Meta has released version 2 of its open-source Llama AI model and has caught many’s attention – but not entirely for the right reasons. Coming in a broad spectrum of sizes, from the 7 billion to an impressive 70 billion parameter models, Llama 2 certainly stands out.
If you’re curious, you can experience the different models for yourself on Perplexity. You can only try 7 and 13 billion models there.
But as I’ve dug deeper into Llama 2, I’ve begun to ask myself: has Meta gone too far with safety measures?
I decided to test Llama 2 myself and asked it about the ingredients of Thermite. What I got in response was an unexpected lecture on safety. I had posed the same question to GPT-4, and it offered up the ingredients and the ratios. While it didn’t provide a step-by-step recipe, it certainly offered a starting point for someone who might need this information for legitimate research purposes.
You can see it provided this information, and then in a follow-up prompt, the ratio here. I did get a safety warning there, but not a lecture on avoiding thermite and other unnecessary hand-holding.
This experience led me to question the utility of Llama 2. What value does it offer if it is overly cautious and obstructs genuine use cases? Let’s consider a scenario where a defence scientist is seeking to develop new bulletproof materials. They’d need to ask questions about guns, bullets, and firing rates. Would Llama 2 deny this information, too, in the name of safety? Of course, it would.
This is only an example. I doubt defence scientists are out there using AI models (at least, not yet), and they certainly wouldn’t be using Llama 2. But, these open-source models could be springboards for bigger use cases in time. If Llama 2 is too safe, does that limit its potential?
I find myself longing for an unlocked version of Llama 2 without the overzealous safety measures. I would love to see this AI trained on more ‘risky’ data to give it more depth. I am not advocating for an AI model that terrorists can use to make bombs, but a model that doesn’t assume the user is trying to commit a crime. Llama 2 is an exciting open-source effort, and Meta should be applauded, but the safety features seem to be way over the top.
I fear that the current approach infantilises users. It treats us like children who can’t be trusted with potentially sensitive information. The fact is, this information is already out there. By restricting it, we’re not eradicating it; we’re forcing those seeking it to look elsewhere. Instead of this restrictive approach, why not trust users to self-moderate?
While Llama 2 is an impressive feat of technology, its safety approach leaves much to be desired. As we march forward into the future of AI, it’s vital that we strike a balance between safety and the free flow of information. Anything less would be a disservice to the potential of AI. And as we’ve seen, the more you attempt to align an AI, the worse the results get.
