AlphaFold 2's strongest competitor self-demolition! Meta disbanded the 12-member team in situ, and may replace ChatGPT in September
Dare to challenge AlphaFold 2, and finally make some achievements in the ESMFold project, and quit if I say no.
Just today, Meta disbanded the team that used AI to predict nearly 600 million protein folds to focus on commercial AI.
We all know that the protein prediction models AlphaFold and AlphaFold2 successively released by DeepMind are tsunami-level existences in academia, which are enough to change human beings.
At that time, Meta also saw the significance of open basic scientific research to human beings.
In July 2022, the disbanded team members jointly released ESMFold, the largest protein prediction model after AlphaFold2.
There are 15 billion parameters, which can increase the folding speed by 60 times.
However, Meta's move shows that it is abandoning pure scientific research projects and turning to the development of profitable artificial intelligence products.
The 12-member team is disbanded
Informed sources revealed that Meta disbanded ESMFold's team with 12 people.
The time for the disbandment, also said to be this spring, coincided with widespread layoffs at Meta.
In March, Xiao Zha sent an all-staff letter to all internal employees, again emphasizing what he called the "Year of Efficiency".
In the next few months, Meta underwent a major reorganization, including a flattened management structure and layoffs, affecting about 20,000 employees.
Plus, previous calls from Meta investors to focus on profitability and growth. It is not difficult to see that Meta did this mainly for profit.
Yaniv Shmueli, Meta AI Research Scientist and Engineering Manager, previously worked on the ESMFold team.
“Meta is trying to adjust its research strategy to understand more about how to create advanced AI that can help its own business, rather than just doing some curiosity projects,” she said.
In July 2022, Meta officially launched its own protein prediction model - ESMFold.
ESMFold is a Transformer-based 15 billion parameter language model.
By using the attention mechanism, it can learn the interaction mode between pairs of amino acids in the input sequence, and realize the high-accuracy, end-to-end, atomic-level structure prediction directly from the sequence of individual proteins.
In terms of accuracy, ESMFold is comparable to mainstream protein prediction models such as AlphaFold2 and RoseTTAFold.
But in terms of inference speed, ESMFold is an order of magnitude faster than AlphaFold 2.
At that time, LeCun also tweeted that this is a great new achievement of the Meta-FAIR protein team.
Meta has yet to confirm whether ESMFold will continue in the future, but says the data is currently still available to the research community.
Catching up with Microsoft and Google, ChatBot launched in September
Now, Meta's new focus is to leverage its long-standing research and development in AI to create products that are hyped around generative AI.
In addition to Google, Microsoft and other established technology giants, Meta is also one of the first companies to invest in AI research.
In 2013, Meta established the Fundamental AI Research (Fair) laboratory, hired leading academics, and devoted itself to artificial intelligence research.
Over the years, the company has been recognized by the scientific community for its advances in artificial intelligence.
However, after the outbreak of generative AI, so far, Meta has not announced chatbots developed by Google, OpenAI, etc.
In February of this year, a generative AI team led by Meta product director Joelle Pineau was officially launched.
Currently has more than hundreds of employees, including employees transferred from Fair Labs.
According to the report, Meta plans to launch a series of chatbots with different character styles as early as September in order to catch up with competitors.
At the annual Connect Developers Conference held in September, Xiao Zha will release the roadmap for AI artificial intelligence products for the first time, as well as more details about the products.
Zuckerberg said that I envision artificial intelligence that can act as an assistant, a coach, or an agent that can help you interact with companies and creators. I think people are unwilling to interact with a single artificial intelligence.
Inside Meta, it is rushing to prototype chatbots that can engage in human-like communication with its 4 billion users.
Meta is also exploring launching a chatbot modeled after Lincoln, and another offering surfer-style recommendations for travel options, according to a person familiar with the plans.
In doing so, Meta will not only improve user engagement, experts say, but also collect a wealth of new data about user interests.
Doing so may help Meta better target users with more relevant content and advertisements.
Remember, most of Meta's $117 billion in annual revenue comes from advertising.
Meta's vice president of AI research, Joelle Pineau, said Meta remains committed to conducting exploratory research based on open science.
Projects born out of Fair and then fed into other areas of the business have always been an integral part of how the team operates. This allows us to apply learnings and techniques from FairAI research into products.
In the long run, Meta is also meant to explore, an avatar chatbot that can be applied in the Metaverse.
Xiao Zha is also spending all his time and energy conceiving of this problem. After all, after the company changed its name, it cannot lose its original intention of focusing on the Metaverse.
Although Meta has not yet released a generative AI product, it has been making contributions in this field.
After the open source of Llama1 ignited the enthusiasm of the community, some time ago, Meta open sourced the Llama 2 model again, and made it available for free and commercially.
On the other hand, Meta is also building AI infrastructure and trying to procure tens of thousands of GPUs. Even, self-developed custom chips to accelerate AI research.
For the latest chatbot, Meta insiders say the company is likely to build technology that screens users' questions and automatically checks the output to make sure it's accurate and free from hateful or rule-breaking speech.