially when that student has perfect memory. Just as with human education, the more you know, the easier it is to learn.
For the first time, there is an AI technology that scales.
[i] The Wall Street Journal’s Joanna Stern recently used ChatGPT to generate an essay which she then submitted as an assignment for a 12th-grade AP Literature class. The essay was given a passing grade (lowest 30% of the class).
I read the essay and thought the passing grade a sad commentary on the low standards of our educational system for a human, but for a mindless algorithm it was amazing.
[iii] This fact is almost universally recognized by the AI community. There are a handful of exceptions that have claimed some LLMs have sentience.