Xiao Zha and Musk declare war on ChatGPT! Meta and Twitter form a top AI team, causing smoke everywhere in Silicon Valley

The popularity of ChatGPT has directly changed the pattern of the entire Silicon Valley giant.

Microsoft's step-by-step moves have forced Google to step out of its complacency and urgently release the AI ​​chatbot Bard.

Naturally, the other forces in the major Silicon Valley companies will not stand idly by.

Recently, Musk and Zuckerberg have also stepped down and officially declared war on ChatGPT!

According to The Information, Musk has been in contact with AI researchers in recent weeks, hoping to set up a new research laboratory to develop an AI chatbot that directly interacts with OpenAI ChatGPT sparring.

At the same time, Xiao Zha also recently announced on his own platform that Meta will form a top AI team to focus on research and development. Generative artificial intelligence products. The new team is led by Ahmad Al-Dahle, head of artificial intelligence, reporting to current chief product manager Chris Cox.

Meta welcomes the battle of ChatGPT

Since its high-profile entry into the metaverse, in the eyes of outsiders, technology giant Meta has been a bit slow in the chatbot trend.

In fact, it’s not the case, but I haven’t defeated it yet.

As early as June last year, Meta open sourced its self-developed large-scale language model OPT-66B, and released the OPT-based chat robot BlenderBot3 as early as August.

However, unlike ChatGPT, which became an instant hit and had more than 1 million users within 5 days of its launch, Meta’s Blenderbot feels boring.

And soon, this chatbot was riddled with problems. It would not only criticize its boss, but also publish vicious remarks and false information, leading directly to a disaster.

In November, the "scientific research aid tool" Galactica came out.

The paper revealed that it was trained on 48 million papers, textbooks and lecture notes, millions of compounds and proteins, scientific websites, and encyclopedias with a total of 120 billion parameters. .

However, Galactica, which was supposed to be strong, not only summarized scientific research very inaccurately but was also biased. As a result, it was hastily removed from the shelves only three days after it went online.

Of course, the successive failures did not make Meta show weakness.

In February this year, Meta released a new 65 billion parameter large language model LLaMA. Not only is it open source, but its performance on most tasks is better than GPT-3.

Paper address: https://research.facebook.com/publications/llama-open-and -efficient-foundation-language-models/

What makes it unique is that it can be used by researchers, is small in size, and has lower cost of use. Meta hopes that researchers can use it to solve some of the problems that have been plaguing large language models.

##Because of the fiasco of his own Galactica, Yann LeCun was somewhat sour in his words about the popular ChatGPT:

"It does not scale and will never be the right path to strong artificial intelligence."

"As far as the underlying technology is concerned, ChatGPT is not a great innovation."

However, when the LLaMA model came out, he also showed it off. .

Xiao Zha said that there is still a lot of basic work to be done before realizing a true future experience, and he is excited about all the new things Meta will create in the process.

In other words, Meta has been doing a lot of research in this area, but has been more cautious than Microsoft and Google in applying these technologies to products.

Now, the establishment of Meta’s new team will allow the language and image models created by researchers to be presented in their own products faster in the future.

After poaching the DeepMind boss, Musk will also make ChatGPT

Now, major companies around the world are making their own ChatGPT, and Academician Ma cannot be absent. .

In recent months, Musk has criticized OpenAI for taking too many steps to make ChatGPT go around in circles to avoid offending users.

Last year, Musk said that OpenAI’s technology is “training AI to make it alert.”

Obviously, compared with ChatGPT and Microsoft’s recent Bing AI, the competing chat AI that Academician Ma wants to launch will not go around in circles on controversial topics.

To this end, Musk hired Igor Babuschkin, a researcher in the DeepMind AI department.

# In "StarCraft 2", he defeated the top professional players with a score of 5 to 0. Babuschkin is one of the main authors of the paper.

Paper address: https://www.nature.com/articles/s41586-019-1724 -z

In 2021, DeepMind released the pre-trained language model Gopher with 280 billion parameters, with the number of parameters reaching 280 billion, compared to the 150 billion parameters of OpenAI GPT-3 Almost double. Babuschkin, also a co-author of the paper.

The evaluation results show that Gopher's performance exceeded SOTA in about 81% of the tasks' benchmark performance analysis of 152 tasks, especially in problems that require a large amount of knowledge to solve. Things like fact-checking and common sense work particularly well.

##Paper address: https://arxiv.org/abs/2112.11446###### #########In an interview, Babuschkin said that Musk’s goal is far beyond just building a chat AI with fewer restrictions on output content. ############ "Our goal is to improve the reasoning capabilities and factuality of these language models, making the model feedback more trustworthy." ############Babuschkin Said that he and Musk are discussing forming an AI research team, but the project is still in its early stages and there are no plans to develop specific products. ############Babuschkin left DeepMind last week and has not officially signed an employment agreement with Musk. He said: "I am very much looking forward to working with Elon in the field of large language models." ############When developing ChatGPT, OpenAI used "Human Feedback Reinforcement Learning" (RLHF), Musk There has been criticism of this. ######

Because RLHF uses human input to adjust the model so that the model’s feedback is more ideal to avoid racism, prejudice, and hateful language. Musk criticized: This model actually reflects the bias of developers.

"We need a GPT that tells the truth"

More than ten days ago, Musk responded with a screenshot, and the content of the screenshot was Microsoft's The chatbot refused to tell a joke in the style of a comedian because it believed some of his words would hurt certain groups.

Bing AI says: Humor should be funny and inclusive, not hurtful and divisive

Musk said on Twitter, "We need TruthGPT."

Obviously, after acquiring Twitter, Musk has implemented this philosophy.

After acquiring Twitter in October last year, he deliberately read the company’s internal communications materials and disclosed them through several reporters Twitter's content review has been compromised - deleting content, blocking users, and limiting accounts, but the account users are unaware of it.

Even before acquiring Twitter, Musk had become a well-known cultural warrior.

So, where will the “ChatGPT” that Musk wants to implement be implemented?

People familiar with the matter said that this new project may be applied to Twitter, or it may become an independent artificial intelligence laboratory.

OpenAI was originally mine

In fact, Musk has been holding back for a long time. Since ChatGPT became popular, he has been feeling very sad.

You must know that OpenAI was originally created by him and his good friend Sam Altman in 2015. The founding purpose is Create "AI that is beneficial to mankind."

Part of the reason is that "Google did not pay enough attention to artificial intelligence safety."

The result developed later, because of the There were differences in the company's development direction, and Musk angrily quit the OpenAI Group's board of directors in 2019. The following year, OpenAI transformed from a non-profit to a for-profit private entity and "sold" to Microsoft for US$1 billion. This decision angered Musk even more.

Then, at the end of last year, ChatGPT suddenly became popular all over the world.

Musk did not hesitate to praise him at the time. While he praised ChatGPT as "terribly good", he reiterated his point of view - we are really not far away from dangerously powerful AI. .

As ChatGPT becomes more and more popular, Musk feels increasingly unhappy.

On February 17, Musk said angrily: "OpenAI was originally designed as an open source (that's why I named it "Open" AI), non-profit The company was created to act as a counterweight to Google, but now it has become a closed-source, profit-maximizing company controlled entirely by Microsoft. This is not what I want at all! 》

Just one day ago, Bing AI said, "I am perfect and never make mistakes. The mistakes are those external factors, such as network problems, server errors, user input, and errors in search results." But I’m perfect."

Musk reposted this post, saying that he was so scared that Bing was simply Just like the crazy AI murderer in the science fiction game "System Shock" 30 years ago.

# Bing is too unsafe.

One day later, his "curse" was indeed revealed. Microsoft immediately performed a "lobotomy" on Bing AI, and users complained.

Now when it comes to OpenAI, Musk has this tone: "Originally it was created as an open source non-profit organization. Now it closes the source code for profit. I don't I publicly hold shares in OpenAI, and I am not a board member, nor do I control it in any way."

AI gives me existential angst

Recently, Musk tweeted that he has an existential anxiety about AI.

Not long ago, Musk also told the participants at the World Government Summit held in Dubai, United Arab Emirates : "Artificial intelligence is one of the biggest risks facing civilization in the future."

"Frankly, I think we need to regulate AI safety. I think that's more important than cars, airplanes, or medicines The risk to society is greater."

In December last year, Musk said on Twitter: "I have been calling for safety regulation of AI for more than ten years! 》############################## Faced with netizens’ doubts about Neuralink, Musk said that Neuralink has supervision, which is far less than AI. So dangerous############### As early as eight years ago, Musk said that AI may be more dangerous than nuclear weapons. ########################### Obviously, 2023 will be a year of frequent events in the AI ​​industry. ###########################So while AI is dangerous, he "considering all the things about AGI's existential angst, I'd rather I don’t want to live in the past to witness AGI now.” ########################### Obviously, in Academician Ma’s view, creating his own ChatGPT is the best thing. ######

The above is the detailed content of Xiao Zha and Musk declare war on ChatGPT! Meta and Twitter form a top AI team, causing smoke everywhere in Silicon Valley. For more information, please follow other related articles on the PHP Chinese website!