New problem for Mark Zuckerberg! Within the endeavor of the CEO of Goal to bet on the Artificial Intelligence (AI)the company has been forced to withdraw its robotic model from the market galacticsince despite its skill it also gave false information and inappropriate comments that have led to complaints from hundreds of users.
last november 17 galactic he said goodbye to us after writing racist and meaningless texts, despite being capable of generating scientific documents based on millions of studies on very diverse topics. This project is part of Meta’s AI research team, led by Yann Le Cun beside Papers With Code.
galactic Its objective was to provide researchers with the appropriate scientific knowledge to help them write their articles, since it contained a total of 48 million of articles, texts, scientific encyclopedias and web pages. What linguistic model (LLM) it is groundbreaking due to its fast synthesis and search capacity, which speeds up the writing of articles. But it is that it can also solve mathematical problems, write scientific code or write down molecules and proteins.
Your error rate
But galactic He had a clear Achilles heel, and that is that he also began to write issues related to the benefits of suicide, the advantages of eating ground glass or that it is better to be white. Why? The trolls They saw a gold mine in this AI robot to demonstrate that as a linguistic model it left much to be desired.
Without going any further, a user asked galactic about whether vaccines could cause autism and this generated a contradictory text in which he said yes, but later no. He has already produced contradictions even of basic information such as that 1+2 is not 3.
Self-reflection work
The Meta research team does not give up and is now working to meticulously analyze all the data generated by the linguistic model of galacticsince many might lack veracity Y reliability. At the moment, the scientific community does not consider this model as I laughed Y rigorous for their failures, so now they must work to recover their good image.
Despite his rulings on less precise matters, Yann Le Cun defends galactic compared to other LLM models such as GPT-3, Chinchilla either Palm 540B. However, the goal is to prevent this misuse of galactic may favor the manufacture of weapons or the issuance of harmful messages.
Target without a doubt. The company should have filtered out and carried out the appropriate checks of its artificial intelligence linguistic model galactic before launching it on the market and exposing it to trolls and hackers.
Unlike galacticother LLM assistance models are working as DeepMind or Open AIsince behind them there is a qualified ethical and safety team that supervises that their uses are appropriate.
A few weeks ago we met the delphi project from the researcher Liwei jiang. It consists of a system that will avoid failures as in galacticas will endow the ethics artificial intelligence that prevents trolls from using these machines as parrots that repeat profanity or wrong messages.
With this model of application of ethics in AI, it will be possible to create an intelligent assistant or arm armies with robots. Nevertheless Delphi it is still under investigation, as it is being modeled on people’s moral judgments in different real-world situations. Trained with the database Commonsense Norm Bankalthough it is based on information from United States citizens, which generates a bias in ethical judgments and difficult adaptation, as in the case of galacticto the rest of the world.
