'LLaMA 2' Released... "Progress as Open Source"
Concerns and Criticism Over Flood of Misinformation and Spam
"I believe that the more open the ecosystem becomes, the more progress will be possible. That is exactly why we released the open-source model 'Llama 2'." - Mark Zuckerberg, CEO of Meta Platforms
On the 18th (local time), Meta released the large language model (LLM) Llama 2 as open source. Since OpenAI's ChatGPT was made public in November last year, a global generative AI craze has swept the world, and Microsoft (MS) and Google have entered a competition to seize the initiative. The core of generative AI training is the LLM. Unlike other companies that have kept LLM technology confidential, Meta has taken a new strategy by releasing its technology as open source.
◆ "Available for Commercial Use" Meta's Hidden Intentions
Meta adopted an open-source strategy from the first time it released 'Llama' in February. What differs this time from the previous release is that while the initial Llama was open-sourced but not allowed for commercial use, this time it is stated that it can be used for free for both research and commercial purposes. Meta also declared partnerships with other companies such as MS, Amazon, Qualcomm, and LG.
Additionally, performance has been significantly upgraded. Llama 2 supports versions with 7 billion to 70 billion parameters and was pre-trained on 2 trillion tokens. The amount of context it can process at once increased dramatically to 4,096 tokens compared to previous models. The more context it can handle, the more information can be processed with a single command. Meta has strengthened its technology while expanding the fields where it can be applied.
Regarding the reason Meta released the LLM developed with its own personnel and resources as open source, CEO Zuckerberg emphasized on his Facebook page that "when software is open, more people can carefully examine it and potentially identify and fix issues." He highlighted the positive aspects for the entire industry.
In fact, from Meta's perspective, external programmers can utilize the open-sourced LLM and provide information on any issues encountered, which can be used for AI experimentation.
Meta's open-source strategy also appears to reflect its intention to widely distribute the technology it developed amid fierce competition among big tech companies like MS and Google for AI leadership. The New York Times (NYT) compared Meta's Llama 2 to Google's decision to release the Android operating system as open source in the past. At that time, Google adopted an open-source strategy to compete with Apple's iPhone, which eventually allowed Android to become the core smartphone software. Similarly, Meta seems to aim to seize leadership starting from the core AI technology, LLM.
MIT Technology Review evaluated that "Meta hopes to gain an advantage over competitors like OpenAI by releasing Llama 2 as open source."
◆ Criticism of "Open Source Flooding with Misinformation" Continues
However, Meta's open-source strategy has also faced criticism in the AI industry. With insufficient AI regulations and many social and economic challenges yet to be resolved, there are concerns that open source could lead to greater harm. There are fears that using open source could cause an explosive increase in spam, financial fraud, and misinformation.
OpenAI, which sparked the generative AI wave, is a representative company voicing such criticism. OpenAI was founded in 2015 as a nonprofit to prevent AI technology from being monopolized by specific companies and initially promised to make its technology public, but recently it has shown a change in stance.
Illya Sutskever, co-founder and chief scientist of OpenAI, said in an interview with IT media The Verge in March that the company's initial plan to open-source its technology was "wrong," adding, "Within a few years, everyone will clearly realize that releasing AI as open source is not wise."
Meta stated that the risks are not significant even though Llama 2 is released as open source. The problems of misinformation and hate speech are already occurring even without open source being public, and Meta's management believes that other tech companies will strengthen their countermeasures to address these issues, according to the NYT.
Meta also explained that it operated a 'red team' in advance to identify and address problems. A red team is an organization that simulates attacks on its own or clients' systems from a hacker's perspective. Meta also created and released a separate guide to help developers use the model responsibly.
However, although Meta released Llama 2 as open source for transparency and industry development, it has been criticized for not disclosing the data used to train the model. Meta released a research paper explaining Llama 2 but only described the training data as "publicly available online data." When Llama was first released in February, the data used was specifically listed in a table, but this time it was different.
The U.S. economic media Business Insider pointed out that this explanation is unusual. In the AI industry, disclosing what information was used in a model is common practice to assess the model's performance.
Business Insider mentioned that between the initial release of Llama and the release of Llama 2 five months later, copyright and other issues related to data usage arose, suggesting that Meta likely decided not to disclose the data used until legal matters are clarified.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
