Almost a year ago, an artificial intelligence robot called BB1.0 (BlenderBot 1.0, hereinafter referred to as BB1.0) was created by Facebook.
At that time, Facebook was full of pride and praised this new robot as the best in the industry, and the AI industry was also full of praise. But the test results within one year made this BB1.0 and the GPT-3 of the same period wear the shining crown of "artificial mental retardation".
After a lapse of one year, Facebook once again strongly promoted BB2.0. As a sequel to the lessons learned, is it a masterpiece that swept away the haze? Or old wine in a new bottle?
First of all, why are they both artificially retarded?
One BB1.0, one GPT-3; the one in front doesn't know that Trump, the top global leader, is no longer the president of the United States, and the one in the back often advises patients to commit suicide "once and for all", two silly "machines", very happy together.
OpenAI's GPT-3 and BB1.0 have a powerful language system - which allows them to express themselves clearly, explain the principles of the East and the West, and the philosophy of the North and the South. But in the Facebook test records, the researchers were extremely troubled by the need to consider the construction of short-term and long-term memory every time they communicated with them. Due to the lack of effective recording nerves, what you tell GPT-3 or BB1.0 on August 3, 2021, they will be forgotten on August 4, 2021.
The barrel effect tells us that it is often the shortest piece of wood that determines the quality of a product.
Due to the flaws of the algorithm and the incorrect parsing, GPT-3 and BB1.0 are often amazing.
GPT-3 advises patients to commit suicide
Delivering completely inaccurate information through assertive statements is what these two little guys do well.
Due to the inability to obtain additional knowledge reserves, it is still difficult for GPT-3 and BB1.0 to communicate effectively with humans in real time. If you ask them how far the latest hot spot artist Wu has developed, it will only give you a list of EXO's It was reported in the early years; you asked him whether the flood in Zhengzhou was effectively controlled, and it only revealed what dishes were in the food exhibition held in Zhengzhou a few years ago.
Of course, most people in foreign countries inquire about news about football stars and social events.
Likewise, GPT-3 and BB1.0 are very stable answers to a lot of old stories that have changed over time.
GPT-3 and BB1.0
Completely unaware of a football star's 2021 transfer news
When Zhang Guan and Li Dai made more and more speeches, the world began to believe that these two silly "machines" still had many flaws, so they were given the nickname of "artificial mental retardation".
The evolution of "artificial mental retardation" - artificial intelligence?
BB1.0 has received resources and long-term research and development in Facebook's policy of not abandoning and not giving up.
BB1.0 has naturally evolved successfully - Facebook announced the sequel to BB1.0 in July: BlenderBot 2.0.
Facebook, a big company, has a publicity team.
High technology, not bad money, all new breakthroughs.
It once attracted the attention of the industry.
keep me healthy, keep me healthy
Its product introduction is - artificial intelligence products that understand consumer behavior through big data and provide personalized recommendations (very long).
BB2.0 can query the Internet for movies, TV shows and other related information through any search engine, and read and write it to its long-term local memory storage. And through continuous learning , remember the previously discussed and entered information.
Continuing Learning Continual AI, a group of artificial intelligence and neuroscience researchers, started research and solved one of the biggest challenges in artificial intelligence. Even though the most advanced artificial intelligence systems were very powerful, they were still subject to statistical limitations. The algorithms were only trained once on the data set and were rarely retrained, which made it difficult for them to learn and acquire new information. Root cause of severe hysteresis.
Building continuous learning technology BB2.0 communication is smoother. When you talk about Marvel movies in one chat, BB2.0 will take the initiative to mention the main characters of movies such as Iron Man in the next conversation.
"Because it knows it's about what you want to talk about."
And the more intelligent BB2.0 after "evolution" is also further in the ability of deep conversation.
BB2.0 uses a retrieval-augmented generative AI model, an approach that generates responses and integrates more extended knowledge beyond that contained in the conversation.
Illinois Institute of Technology, a famous American school , will recruit students in the spring of 2022. The exclusive experience camp in China is here. Let’s experience the benefits of an online master’s degree in computer science! Two weeks for a limited time, limited to 30 people! The whole process is free , click the link below to lock the experience camp qualification!
During a human conversation with BB2.0, this AI model combines an information retrieval component with a text generator to find closely related data in its long-term memory and by searching internet documents.
After BB2.0 utilizes the neural network module to give the context of the dialogue context, the chatbot adds this part of the retrieved data to the dialogue history, and it is up to the algorithm to decide when to consider referencing this knowledge.
In this way, it appears that BB2.0 is omniscient and omniscient; It is more inclined to the level of "artificial intelligence" that the general public thinks.
Security Plus Code
In terms of security, Facebook has also improved and upgraded BB2.0.
In the past, artificial intelligence models always made the same mistakes. Not only BB1.0 and GPT-3, but most AIs will be affected by big data and have some unfriendly and incorrect tendencies when collecting massive data sets.
For example putting some "kitchen" near "feminine pronoun" and "Islam" near "terrorism"; these are worldly prejudices and a serious departure from our values.
A Stanford University research paper shows that AI models also associate "Jew" with "money" and "chimpanzee" with "black"; this is unacceptable racial discrimination.
In order to contain and avoid this situation, Facebook has imposed a safety formula on BB2.0.
This security formula uses automatic classifier measurement to detect key words and sensitive text information, which reduces the possibility of chatbots making "inappropriate" responses by 90% . In avoiding public relations crises that do not conform to values, BB2 .0 is 74.5% more secure .
"The way BlenderBot 2.0 uses the internet and long-term memory to respond to conversations opens up a new way of doing things that will better address security concerns," Facebook research scientist Jason Weston and research engineer Kurt Shuster wrote in a blog post.
write at the end
Let's go back to the question mentioned at the beginning:
BlenderBot 2.0, as a continuation of the previous experience, is a masterpiece that swept away the haze? Or old wine in a new bottle?
What we can be sure of is that BB2.0 adopts continuous learning and breaks through the shortcomings of previous AI models, which is a definite improvement. In the follow-up session analysis strengthening, as well as the increase in security measures, is also an undeniable increase.
Everyone's measurement standards for new AI products are different, and it is a kind of evaluation standard for a breakthrough or a small breakthrough. But as Jason and Kurt said, every little progress we make is moving towards the ultimate direction of artificial intelligence, so that agents can finally understand communication and dialogue as well as humans.
Therefore, BlenderBot 2.0 may also be a real improvement. What is your opinion?