Large Language Models (LLMs): an Ontological Leap beyond AI

More than the quasi-human interaction and the practically infinite use cases that could be covered with it, OpenAI’s ChatGPT has provided an ontological jolt of a depth that transcends the realm of AI itself.

Large language models (LLMs), such as GPT-3, YUAN 1.0, BERT, LaMDA, Wordcraft, HyperCLOVA, Megatron-Turing Natural Language Generation, or PanGu-Alpha represent a major advance in artificial intelligence and, in particular, toward the goal of human-like artificial general intelligence. LLMs have been called foundational models; i.e., the infrastructure that made LLMs possible –the combination of enormously large data sets, pre-trained transformer models, and the requirement of significant computing power– is likely to be the basis for the first general purpose AI technologies.

In May 2020, OpenAI released GPT-3 (Generative Pre-trained Transformer 3), an artificial intelligence system based on deep learning techniques that can generate text. This analysis is done by a neural network, each layer of which analyzes a different aspect of the samples it is provided with; e.g., meanings of words, relations of words, sentence structures, and so on. It assigns arbitrary numerical values to words and then, after analyzing large amounts of texts, calculates the likelihood that one particular word will follow another. Amongst other tasks, GPT-3 can write short stories, novels, reportages, scientific papers, code, and mathematical formulas. It can write in different styles and imitate the style of the text prompt. It can also answer content-based questions; i.e., it learns the content of texts and can articulate this content. And it can grant as well concise summaries of lengthy passages.

OpenAI and the likes endow machines with a structuralist equipment: a formal logical analysis of language as a system in order to let machines participate in language. GPT-3 and other transformer-based language models stand in direct continuity with the linguist Saussure’s work: language comes into view as a logical system to which the speaker is merely incidental. These LLMs give rise to a new concept of language, implicit in which is a new understanding of human and machine. OpenAI, Google, Facebook, or Microsoft effectively are indeed catalyzers, which are triggering a disruption in the old concepts we have been living by so far: a machine with linguistic capabilities is simply a revolution.

Nonetheless, critiques have appeared as well against LLMs. The usual one is that no matter how good they may appear to be at using words, they do not have true language; based on the primeval seminal trailblazing work from the philologist Zipf, criticism have stated they are just technical systems made up of data, statistics, and predictions.

According to the linguist Emily Bender, “a language model is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot. Quite the opposite we, human beings, are intentional subjects who can make things into objects of thought by inventing and endowing meaning.“

Machine learning engineers in companies like OpenAI, Google, Facebook, or Microsoft have experimentally established a concept of language at the center of which does not need to be the human. According to this new concept, language is a system organized by an internal combinatorial logic that is independent from whomever speaks (human or machine). They have undermined one of the most deeply rooted axioms in Western philosophy: humans have what animals and machines do not have, language and logos.

Some data: monthly, on average, humans publish about seventy million posts on the content management platform WordPress. Humans produce about fifty-six billion words a month, or 1.8 billion words a day on this content management platform. GPT-3 -before its scintillating launch- was producing around 4.5 billion words a day, more than twice what humans on WordPress were doing collectively. And that is just GPT-3; there are other LLMs. We are exposed to a flood of non-human words. What will it mean to be surrounded by a multitude of non-human forms of intelligence? How can we relate to these astonishingly powerful content-generator LLMs? Do machines require semantics or even a will to communicate with us?

These are philosophical questions that cannot be just solved with an engineering approach. The scope is much wider and the stakes are extremely high. LLMs can, as well as master and learn our human languages, make us reflect and question ourselves about the nature of language, knowledge, and intelligence. Large language models illustrate, for the first time in the history of AI, that language understanding can be decoupled from all the sensorial and emotional features we, human beings, share with each other. Gradually, it seems we are entering eventually a new epoch in AI.

Humanos y máquinas, un futuro con inteligencia artificial. AI Event at Fundación Telefónica

Last Thursday, April 7th, I was invited as CEO of hAItta to participate and give a speech about Natural Language Processing: the pending challenge in AI, at the event “Humanos y máquinas, un futuro con inteligencia artificial” in Fundación Telefónica.

It was a great pleasure for me to share the floor with Mr. Ramón López de Mántaras. director y profesor de investigación del Instituto de Investigación en Inteligencia Artificial del CSIC, Mr. Álvaro Otero, COO de BigML, Mr. Juan García Braschi, country manager España y CFO en Cabify, and Mr. Javier Placer Mendoza, CEO de Telefónica Open Future.

Bringing AI closer to society.

Full video of the event “Humanos y máquinas, un futuro con inteligencia artificial. Mar de datos 2016. Fundación Telefónica”

Summary video of the event and later interview

Winograd Schema Challenge: A Step beyond the Turing Test

The well-known Turing test was first proposed by Alan Turing (1950) as a practical way to defuse what seemed to him to be a pointless argument about whether or not machines could think. He put forward that, instead of formulating such a vague question, we should ask whether a machine would be capable of producing behavior that we would say required thought in people. The sort of behavior he had in mind was participating in a natural conversation in English over a teletype in what he called the Imitation Game. The idea, roughly, was the following: if an interrogator was unable to tell after a long, free flowing and unrestricted conversation with a machine whether s/he was dealing with a person or a machine, then we should be prepared to say that the machine was thinking. The Turing test does have some troubling aspects though.

Leer más »

Once upon a Time in 1954…

Die Grenzen meiner Sprache bedeuten die Grenzen meiner Welt

(The limits of my language are the limits of my world)

Ludwig Wittgenstein, Tractatus Logico-Philosophicus

… In a cold day of January it took place in Washington DC the Georgetown-IBM experiment, the first and most influential demonstration of automatic translation performed throughout the history. Developed jointly by the University of Georgetown and IBM, the experiment implied the automatic translation of more than 60 sentences from Russian into English. The sentences were chosen precisely; there was no syntactic analysis, which could manage to identify the sentence structure. The approach was mainly lexicographic, based on dictionaries in which a certain word had a link to some particular rules.

That episode was a success. Story has it that the level of euphoria amongst the researchers was such that it was stated that within three or five years the problem of the automatic translation would be solved… That was more than 60 years ago and the language problem –the comprehension and generation of messages by the machine- is still pending. Probably this is the last frontier which separates the human intelligence from the artificial intelligence.

Leer más »

Érase una vez en 1954…

Die Grenzen meiner Sprache bedeuten die Grenzen meiner Welt

(Los límites de mi lenguaje son los límites de mi mundo)

Ludwig Wittgenstein, Tractatus logico-philosophicus

… En un frío día de enero tenía lugar en Washington D.C. el experimento Georgetown-IBM, la primera y más influyente demostración de traducción automatizada de la historia. Desarrollado conjuntamente por la Universidad de Georgetown e IBM, el experimento implicaba la traducción automatizada de más de 60 oraciones del ruso al inglés. Las oraciones se escogieron de manera precisa; no había ningún análisis sintáctico que pudiera llegar a detectar la estructura de la oración. El enfoque fue eminentemente lexicográfico, basado en diccionarios en los que una palabra determinada tenía una conexión con unas reglas específicas.

Aquello fue un éxito y, cuenta la historia que, tal fue la euforia entre los investigadores, que se llegó a afirmar que en un plazo de entre tres y cinco años el problema de la traducción automatizada quedaría resuelto… Eso fue hace algo más de 60 años y el problema del lenguaje, de la comprensión y generación de mensajes por parte de una máquina, aún está sin resolver. Probablemente es la última barrera que separa a la inteligencia humana de la inteligencia artificial.

Leer más »