Chatting is an art, but also a skill. Sometimes, through high-efficiency and skillful chat, there are often unexpected gains, which people affectionately call "white whoring". That's right, as long as the skill is deep, the quality of "white prostitution" is better. Nothing in the world will go beyond communication and reach the result directly. This is why scientists of artificial intelligence solutions are keen to solve the current problems of users in the form of dialogue.
Google released a new model LaMDA at this year's I/O conference. Its function is roughly the same as the logic of today's task-based AI, but it is also an important node in the development of AI. Google and Alphabet CEO Sundar Pichai's view of LaMDA is: "It aims to have a conversation with any subject."
Entering the 21st century, artificial intelligence has crossed the stage of answering non-questioning questions. What is even more outrageous is that different consumer-grade devices are equipped with AI that will not only wait for questions like chicken blood, but will also wait for questions to appear suddenly, giving users a surprise. . To put it simply, most of today's AI still lacks the injection of "soul", but is excited to search for the answers users want to know.
LaMDA is different. If it is put on the market, we may imagine it as when discussing a question with AI, skipping the current question according to the next question, and accurately giving the answer to the next question, and will not "What's the date today?", "How to fish?" and so on, the basic logic does not fit the question. Although Google’s model is not much different from the market’s products, Google has injected a bit of soul into AI, pushing the industry a small step forward.
This is because the training model has a neural network architecture (similar to language models such as BERT and GPT-3), and during the training, it learned the nuances of how to distinguish open dialogue from other forms of language. Pichai said that LaMDA has integrated these concepts from its training data, so it can respond wisely and keep the conversation open. "There is no dialogue going the same way twice, and the model can conduct dialogue without retraining." Pichai said.
Although the emergence of LaMDA makes it no longer difficult for people to talk with robots and will not enter a dead end, it is a manifestation of AI technology progress. But Google itself is very concerned about the principle of AI, that is, AI may be the same as all tools, or it may be abused. In the Google blog "LaMDA: Our Breakthrough Conversation Technology", it is clearly pointed out that through internalizing bias, it may cause hate speech or copy misleading information. Even if the language trained by censorship is still used, it still exists. The possibility of abuse.
LaMDA is the so-called Large Language Model (LLM), which is a deep learning algorithm trained on large amounts of text data. But because of this, experts believe that models of racism, gender perspectives, and insulting ideas may be embedded. And because of the dissemination, automatic learning, introduction, etc., it is easy to make people mistakenly think that the output of opinions is artificial.
Fortunately, Google has made a lot of efforts to ensure LaMDA's high standards of fairness, accuracy and privacy. "When creating technologies such as LaMDA, our first priority is to ensure that we minimize such risks. We have been researching and developing these technologies for many years, so the issues involved in machine learning models (e.g. unfairness The biases) are very familiar.”
At the same time, more than 500 researchers around the world are racing to understand the functions and limitations of these models. Under the BigScience project led by Huggingface, a startup that uses an "open science" approach to understand natural language processing (NLP), they seek to build an open source LLM that will become a shared resource for the scientific community. The goal is to obtain as many scholarships as possible in a single key year. Their central question is: how should LLM be developed and when should LLM be deployed to reap its benefits without causing harmful consequences?
Thomas Wolf, chief scientific officer of Huggingface, said: "We can't really stop this madness of large language models that everyone wants to train." "But what we can do is try to nudge this in a more favorable direction in the end."
Write at the end
If Google will integrate LaMDA into existing products, these products may become more useful to users. For example, LaMDA can solve the interaction of children and the elderly in the form of conversation; in the built-in car, it is more convenient to change the destination or query the road conditions; in addition, it will also understand the meaning of different words And applications, etc. This is a change in To C users' search behavior, and of course it may also be a direction for To B companies to adjust content and products.
But this is not the final form, and Google has left more room for imagination in the future. "Sensitivity and specificity are not the only qualities we pursue in models such as LaMDA. We are still evaluating dimensions such as'interest' by evaluating whether the response is insightful, unexpected or witty."
