Article reprinted from: Machine Heart
There are pros and cons to open source and closed source in AI, especially in the era of large models. What’s important is how to write good articles during use.
Image source: Generated by Unbounded AI
People have always been divided over the choice of open source and closed source in the field of AI, but in the era of big models, the powerful force of open source has quietly emerged. According to an internal document leaked by Google, around open source models such as Meta's LLaMA, the entire community is rapidly building models similar to the capabilities of OpenAI and Google's big models.
Meta is undoubtedly at the heart of the open source world, and continues to make open source efforts, such as the recent release of Llama 2. However, the tree that stands out in the forest will inevitably be blown down by the wind, and Meta has recently been in "trouble" because of open source.
Outside Meta’s San Francisco office, a group of protesters holding signs gathered to protest Meta’s strategy of publicly releasing AI models, claiming that these released models have caused “irreversible proliferation” of potentially unsafe technologies. Some protesters even compared the large models released by Meta to “weapons of mass destruction.”
The protesters, calling themselves “concerned citizens,” were led by Holly Elmore, an independent advocate for the AI Pause movement, according to her LinkedIn profile.
Photo credit: MISHA GUREVICH
She noted that if a model is proven to be unsafe, the API can be shut down, such as large models such as Google and OpenAI that only allow users to access them through the API.
In contrast, Meta's LLaMA series of open source models make the model weights available to the public, allowing anyone with the right hardware and expertise to replicate and tweak the models themselves. Once the model weights are released, the publishing company no longer has any means to control how the AI is used.
In Holly Elmore's view, releasing model weights is a dangerous strategy, as anyone can modify the model and these models cannot be recalled. "The more powerful the model, the more dangerous this strategy is."
Compared to open source, large models accessed via APIs often have various safety features, such as response filtering or specific training to prevent the output of dangerous or offensive responses.
If the model weights are released, it becomes much easier to retrain the model to jump over these "guardrails". This makes it more possible to use these open source models to create phishing software and carry out cyber attacks.
Photo credit: MISHA GUREVICH
Because she believes that part of the problem is that "there are not enough security measures for model release" and there needs to be a better way to ensure model security.
Meta has not yet made any comments on this. However, Yann LeCun, chief AI scientist at Meta, seemed to respond to the statement that "open source AI must be banned" by showing the grand occasion of the open source AI startup community in Paris.
There are also many people who disagree with Holly Elmore and believe that an open strategy for AI development is the only way to ensure trust in technology.
Some netizens said that open source has both advantages and disadvantages. It can allow people to gain greater transparency and enhance innovation, but it also faces the risk of abuse (such as code) by malicious actors.
As expected, OpenAI was once again mocked, “It should return to open source.”
Many people are concerned about open source
“Wide release of advanced AI models will be particularly problematic in the future because preventing their misuse will be nearly impossible,” said Peter S. Park, a postdoctoral researcher in AI safety at MIT.
However, Stella Biderman, executive director of EleutherAI, a nonprofit artificial intelligence research organization, said: "So far, there is little evidence that open source models have caused any specific harm. It is also unclear whether simply putting a model behind an API will solve the security problem."
“The basic elements of building an LLM are publicly available in freely available research papers, and anyone in the world can read the literature and develop their own model,” Biderman said.
She added: “Encouraging companies to keep model details secret could have serious consequences for transparency, public awareness and scientific development in the field, particularly for independent researchers.”
Although everyone is already discussing the impact of open source, it remains to be seen whether Meta’s approach is truly open enough and whether it can leverage the advantages of open source.
“The concept of open source AI has not been properly defined,” said Stefano Maffulli, executive director of the Open Source Initiative (OSI). “Different organizations use the term to refer to different things — meaning different levels of ‘publicly available’ — and that confuses people.”
Maffulli pointed out that for open source software, the key issue is whether the source code is publicly available and can be reproduced for any purpose. However, if you want to reproduce an AI model, you may have to share the training data, the way the data is collected, the training software, the model weights, the inference code, etc. Among them, the most important thing is that the training data may have privacy and copyright issues.
OSI has been working on a precise definition of "open source AI" since last year, and is likely to release an early draft in the coming weeks. But in any case, he believes that open source is essential to the development of AI. "If AI is not open source, we can't have trustworthy and responsible AI," he said.
In the future, the disagreement about open source and closed source will continue, but open source can no longer be stopped.
Original link: https://spectrum.ieee.org/meta-ai