Hastily removed from the shelves, the boss is disdainful, why does Facebook avoid the ChatGPT battle?

​Author | Xu Jiecheng

Since the release of ChatGPT in December last year, topics about large language models (LLM) and chatbots have almost dominated the entire Internet. Major technology giants soon realized the unlimited business opportunities it could bring.

Microsoft was the first to take action by investing an additional US$10 billion to integrate ChatGPT into its own search engine Bing; Google, which was slightly more cautious, launched its self-developed chat robot Bard after spending a certain amount of time and energy. Not to be outdone, domestic technology giant Baidu also recently announced that it will enter the melee with its chatbot Wenxinyiyan in March this year.

But a technology giant that has always been aggressive in the past seems to have given up on this "ChatGPT battle". This company is Meta led by Zuckerberg. Surprisingly, Meta may be the only tech giant that hasn’t jumped on the chatbot bandwagon yet.

1. Open high and go low, fight and lose repeatedly

Careful study of Meta performance The reason for the abnormality is most likely due to many failed attempts in the past - in fact, as early as June 2022, Meta open sourced its self-developed large-scale language model OPT-66B, and released a new language model based on OPT-66B in August of the same year. The chatbot BlenderBot3 is 3 months earlier than ChatGPT. It can be said that Meta is one of the first companies to get involved in LLM chatbots.

Although BlenderBot3 was only released in the United States at that time, the sensation it caused at the time was no less than that of ChatGPT today. Just hours after its release, Twitter and Reddit are already filled with screenshots of people having interesting conversations with BlenderBot3.

However, soon, this "successful" chat robot headed for a disaster. A large number of users found that BlenderBot3 would publish vicious remarks and false information, and even questioned Zuckerberg's business strategy, calling it "unethical." This caused a large number of users to gradually lose trust in BlenderBot3. In the end, Meta watched helplessly as the "big baby" it spent a lot of money to build gradually declined.

Of course, one failure did not extinguish Meta’s enthusiasm for LLM. After regrouping, Meta teamed up with Papers with Code in November 2022 to release another robot, Galactica, based on a large language model. Compared with the previously failed BlenderBot3, Galactica has a more specific application field - ghostwriting papers.

According to the official introduction, Galactica is trained from 48 million papers, textbooks and other materials, whether it is ghostwriting paper abstracts, introductions, formulas, or even references. Not only that, in addition to text generation, Galactica can also perform multi-modal tasks involving chemical formulas and protein sequences.

But this time, Meta still failed to solve the problem of LLM generation accuracy. Although Galactica’s book strength seems to be very strong, there are a lot of errors and even It's fake content. In order to prevent the impact from spreading further, Meta had to hurriedly remove Galactica from the shelves only three days after its release.

2. The helmsman’s belief is shaken

Successive failures seem to have shaken the belief of Meta AI helmsman Yann LeCun, Turing Award winner and Meta’s chief AI scientist, in LLM There was some vacillation. The recent news of ChatGPT and Google Bard errors seems to have given LeCun some support.

Whether it is to protect his own face, or he really realized the fatal flaw of the LLM robot from two failures, today LeCun has changed from the original LLM supporter For the LLM bashers.

As the popularity of related topics continues to increase, LeCun has also begun to actively expose the shortcomings of large language models and chat robots through various channels. In a recent online discussion organized by Collective Forecast, LeCun said that although they are revolutionary in the public eye, in terms of the underlying technology, today's chatbots are not that great of an innovation.

In addition, LeCun has expressed his disdain for ChatGPT on Twitter many times: it cannot scale and will never be the right path to strong artificial intelligence. LLM that scales up auto-regression simply cannot bring chatbots to the level of human intelligence. I don't think ChatGPT does more right than correcting grammar, completing sentences, or summarizing articles.

LeCun believes that small companies like OpenAI have nothing to lose, and they can certainly use immature technologies and products to create hype for themselves. But it is obviously very unwise for large companies to choose to wade into this muddy water, especially after everyone has seen the failed attempt that cost Google $100 billion.

3. Meta’s layout and trade-offs

In fact, Meta’s investment in the field of artificial intelligence has always been at the forefront of major technology companies for a long time. At the forefront, most of the innovations of the Meta artificial intelligence team have entered their advertising business, and until now, they are still working hard to transform self-developed artificial intelligence models and algorithms into products that can bring revenue.

LeCun pointed out that Meta has long been criticized for spreading false information due to mistakes by BlenderBot and Galactica. Today, Meta hopes to more strictly control the tools and content they publish, instead of blindly using chatbots to sneak into the current "artificial intelligence craze" and make the same mistakes again.

Whether it is due to the pain caused by previous failures or the change in the concept of the person at the helm, Meta is indeed avoiding this globally focused battle. In an interview about generative artificial intelligence, LeCun said: Zuckerberg’s long-term dream about the metaverse is still in progress, and he also agrees with the fact that generative artificial intelligence may be the best in the metaverse. realization. When it comes to large-scale language models and chatbots, they now seem more willing to sit on the mountain and watch the tigers fight from a distance, actively looking for the mistakes that Google, Microsoft or OpenAI are making, and learn enough experience from them.

Reference link:

https://analyticsindiamag.com/meet-the-ai-genius-who-is-obsessed-with-llms/

http://www.myzaker.com/article/63e3902e8e9f094fe76b7af7/

https://analyticsindiamag.com/why-meta-took-down-its-hallucinating-ai-model-galactica/

The above is the detailed content of Hastily removed from the shelves, the boss is disdainful, why does Facebook avoid the ChatGPT battle?. For more information, please follow other related articles on the PHP Chinese website!