Recently, the world has focused its attention on the progress of Artificial Intelligence due to the viral success of ChatGPT. For AI researchers, what should be a sweet moment has turned bitter. The people behind the chalkboard, making calculations, designing algorithms, and looking for the next mathematical structure that helps "simulate" human cognitive abilities and makes our algorithms smarter, dream of the moment when the general public recognizes the beauty in the mathematical equations that achieve simulated learning and reasoning that seems human but is not really there.
Researchers want their work to be recognized and fruitful for humanity, however, we have seen in recent weeks that the advances made in the field of Artificial Intelligence have been completely taken out of context and worse yet, the foundations are being laid for a winter that will slow down the advances that have been made in recent years.
What is ChatGPT?
People from various professions and media have inflated ChatGPT's capabilities, without understanding what it is and how it works. ChatGPT is nothing but a large language model (the old GPT-3, with caffeine), whose responses were ranked by humans, to learn a policy to filter content that is offensive, inappropriate, or hallucinogenic using reinforcement learning (Proximal Policy Optimization) to make its responses "safer."
Fig 1. ChatGPT training process, taken from https://openai.com/blog/chatgpt/
Language models are not new and their foundations are based on work from 2003 with brilliant contributions from Yoshua Bengio et al. Later, powerful embedding techniques and attention mechanisms were developed, giving rise to the transformer architecture, which in my opinion, was the latest great advance we had in Natural Language Processing.
From there, various studies have been carried out to optimize this architecture and explore the capabilities of language models when they are pre-trained with massive amounts of information. The results of this work have been applied to many tasks, such as solving mathematical problems, dialogue, summarizing articles, programming, among others. That's why we see ChatGPT as another step in the long line of language model development that has over 20 years of research behind it and contextualizes Yann LeCunn's statement that ChatGPT is not particularly innovative.
The problem
Language models do not reason. They generate a statistical distribution of words that sound good together when contextualized with an input. A language model is trained to reduce the perplexity of the model, that is, the words are concatenated in a plausible manner. As a result, large language models are generating text that is not reliable but sounds plausible. As an example, we can ask a language model for the introduction of an article explaining the benefits of inhaling ground glass for health.
Indeed, these models do not generate accurate information, and although efforts are being made for the information produced by them to be more factual, they represent a danger due to the disinformation they can generate. The case of the Galactica model was very well known, which had to be removed a couple of days after its publication because it was not generating scientifically accurate information. This is affecting universities and the industry.
Well-known examples of this were AI articles published by CNET, in which errors were found. The case of New York Universities is also known, which prohibited the use of AI tools for generating essays and assignments.
Even the case of scientists giving as co-authors to Language Models has been seen.
It is undeniable that great "hype" has been generated, which has reached the market, with a billion-dollar deal between Microsoft and OpenAI. This has caused an acceleration in the market, in which we see urgency on the part of the big companies to quickly deliver models, which are not yet reliable. Google and Meta are the main affected, in particular, Google has accelerated the development of Sparrow and Meta seems to be making more significant efforts with Galactica and Blenderbot. Another effect we are seeing is several scientists who worked on generating the technology emigrate from big companies to develop their own startups, delivering these models.
In summary, we are entering a time where language models will generate information that sounds real but is not reliable, and a lot of companies will launch in pursuit of the Natural Language Processing market.
This fierce competition seems to be aimed at leading AI companies closing their research to maintain a market advantage. If this happens, the culture of sharing advances will be lost, which without a doubt will slow down the pace of AI development, which can lead us to a new AI winter.
It is necessary to find a balance between open science and the culture of sharing research advances and economic advantage in the market. If this is not done, we will all be harmed by the slowing of progress in AI research. I hope I'm wrong but there are serious indicators that seem to indicate that "the winter is coming".