Crush GPT-4! Google DeepMind CEO self-revealed: the next generation of large models will be combined with AlphaGo

Google DeepMind CEO Hassabis has newly revealed: the new Gemini model will be combined with AlphaGo and large language models, and the cost is expected to be tens of millions of dollars, even hundreds of millions.

Google, it's really broken.

Is the legendary Gemini, which combines AlphaGo and GPT-4 models, finally coming?

One is the AI system that uses reinforcement learning to defeat human Go champions and create history, and the other is the strongest multi-modal large model that currently dominates almost all the big model lists, riding the dust, and the two AIs are simply invincible!

Google DeepMind CEO Hassabis recently told foreign media Wired that Gemini is still under development and will take several months, and Google DeepMind is ready to spend tens of millions of dollars, even hundreds of millions.

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

Too long to look at the version

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

Gemini is a large language model, similar to GPT-4

It is expected 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 problems by learning from trial and error

The tree search method helps to explore and remember possible movements in a scene, such as in a game scene

DeepMind's extensive experience in reinforcement learning will bring new capabilities to Gemini.

Other fields of technology, such as robotics and neuroscience, will also be integrated into Gemini

The next algorithm is to surpass ChatGPT

According to OpenAI CEO Sam Altman, GPT-5 is still some time away from release and will not start training for at least 6 months. The release date of the Gemini has not yet been determined, but it could be within a few months.

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

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

The scene where AlphaGo defeated Go world champion Lee Sedol in 2016 is still vividly remembered.

Hassabis said, "Gemini combines some of the advantages of the AlphaGo system with the amazing linguistic capabilities of large language models, so to speak. And we have some other interesting innovations."

Gemini is said to have multimodal capabilities not available in previous models and is very efficient at integrating tools and APIs. And Gemini will come in a variety of scales designed to support future innovations in memory and planning.

In March, there was a saying that Gemini would have a trillion parameters like GPT-4. What's more, Gemini is said to be trained using tens of thousands of Google TPU AI chips.

At Google's developer I/O conference last month, Google mentioned that from the beginning, Gemini's goal was to be multimodal, efficient integration tools and APIs.

At that time, Google's preview was: "Although it is still early, we have already seen multimodal capabilities in Gemini that have never been seen in previous models, which is too impressive."

The technology behind AlphaGo is reinforcement learning, which is the first technology pioneered by DeepMind.

RL agents interact with the environment over time, learning strategies through trial and error, maximizing long-term cumulative rewards

Through reinforcement learning, AI can adjust its performance through trial and error, and thus learn to deal with tricky problems, such as choosing how to take the next step in a Go game or video game.

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 by tech giants.

In 2014, DeepMind used reinforcement learning to let AI learn to play simple video games, a surprising achievement that directly led to DeepMind's acquisition by Google.

Google's bet was right.

Over the next few years, DeepMind produced a stunning result every once in a while.

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

In 2016, the groundbreaking AlphaGo directly ignited the boom 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 was comparable to laboratory technology, basically solving the problem of protein folding.

In June, AlphaDev created an entirely new sorting algorithm that could revolutionize the efficiency and outcomes of computer science.

Compared with OpenAI's more general route, DeepMind has been deeply engaged in vertical fields 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.

fight to the death

It's clear that Gemini is Google's backwater.

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

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

To counter ChatGPT, Google threw multiple actions in a row, such as launching Bard and integrating generative AI into search engines and other products.

In order to concentrate on big things, in April, Google simply merged Hassabis' DeepMind and Google's main artificial intelligence lab, Google Brain, into Google DeepMind.

For the new team, Haasabis is clearly very confident. He said the new team brings together two forces that have been crucial to recent AI advances.

"If you look at where we are in the human space, you believe that 80 or 90 percent of the innovation in the future will come from one of the teams. Both teams have achieved fantastic results over the past decade."

New ideas

Training a large language model like OpenAI's GPT-4 requires feeding a large selection of datasets from books, web pages, and other sources into a "Transformer."

Transformer uses patterns in the training data to expertly 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 bigwigs or artificial intelligence experts.

Musk: The essence of AI technology is statistics

OpenAI's breakthrough in the GPT series of models is to radically adopt RLHF to strengthen the capabilities of the model on top of the core technology of Transformer.

DeepMind also has a lot of experience in reinforcement learning.

This gives good reason to expect the innovative power that Gemini may demonstrate in the future.

More critically, Hassabis and his team will also try to augment the capabilities of large language models with core technologies from other areas of AI.

DeepMind's technology accumulation is very extensive.

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

AI bigwigs like LeCun, for example, say that Transformer limits the capabilities of language models too much to the scope of text.

Like humans and animals, learning from the world's physical experience may be the optimal solution for developing artificial intelligence.

Perhaps in Gemini, AI will show potential in other directions.

Uncertain future

Hassabis' mission is to accelerate the development of Google's AI technology while managing unknown and potentially serious risks.

The rapid progress of big language models has led many AI experts to worry about whether this technology will open Pandora's box and impose an unacceptable price on human society.

Hassabis said that the potential benefits of artificial intelligence to human society are immeasurable.

Humans must continue to develop this technology.

A mandatory moratorium on the development of AI technology is completely inoperable.

But that doesn't mean Hassabis and his DeepMind will push the technology forward recklessly.

After all, Google and DeepMind ceded the leadership of AI technology to OpenAI.

A large part of the reason is the "too responsible" attitude towards AI development.

Netizen: Not optimistic

But for the future release of Gemini, because considering Google's conservative attitude, most netizens seem to be not very optimistic.

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

I bet 10 pieces Google will never release this thing.

If anyone has followed Google's project, they will find that they usually blow a bull for a while, then publish nothing, and then cut the project a year later.

However, for Google's contribution in the current big language model, netizens still recognize it.

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

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

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

Still, he thinks Google will only advance the technology with ideas that improve its existing products, rather than launch entirely new products.

Resources:

https://the-decoder.com/deepmind-founder-shares-details-on-gemini-googles-next-gen-response-to-gpt-4/