The Clear And Present Danger of Open LLMs

When Meta announced its latest large language model (LLM) in mid-April, it was a seminal moment in the relatively nascent field of generative AI.

For a start, LLama 3 set new benchmarks for accuracy and versatility to quickly establish itself as a pivotal tool in AI research and application. What’s more, anyone could download it, and with the right equipment, run and fine-tune it from their private systems.

Yet its release also caused disquiet among some in the industry who questioned the safety of democratizing AI this early in the technology’s developmental process, as noted by a report in Fast Company.

Available forever

On one hand, an open LLM like Llama 3 promotes transparency and could enhance public trust in the technology compared to an LLM operating in a closed architecture. And competition is always a good thing, isn’t it?

“With open-source LLM, organizations will have more capabilities to develop and deploy AI-powered solutions,” Rajiv Garg, a professor at Emory University’s Goizueta Business School was quoted as saying.

“Llama 3 is a solid model that will reduce the entry barrier significantly for most organizations – especially the ones that want to fine-tune these models with their internal data.”

Yet there is no doubt that these models can be exploited for harmful purposes, including the spreading of misinformation or propaganda. While problems or unsafe responses can be quickly rectified on ChatGPT or Claude, the same cannot be said of Llama 3.

In effect, there is no taking back Llama 3 once it is released.

An openly available model

It is worth noting that Llama 3 is best described as an “open weights” model. That means that the model offers some transparency regarding how it makes calculations. However, the full details to create it, such as the datasets used to train it, are not publicly available.

Indeed, Meta itself described Llama 3 as “the most capable openly available LLM to date”. On the other hand, Singapore’s SEA-LION that we wrote about last month is open-source. This means that weights and codes are released.

Whether open or open-source, the biggest fear revolves around their potential use for malicious purposes such as generating deepfakes or propaganda. Coupled with the problem of uncontrolled proliferation, and the potential for widespread impact on societies is no longer so far-fetched.

For now, a much larger “400B” version of Llama 3 is still being trained. Of course, it won’t run on the typical home system, no matter how powerful the GPU is. But who is to say that AI hosting firms won’t offer access to it on a pay-per-use basis?

Ultimately, regulation and guidelines around responsible AI solutions are necessary. The question is: Will the regulations come in time?

Image credit: iStock/ynny