Open-Source LLMs Are Far Safer Than Closed-Source LLMs While everyone agrees that we don't have AGI yet, a reasonable number of researchers believe that future versions of LLMs may become AGIs The crux of the argument that closed-source companies use is that it is dangerous to open-source these LLMs because, one day, they may be engineered to develop consciousness and agency and kill us all. First and foremost, let's look at LLMs today. These models are relatively harmless and generate text based on next-word predictions. There is ZERO chance they will suddenly turn themselves on and go rogue. We also have abundant proof of this - Several GPT 3.5 class models, including Llama-2, have been open-sourced, and humanity hasn't been destroyed! Some safety-ists have argued that the generated text can be harmful and give people dangerous advice and have advocated heavy censorship of these models. Again, this makes zero sense, given that the internet and search engines are not censored. Censor uniformly or not at all. Next, safety-ists argue that LLMs can be used to generate misinformation. Sadly, humans and automated bots are more competent than LLMs at generating misinformation. The only way to deal with misinformation is to detect and de-amplify it. Social media platforms like Twitter and Meta do this aggressively. Blocking LLMs because they can generate misinformation is the same as blocking laptops because they can be used for all kinds of nefarious things. The final safety-ist argument is that bad actors will somehow modify these LLMs and create AGI! To start with, you need a ton of compute and a lot of expertise to modify these LLMs. Random bad actors on the internet can't even run inference on the LLM, let alone modify it. Next, companies like OpenAI and Google, with many experts, have yet to create AGI from LLMs; how can anyone else? In fact, there is zero evidence that even if future LLMs (e.g., GPT-5) don't hallucinate or mimic human reasoning, they will ever have agency. The opposite is true - there is tons of evidence to show they are just extremely good at detecting patterns in human language, creating latent space world models, and will continue to be glorified next-word predictors. Just like a near 100% accurate forecasting model is not a psychic, an excellent LLM is not AGI! Open-sourcing models have led to more research and a better understanding of LLMs. Sharing research work, source code and other findings promotes transparency, and the community quickly detects and plugs security vulnerabilities, if any. In fact, Linux became the dominant operating system because the community quickly found and plugged vulnerabilities. AFAICT, the real problem with LLMs has been research labs that have anthropomorphized them (attribute human-like quality), causing FUD amongst the general population who think of them as "humans" that are coming for their jobs. Some AI companies have leveraged this FUD to advocate regulation and closed-source models. As you can tell, the consequences of not open-sourcing AI can be far worse - extreme power struggles, blocked innovation, and a couple of monopolies that will control humanity. Imagine if these companies actually invented AGI. They would have absolute power, and we all know - absolute power corrupts absolutely!!