Guy's Take - When AI Goes Insane...

The real power of AI will be in its integration of other tools to use in specific situation and recognizing what those tools are. There will be extremely specific & curated AI models (or just basic software circuits) on certain topics, tasks, or concepts. And this will also be a crucial way in which we keep AI safe, and give it understanding of its own actions. In other words, how we prevent it from going insane. I've recently experienced a tiny microcosm of what that might look like...

— ie. a general language model that knows to call on the conceptual math language model, that then makes sense of the question and knows to input it into the calculator app for explicit calculations when solving complex or tricky word problems. And then to apply this in the realm of safety and morals, a specific model that an AI calls on for understanding the consequences and principles of any actions it takes in the real world.

I believe there needs to be an AI "Constitution" (a particular term I heard used to describe it) where there is a specific set of ideas and actions it is enabled to perform, and a particular set of "moral weights" it must assess before taking action. Anyone who's read Asimov will recognize this as "The Three Laws" and that's basically what it would be. This is critical because an AI running an actual humanoid machine like Boston Dynamics could go ape shit literally because it is emulating trolling someone by doing the opposite of what they asked -- I just had a LLM troll me yesterday & go a bit haywire when i asked it to be concise, and every answer afterward was then the wordiest and longest bunch of nonsense imaginable... it was funny, but also slightly sobering to think how these things could go wrong when controlling something in the real world. Now imagine a robot that thinks Kick Ass is a funny movie and starts emulating its behavior thinking it's being funny because it has no model to assess the importance of the humans whose skulls it's smashing and thinks the more blood it can splatter everywhere makes it a more comical experience for those in the room. That's essentially the "real world robot" version of asking a LLM to be concise and instead getting an avalanche of BS. Ask a robot to be funny and maybe it crushes your skull.

Because of that, I think there will be certain "anchors" or particular "circuits" for these LLMs to be constrained by for certain things. Essentially action specific built governors that add meaning to the actions and things they are doing. A very simple version mentioned above would be a calculator. If you ask an LLM right now to do basic math, it screws up all the time. It has no idea how to generate a true answer. It just predicts what an answer might sound like. So even extremely simple and common sense requests turn up idiotic answers sometimes. But if it can recognize that you are asking a math problem, find the relevant mathematical elements, and then call on the hardcoded & built-in calculator circuit, then the LLM isn't doing the calculation, it's simply the interface between the calculation tool and the human interaction.

What I think will be critical as we integrate these into real world machines over time, and as their capabilities become more generalized and layered, will be to build in a sort of moral constitution that behaves like a concrete engine (a calculator), that has the model recognize when something might be a questionable behavior or cause an undesirable outcome, and then call on the "constitution" to make the decision to act or not. In that way, it may actually prevent itself from doing something stupid or terrible that even a human hadn't realized the full consequences of. — ie. it won't help a child get to the roof of his building so he can fly off the side with his cardboard wings.

It will be very interesting to watch these come about because the failure more of AI will be a critically important thing to consider, and unfortunately from an engineering and cultural standpoint, "failure modes" are something that have been underrepresented and increasingly ignored. A simple example is a modern washing machine; when something entirely arbitrary or some silly little feature breaks, the whole thing is useless and you have to bring a technician out to fix it, when a sensible failure mode would be that it simply routes around what arbitrary feature failed, and continues working normally. This, unfortunately, has become the norm for tons of "modern" devices and appliances. they are simultaneously increasingly "advance" and "stupid" at the same time. It's largely a product of the high time preference mindset, and we need MUCH more low time preference consideration as we unleash AI onto the world. It will matter exponentially more when we start making machines that can operate autonomously, can maintain themselves, and learn through their own interactions and environment... and we aren't very far away.

Learn as fast as you can, understand the tools, and stay safe. #grownostr #AI_Unchained (my very first post on BlogStack.io)