DeepMind restarts the 'Second Counterattack' and announces that it will challenge Lee Sedol!

​Author | Xu Jiecheng

Reviewer | Yun Zhao

When we mention artificial intelligence, DeepMind is a name that cannot be bypassed no matter what. As one of the most famous AI research companies in the world, it was the AlphaGo Go robot that defeated Lee Sedol in one fell swoop and created the most dazzling milestone in the development of artificial intelligence. Now, this great artificial intelligence company seems to have become the last trump card of its owner Google in the counterattack of chatbots.

Recently, Behnam, co-leader of Google’s Blueshift large model team, announced that the Blueshift team will cooperate with DeepMind to further enhance both parties’ large-scale language model research and development capabilities. The merged team will be led by DeepMind chief scientist Oriol Vinyals, who is the main leader of many top projects such as AlphaCode, AlphaFold and Google Research.

1. This move may be due to the “disharmony” between Google and DeepMind

Although this cooperation between Google and DeepMind seems logical on the surface, in fact Since its acquisition in 2014, the relationship between DeepMind and Google has been difficult to say harmonious.

DeepMind CEO and co-founder Demis Hassabis

For a long time, DeepMind internally believed that Companies need to have the same institutional structure as non-profit organizations, and their own research and innovation in artificial intelligence should be used to benefit all mankind rather than be controlled by private companies. And some researchers at DeepMind also believe that they should be academics, and when dealing with Google, they are very disdainful of the latter's complicated bureaucracy. In this relationship, DeepMind is more like an arrogant and rebellious genius boy, while Google is more like an old father with deep pockets but unable to influence his children's thoughts.

And just in 2021, DeepMind had a "separation" with Google, although in the end Google relied on its strong financial resources and control to retain the wannabe An independent "dream team", but since then, the relationship between DeepMind and Google has become increasingly awkward, with some DeepMind employees even using encrypted messages to communicate in order to avoid Google's monitoring. Now it seems that if Google chooses to activate DeepMind at such a time point, it is most likely that it is ready to show the world its last and most powerful trump card.

2. Challenging ChatGPTD DeepMind’s killer tool will be released within the year

As Google’s ultimate choice, what kind of strength does DeepMind have in large-scale language model research? Can this cooperation between the two help Google win this turnaround?

In fact, as early as September last year, DeepMind published a paper about its chat robot Sparrow. According to reports, the basic model used by Sparrow has more than 70 billion parameters and adopts a reinforcement learning framework based on human feedback. And to control the accuracy and security of the content it generates, Sparrow is able to conduct Internet searches through Google and verify the content it generates when talking to users or answering questions.

In addition, in order to ensure the absolute safety of the model's behavior, DeepMind researchers also developed a set of proprietary dialogue rules for the model and used dialogue testing to induce Sparrow to violate these rules. After obtaining information data involving illegal conversations, the researchers used it to independently train a new control model to promptly limit Sparrow when it might make a mistake.

Sparrow dialogue processing model

According to the final test results, when asked factual questions At this time, the answers provided by Sparrow are very credible, and more than 78% of the generated content is supported by definite evidence. And this is also the most lacking capability in ChatGPT, Bard or New Bing.

Although Sparrow (70 billion parameters) cannot be compared with ChatGPT (175 billion parameters) in terms of model parameters, the important point is that Google currently also has a nowhere The large monster placed - Bard (137 billion parameters). If Google and DeepMind can integrate the capabilities of Bard and Sparrow well in this cooperation between the two parties, it will most likely subvert today's chatbot market. What is even more noteworthy is that not long ago, DeepMind founder and CEO Hassabis publicly stated that the internal beta version of Sparrow will be officially released in 2023. In this regard, some industry insiders predict that the Sparrow mentioned by Hassabis will most likely be the reskinned and optimized Bard.

3. The Chat war is getting more intense

Since the failure at the Bard conference, Google has been trying to find ways to save the situation. Now it seems that this cooperation with DeepMind should be a very right choice. But even so, time is running out for Google.

Today’s chat robot market has shown an obvious trend of competition among the best. Microsoft, which had relied on New Bing to gain a first-mover advantage, soon thereafter limited the number of daily conversations it could have with users due to problems with New Bing generating malicious content. However, recent news shows that with the support of OpenAI, Microsoft has begun to solve this problem and will relax or even lift New Bing's conversation restrictions in the short term.

Just in the past few days, Meta Company has also announced a new large-scale language model called LLaMA-13B, and claimed that it is only one-tenth of the scale of ChatGPT. In this case, it has stronger performance than GPT-3 on most benchmarks. In addition, according to people familiar with the matter, Silicon Valley Iron Man Musk is also preparing for a new artificial intelligence research laboratory in the near future, hoping to develop a similar competitor to ChatGPT, and the first researcher of the laboratory has worked at OpenAI. With DeepMind.

Although compared to these rising stars, Google currently has a huge advantage in model and product maturity, and the addition of DeepMind is bound to bring it many crucial help. But in a dangerous situation where there are interceptions in front and pursuers in front, can Google rely on the ace card of DeepMind to break out of the siege and return to the throne this time? Regarding this question, we cannot see a clear answer yet.

Reference link:

https://analyticsindiamag.com/google-does-not-deserve-deepmind/

https://www .deepmind.com/blog/building-safer-dialogue-agents

https://m.thepaper.cn/baijiahao_14479005

https://openreview.net/pdf?id=YoS -abmWjJc

The above is the detailed content of DeepMind restarts the 'Second Counterattack' and announces that it will challenge Lee Sedol!. For more information, please follow other related articles on the PHP Chinese website!