Maybe it's cynical, but I feel similarly and also wonder, what do these people see that I don't? Smart people with supposedly a lot of inside information. But when I hear them talking about LLMs with words like "thinking" or "intelligence", it seems like it either has to be fearmongering for ulterior motives, or otherwise I live in a completely different branch of reality from them.
Their models are still, and based on this technology always will be, token predictors. How is a text generation engine going to directly endanger human lives?
It won't. It only can if someone hooks up tools to them that can perform damaging actions. Someone has to build that toolset. Someone has to give those tools the affordances to take harmful actions. And then someone has to write a "harness" that uses the token predictor as an authoritative input for instructions. Doing so falls somewhere on a spectrum between gross negligence and malevolence.
From a distant enough viewpoint it might not look like a big difference, but I think it crucially shifts the responsibilities. What is the actual difference between driving a potentially dangerous system with inputs from an LLM versus inputs from a fuzzer, or a website with injection vulnerabilities? I think it's the ones building and configuring the tools downstream that are responsible for knowing the risks and build under the assumption of the worst happening.
Heaven knows I'm not one to defend the big LLM companies, but I don't see how this is OpenAI's responsibility. They're building and selling an autocomplete system. How is it their fault if people ask the autocomplete for harmful commands and then feed them into an over-privileged system to run them? (I know the waters are muddied a bit by some labs acting like this themselves right now).
As much as I don't agree with the authors framing of and terminology around LLMs, reading the article generously, maybe this is exactly what he felt, too. That it's not at OpenAI where his concern and insight is needed most, because the token predictor isn't actually the dangerous part. Maybe he'll look where he can bring safety consciousness downstream, where people are hooking up these outputs into systems and in ways that they shouldn't. Don't let the LLM pick your bombing targets. Don't give the LLM sufficient rights to the hospital's file storage to be capable of ransomware-ing it. Don't strap the loaded machine gun to the dog.
Their models are still, and based on this technology always will be, token predictors. How is a text generation engine going to directly endanger human lives?
It won't. It only can if someone hooks up tools to them that can perform damaging actions. Someone has to build that toolset. Someone has to give those tools the affordances to take harmful actions. And then someone has to write a "harness" that uses the token predictor as an authoritative input for instructions. Doing so falls somewhere on a spectrum between gross negligence and malevolence.
From a distant enough viewpoint it might not look like a big difference, but I think it crucially shifts the responsibilities. What is the actual difference between driving a potentially dangerous system with inputs from an LLM versus inputs from a fuzzer, or a website with injection vulnerabilities? I think it's the ones building and configuring the tools downstream that are responsible for knowing the risks and build under the assumption of the worst happening.
Heaven knows I'm not one to defend the big LLM companies, but I don't see how this is OpenAI's responsibility. They're building and selling an autocomplete system. How is it their fault if people ask the autocomplete for harmful commands and then feed them into an over-privileged system to run them? (I know the waters are muddied a bit by some labs acting like this themselves right now).
As much as I don't agree with the authors framing of and terminology around LLMs, reading the article generously, maybe this is exactly what he felt, too. That it's not at OpenAI where his concern and insight is needed most, because the token predictor isn't actually the dangerous part. Maybe he'll look where he can bring safety consciousness downstream, where people are hooking up these outputs into systems and in ways that they shouldn't. Don't let the LLM pick your bombing targets. Don't give the LLM sufficient rights to the hospital's file storage to be capable of ransomware-ing it. Don't strap the loaded machine gun to the dog.