Everything related to artificial intelligence is developing too quickly. Within less than a week of Meta launching its AI model LLaMA2, startups and researchers have already developed chatbots and AI assistants using it. Some companies are beginning to roll out products using this model; it’s only a matter of time.

In my previous article, I focused on the threats that LLaMA2 could pose to OpenAI, Google, and other companies. Having a flexible, transparent, and customizable free large model can help small companies create AI products and services faster. In contrast, large, complex proprietary models like OpenAI‘s GPT-4 are somewhat lacking in this regard.

But for me, the truly impressive aspect is Meta‘s level of openness. It will allow a broader AI community to download and adjust the model. This will help make it safer and more efficient. Crucially, when it comes to the inner workings of AI models, it can prove that transparency has advantages over secrecy. This is timely and important.

Tech companies are racing to deploy their AI models into the real world, and we are seeing more and more products embedding generative AI. However, the most powerful models, such as OpenAI‘s GPT-4, are tightly guarded by their creators. Developers and researchers have limited access to these models through their paid website, without knowing the details of their inner workings.

This lack of transparency could lead to subsequent issues, as highlighted by a recent new preprint paper — researchers from Stanford University and UC Berkeley found that GPT-3.5 and GPT-4 performed worse than a few months ago in solving math problems, answering sensitive questions, generating code, and performing visual reasoning.

(Source: STEPHANIE ARNETT/MITTR | MIDJOURNEY (WORKERS, COGS))

Arvind Narayanan, a computer science professor at Princeton University, wrote in his assessment that these models lack transparency, making it difficult to say why this is the case, but nonetheless, caution should be exercised regarding the research findings.

They are more likely to be caused by “the way authors evaluate” rather than direct evidence that OpenAI has made the models worse. He believes that researchers did not consider that OpenAI has already fine-tuned the models to perform better, which may have unintentionally led some prompt techniques to no longer work as they did before.

Sasha Luccioni, a researcher at the AI startup Hugging Face, stated that companies that built and optimized products using a certain iteration of the OpenAI model are likely to see their products suddenly fail and experience performance issues.

When OpenAI fine-tunes its models in this way, for example, products built relying on specific prompts may stop working as they did before. She added that this closed model lacks accountability. “If you have a product and you change something in it, you should inform your customers.”

Open models like LLaMA2 will at least clearly indicate how the company designed the model and what kind of training techniques were used. Unlike OpenAI, Meta shared all “recipes” of LLaMA2, including how it was trained, what hardware was used, how data was annotated, and the techniques used to reduce harmful content. Luccioni said that those conducting research and building products on existing models need to know exactly what they are doing.

She said, “Once you can access and use this model, you can conduct various experiments to ensure you achieve better performance, or reduce biases, or whatever you want.”

Ultimately, the debate around the openness and closed nature of AI will come down to one point: who is in charge. With open models, users have more power and control. As for closed models, you will be at the mercy of their creators.

The release of such an open and transparent AI model by a large company like Meta feels like a potential turning point in the generative AI gold rush.

If products built on heavily marketed proprietary models suddenly crash embarrassingly, and developers do not know why, then an open, transparent AI model with similar performance may become a more attractive and reliable choice.

Meta is not doing this for charity. It can gain a lot from allowing others to explore the flaws in its models. Ahmad Al Dahle, vice president leading generative AI at Meta, stated that the company will learn from the broader external AI community and leverage this knowledge to continuously improve its models.

In any case, this is a step in the right direction, Luccioni said. She hopes that Meta’s actions will put pressure on other tech companies with AI models to consider a more open path.

“I am glad to see that Meta can maintain such an open attitude,” she said.

Author Bio: Melissa Heikkilä is a senior reporter for MIT Technology Review, focusing on AI and how it is changing our society. Previously, she wrote about AI policy and politics for POLITICO. She has also worked at The Economist and served as a news anchor.

Support: Ren

Operations/Layout: He Chenlong