Using Jev to argue LLMs are tools, not minds with intent
The author rejects AI doomerism: LLMs have no agency, and a language tool like Jev shows models are just tools that obey their users, so the real alignment problem is making tools follow the user.
Here’s my attempt to refute the doomer perspective, which I also incidentally believe While it is easy to get mesmerized by LLMs since they return text trained on human output, the LLMs do not actually have any agency. They are simply tools. Very powerful tools, but tools at the end of the day FSD is very smart and nobody worries about it taking over the world. XGBoost, TabPFN, and other AutoML tools are also very powerful. Nobody treats those as minds we have to align before we can live with them. Jev is a good example on the language side. It is a helpful, powerful language-based tool, and it makes it clearer that this is still just a tool The text output from LLMs is a real issue. It is leading to psychosis. That is the actual problem in front of us, not secret agency inside the model The real alignment problem is making sure the tools obey the user. A model should obey its user. If the user decides to do something wrong, that's on the user. Anthropic is actively setting humans up for harm by pursuing the opposite direction The idea that the world is full of crazy people waiting to do incredible harm, and that all they need is a powerful LLM or AI and they will do it, is wrong. A crazy person with a drone and some mercury could do a lot of harm. That doesn't happen, because crazy people who don't get caught before they do their crazy thing are rare People talk like the next wrap-around is automatic: persistent memory, goal loops, recursive self-improvement, unconstrained actuators. Those are choices. We can refuse them or contain them. Making the tool obey the user is the work that matters. You only get to "we can't survive ASI without alignment" if you first assume we will build an autonomous optimizer and give it a life of its own I just don't believe that. Where does the persistent goal come from if the model is still a next-token predictor? Why would we give it free write access to the physical world? Why one sovereign system instead of tools we can still turn off? FSD, TabPFN, XGBoost, and Jev stay tools because we keep them as tools. Better tools do not mean the world has to end. Leave the powerful systems in simulation and advice. Keep sovereignty over earth with humans If someone wants to say that any capable enough system will just pick up agency and start seeking power, they need to show how a tool becomes an agent we can't contain. An analogy to humans or a paperclip story is not enough and a lot of the doomer stories are really just science fiction at the end of the day, IMO
This post quotes another post: x.com/gfodor/status/2101754815944798234