Crush GPT-4! Google DeepMind CEO reveals: the next generation of large models will be integrated with AlphaGo

**Source:**Xinzhiyuan

**Guide: **Google DeepMind CEO Hassabis made a new revelation: the new Gemini model will be combined with AlphaGo and the large language model, and the cost is expected to be tens of millions of dollars, or even hundreds of millions.

Google, it’s really overwhelmed.

Is the legendary Gemini, which merges AlphaGo and GPT-4-like large models, finally coming?

One is the AI system that used reinforcement learning to defeat the human Go champion and created history. The other is the most powerful multi-modal large model that dominates almost all large-scale model lists. The combination of the two AIs is almost invincible up!

Google DeepMind CEO Hassabis recently said to foreign media Wired that Gemini is still under development and it will take a few months, while Google DeepMind is ready to spend tens of millions of dollars, or even hundreds of millions.

Previously, Sam Altman revealed that the cost of creating GPT-4 exceeded $100 million. Google DeepMind, of course, cannot lose.

Too long to read version

Gemini will combine AlphaGo with the language functions of large models such as GPT-4, and the system’s ability to solve problems and plan will be greatly enhanced.

• Gemini is a large language model, similar to GPT-4
• Estimated to cost tens to hundreds of millions of dollars, comparable to the cost of developing GPT-4
• In addition to AlphaGo, there will be other innovations

Gemini will integrate AlphaGO using reinforcement learning and tree search.

• Reinforcement learning allows AI to solve challenging puzzles by learning from trial and error
• Tree search method helps to explore and remember possible moves in the scene, such as in game scenes

DeepMind’s extensive experience in reinforcement learning will bring new features to Gemini.

• Other fields of technology (such as robotics and neuroscience) will also be integrated into Gemini

Next algorithm, to surpass ChatGPT

According to OpenAI CEO Sam Altman, GPT-5 is still a few days away from release, and training will not begin for at least six months. The release date of Gemini has not yet been determined, but it may be within a few months.

Gemini, which is still under development, is also a large language model for processing text, which is similar in nature to GPT-4.

But Google DeepMind CEO Demis Hassabis said that Gemini will incorporate the technology used in AlphaGo, which will give the system new planning and problem-solving capabilities.

In 2016, the scene where AlphaGo defeated the Go world champion Lee Sedol is still vivid.

Hassabis said, “It can be said that Gemini combines some of the advantages of the AlphaGo system with the amazing language capabilities of the large language model. And, we have some other interesting innovations.”

Gemini is said to have multimodal capabilities not found in previous models and is very efficient at integrating tools and APIs. Moreover, Gemini will be available in a variety of sizes, designed to support future innovations in memory and planning.

In March, it was said that Gemini would have a trillion parameters like GPT-4. Moreover, it is said that Gemini will use tens of thousands of Google TPU AI chips for training.

At the Google Developers I/O conference last month, Google mentioned that from the very beginning, Gemini’s goal was multi-modal, efficient integration tools, and APIs.

At the time, Google’s announcement was: “Although it is still early days, we have already seen in Gemini a multimodal capability that has never been seen in previous models, which is very impressive.”

The technology behind AlphaGo is reinforcement learning, a technology pioneered by DeepMind.

RL agents interact with the environment over time, learning policies through trial and error that maximize long-term cumulative rewards

Through reinforcement learning, AI is able to adjust its performance through trial and error and receiving feedback, thus learning to deal with very difficult problems, such as choosing how to take the next move in Go or video games.

In addition, AlphaGo also uses the Monte Carlo Tree Search (MCTS) method to explore and remember all possible moves on the board.

This isn’t the first time Hassabis has stirred up a massive AI gold rush among tech giants.

In 2014, DeepMind used reinforcement learning to allow AI to learn to play simple video games. This achievement was astonishing, and DeepMind was directly acquired by Google.

Google’s bet turned out to be right.

In the next few years, DeepMind produced a result that shocked the world every once in a while.

Deep learning and reinforcement learning are solving many classic artificial intelligence problems, such as logic, reasoning, and knowledge representation

In 2016, the earth-shattering AlphaGo directly ignited the upsurge of deep learning and the first round of AI industry.

In 2017, AlphaGo Zero quickly surpassed AlphaGo without using human data.

AlphaGo Zero

In 2020, AlphaFold’s prediction of protein structure is comparable to laboratory technology, basically solving the protein folding problem.

In June of this year, AlphaDev created a new sorting algorithm, which may completely change the efficiency and results of computer science.

Compared with OpenAI’s more general route, DeepMind has been deeply involved in the vertical field for many years.

Where is the next big leap forward in language models? Gemini may point the way to the next generation of language models.

Last Stand

Clearly, Gemini is Google’s last stand.

Many of the technologies pioneered by Google, such as the Transformer architecture, have made possible the recent deluge of AI.

Because it is too cautious about the development and deployment of technology, it temporarily lags behind in the face of competition from ChatGPT and other generative AI.

In order to fight against ChatGPT, Google has continuously thrown multiple actions, such as launching Bard, and integrating generative AI into search engines and other products.

In order to concentrate on major tasks, in April, Google simply merged Hassabis’ DeepMind and Google’s main artificial intelligence laboratory, Google Brain, into Google DeepMind.

For the new team after the fit, Haasabis is obviously very confident. The new team, he says, brings together two forces that have been critical to recent advances in artificial intelligence.

“If you look at where we are in artificial intelligence, you would believe that 80% or 90% of the innovation in the future will come from one of the teams. Both teams have produced extremely good results in the past ten years. "

NEW IDEA

Training a large language model like OpenAI’s GPT-4 requires feeding into a “Transformer” a large curated dataset from books, web pages, and other sources.

The Transformer uses patterns in the training data to adeptly predict every letter and word that should appear in subsequent text.

This seemingly simple mechanism is very powerful in answering questions and generating text or code.

But this seemingly simple technical principle has also been criticized by many industry leaders or artificial intelligence experts.

Musk: The essence of current AI technology is statistics

LeCun: The current level of intelligence of AI is not as good as that of dogs

OpenAI’s breakthrough in the GPT series of models is based on Transformer’s core technology, and it aggressively uses RLHF to strengthen the model’s capabilities.

And DeepMind also has very rich experience in reinforcement learning.

This gives people very good reasons to look forward to the innovative capabilities that Gemini may demonstrate in the future.

More importantly, Hassabis and his team will also try to use core technologies in other fields of artificial intelligence to enhance the capabilities of large language models.

DeepMind’s technology accumulation is very extensive.

From robotics to neuroscience, they have a wide variety of gear in their arsenal to pick from.

For example, AI bigwigs like LeCun said that Transformer limits the ability of the language model too much to the scope of the text.

Like humans and animals, learning from the physical experience of the world may be the best solution for developing artificial intelligence.

Perhaps in Gemini, artificial intelligence will show potential in other directions.

Uncertain Future

Hassabis is tasked with accelerating the development of Google’s AI technology while managing unknown and potentially serious risks.

The rapid progress of large language models has caused many artificial intelligence experts to worry whether this technology will open Pandora’s box and make human society pay an unacceptable price.

Hassabis said that the benefits that artificial intelligence may bring to human society are immeasurable.

Humanity must continue to develop this technology.

Mandatory suspension of the development of AI technology is completely unworkable.

But that doesn’t mean Hassabis and DeepMind, led by him, will advance the technology recklessly.

After all, the reason why Google and DeepMind handed over the leadership of AI technology to OpenAI.

A large part of the reason is the “overly responsible” attitude towards AI development.

Netizen: not optimistic

But for the release of Gemini in the future, considering Google’s conservative attitude before, most netizens seem to be less optimistic.

When do you think this AGI-like model will be released?

I bet 10 bucks that Google will never release this thing.

If anyone has paid attention to Google’s projects, they will find that they generally brag for a while, then release nothing, and then kill the project a year later.

However, netizens still recognize Google’s contribution to the current large language model.

Netizen A: The large language model technology used by OpenAI is basically invented by Google

Netizen B: Yes, but Tesla cannot make a fortune, but Edison can.

This netizen is very optimistic that DeepMind will use its experience in reinforcement learning to make breakthroughs in large language models.

However, he still believes that Google may only use the idea of improving its existing products to advance this technology, rather than launching brand new products.

References: