Meta recently released its latest AI model, Llama 3, claiming that this is the most powerful open-source model available. At the same time, Meta's chief AI scientist Yann LeCun said that Meta is also developing a more powerful new generation of Llama, and hinted that it has the potential to surpass the best closed AI models. Less than a week later, OpenAI announced the expansion of its enterprise-grade capabilities for API customers, further enriching the Assistants API capabilities, and introducing new tools designed to enhance security and management controls, as well as manage costs more effectively. In today's AI era, the showdown between Meta and OpenAI seems to be evolving into a battle between "open source" and "closed source".
"The Strongest Open Source Alpaca"
"Llama 3", also known as "Alpaca 3", is described by the producer Meta as the most powerful open-source AI model. How powerful is such a "strongest open source alpaca"? Meta has released a total of two versions of "Alpaca 3". One version has 8 billion parameters and the other has 70 billion parameters. "Parameters" is an industry term, and generally the larger the number, the more powerful the model. Yang Likun said that Meta is currently developing models with larger parameters, which can have more than 400 billion parameters in the future. "It will take a long time to adjust the models, but there will be multiple variants of these models in the coming months," Yang said at the Imagination in Action conference at MIT. Therefore, it is not yet clear when the model with more than 400 billion parameters may be released.
In addition, Meta has also launched an AI assistant "Meta AI" based on "Alpaca 3". "Tests have shown that these models are superior to previous open-source models with similar numbers of parameters," Meta said. Although major AI developers such as Google and OpenAI are currently retaining closed-source technology, Yang predicts that open-source AI models will progress faster. "Theoretically, they [open-source AI models] will drive AI to human-level intelligence much faster," he said. ”
Yang argues that broader software industry observations suggest that open source technology will eventually succeed. "Open source has become the dominant factor in software infrastructure. This is because it allows improvements to be shared more quickly and the code can be reviewed more thoroughly. More people review the code, which means better AI improvements. "The infrastructure needs to be open source – that way progress will be faster. ”
According to Wired, Meta's open-source AI strategy has accelerated the current wave around generative AI. When Meta released the "Alpaca 2" model in July last year, many startups, researchers, and business owners were able to download more powerful AI models to experiment with and build. While Google and OpenAI's GPT-4 and Google's Gemini are only available through paid APIs, Google made a smaller open-source model called Gemma available for download in February of this year. There are a handful of other companies that have chosen to join Meta's ranks and release powerful open-source AI models. Just a few weeks ago, startup Databricks launched a new model that is the most powerful open-source variant. If Meta's reported benchmark scores are correct, "Alpaca 3" seems to be even more powerful.
How does OpenAI respond?
While OpenAI appears to be dismissive of competing with Meta, less than a week after Meta launched Alpaca 3, OpenAI announced the expansion of its enterprise-grade capabilities for API customers. OpenAI has further enriched the capabilities of its Assistants API and launched new tools. This is done to enhance security and management control, as well as to manage costs more effectively. Olivier Godement, head of product at OpenAI, said in a media interview: "OpenAI is still leading the way when you talk to developers and businesses about meaningful AI model work. "That being said, we always welcome more competition – it's the way to make everyone better. ”
In this major security upgrade, OpenAI introduced private links. It's a way to improve safety. This also enables direct communication between Microsoft's Azure cloud services and OpenAI. OpenAI says this allows the data to minimize "exposure to the open internet" for customer data and queries sent through APIs. This addition complements the existing security stack, including SOC 2 Type II authentication, single sign-on (SSO), AES-256 data encryption at rest, TLS 1.2 transit encryption, and role-based access control.
In addition, OpenAI has introduced native multi-factor authentication (MFA) to strengthen access controls to meet growing compliance requirements. For healthcare companies that need to be HIPAA compliant, OpenAI will continue to offer business partnership agreements to eligible API customers, as well as a zero data retention policy.
Of course, OpenAI is playing more than just security. OpenAI has also upgraded its most important enterprise feature, the Assistants API. This upgrade allows the Assistants API to handle more than 500 times the original files. The upgraded Assistants API adds enhanced file retrieval capabilities, allowing each assistant to handle up to 10,000 files. Previously, there was a limit of 20 files per assistant. The Assistants API allows enterprises to deploy customized fine-tuning models and provide AI conversational assistants to their applications by retrieving enhanced generation (RAG) calls to specific documents.
Business media outlet VentureBeat reported that OpenAI's direction to improve its products is more focused on enterprise-grade security, management control, and cost management. These updates show OpenAI's intention to provide businesses with a "plug-and-play" experience to fend off the rise of open-source models like Alpaca 3 and Mistral. The open-source model may require more setup from the enterprise.
"OpenAI不Open"
It's just that as OpenAI has been running wild in the direction of "closed source", this company has also been criticized and ridiculed because of its name. OpenAI doesn't seem to be as open as it originally thought. Tesla and Twitter CEO Elon Musk was one of the founders of OpenAI when it was founded in 2015 and donated large sums of money. Last month, he noticed the changes in OpenAI, tweeting: "OpenAI was created as an open-source (that's why I named it "open" AI), non-profit company, with the aim of becoming a rival to Google. But now it's turned into a closed-source, profit-oriented company that's actually controlled by Microsoft. This is by no means what I want. The billionaire recently complained, "I donated about $100 million to a nonprofit, but it turned into a $30 billion for-profit company."
Another reason why OpenAI doesn't share the inner workings of GPT-4 could be legal liability. AI language models and image generators need to be trained using a lot of data from the internet. This may include copyrighted data. Ilya Sutskever, chief scientist and co-founder of OpenAI, did not answer when asked if OpenAI could be sure that its training data did not contain pirated material, according to The Verge. Fortune magazine also confirmed with OpenAI, but did not receive a response.
Among OpenAI's critics is Emad Mostaque, founder and CEO of Stability AI. This company has launched Stable Diffusion, an open-source image generator. Mostak has been making fun of OpenAI's name with this fact, criticizing OpenAI for not taking an open-source approach. He tweeted an appeal to OpenAI employees: "Any OpenAI employee who really wants to participate in the open source AI project can sign up with us. We'll give you a matching salary and benefits. You can choose any open source AI project to work on, whether it's ours or someone else's. Cooperative, open, prioritizing good results over self-interest. He changed the company URL in his bio to "actuallyopenai.com" (true open-source AI).
OpenAI's Sutskev argues that "open-source AI is unwise" and that this view is likely to gain more adherents, citing an increase in malicious use of AI. Wired magazine quoted people in the AI field as saying that Meta's "alpaca" model is not as open as imagined. Meta's open-source license on the model imposes some restrictions on what researchers and developers can build. Despite this, Wired magazine still believes that "Alpaca 3" is likely to trigger a new round of AI experiments. Clement Delange, CEO of HuggingFace, an AI technology sharing platform, said developers have created more than 30,000 variants of Llama 2. "I'm sure we'll also see a range of new models based on Llama 3," he said. "Meta has made an awesome community leadership move. ”
"Open source" or "closed source", both technology development directions have a lot of supporters. The competition between OpenAI and Meta is a reappearance of this long-standing problem in the AI generation. Still, as Fortune magazine quipped at the end of the article, "Criticism of OpenAI's name is likely to continue." ”
News sources: Wired Magazine, Fortune Magazine, VentureBeat, the Verge, some of the pictures are from the Internet
iWeekly Weekend Pictorial Exclusive Manuscript, please do not reprint without permission