4 ms·
Unlike human beings, LLMs have no explicit world model, they're an LLM. Why would a politician be replaced by an LLM? It seems like a highly inappropriate use.
by hammyhavoc 3y ago
Unlike human beings, LLMs have no explicit world model, they're an LLM. Why would a politician be replaced by an LLM? It seems like a highly inappropriate use.
- AnimalMuppet 3y agoHmm. If you trained an LLM on all the feedback a Senator receives from constituents, that could be pretty effective. Of course, once word gets out that that's what's going on, it will turn into a contest of which side of an issue can spam the Senator's office harder...
- s3p 3y agoI'm not sure this answers OP's question.
- hammyhavoc 3y agoWhy would it be effective? If it only deals in language, but not an understanding of human issues, economies et al, it would be ineffective and inappropriate. > it will turn into a contest of which side of an issue can spam the Senator's office harder An LLM wouldn't even be required for analysing sentiment and recurring topics, this is how a lot of brands have handled millions of inbound tweets all these years. To imply that a senator is just the political equivalent of a customer service agent is disingenuous when they themselves draft and propose legislation et al, and that legislation hopefully comes from a nuanced understanding of what it is that they're targeting, which an LLM wouldn't have.
- AnimalMuppet 3y agoI don't want an LLM writing legislation. That way lies disaster. (And that disaster is uncomfortably likely to happen within 10 years...) But I'm cynical enough to think that training an LLM on constituents' comments, and then having it write a senator's speeches, could be effective. The speeches would lack specifics, but the tone would echo the constituents' concerns. That might be good enough to get re-elected. (As I said, I'm cynical. It shows up in me thinking that fine-sounding speeches are likely enough to get a senator re-elected.)
- seydor 3y agoNeural networks are great at optimizing when the objectives are agreed upon and well known, and will not abuse power for their own gain (that is one objective). Why would you want humans instead?
- hammyhavoc 3y agoBecause human beings understand the way the world works, but an LLM only deals in language, whereas societies, economies et al are incredibly complex and nuanced, and also not defined by language. Whilst a hypothetical application of a neural network politically may not abuse power for its own gain, that doesn't preclude it from making uninformed decisions, just like being a human being doesn't preclude someone from making poor decisions. How would constituents hold a neural network accountable? All kinds of shitty things could be done "because the neural net says so", and that lack of accountability is in and of itself a fundamental problem, especially the lack of human-readable data set as it is stored as tokens, it's only the output that is human-readable, so we wouldn't even be able to tell in what ways it had been manipulated. The ultimate patsy and scapegoat for real human beings to take advantage of.
- InitialLastName 3y ago> when the objectives are agreed upon and well known, The above is inherently not a feature of the problems that political systems are intended to solve.
- seydor 3y agowhat are political systems trying to solve?
- InitialLastName 3y agoIf the solution to a problem is obvious and agreed upon by everyone, you don't need political mechanisms to make a decision. Politics is best-suited to issues of resource contention and coordination that aren't distillable to a clear set of goals that everyone agrees on.
- patrickas 3y agoLLMs do have explicit world models that can be even manipulated. There are many recent papers on the subject.
- hammyhavoc 3y agoChatGPT itself tells me it has no explicit world model.
- NumberWangMan 3y agoI mean, just like you can create 1-line python script that claims "I am an AGI" and have that fact be false, you can have ChatGPT tell you it has no explicit world model, while exhibiting behaviors that can only really be explained by it having some sort of model of the world inside it. Fine-tuning is like a PR agent teaching someone what sorts of things not to mention on TV even though they may be true.
- hammyhavoc 3y agoWhat are some of these behaviours?
- NumberWangMan 3y agoPatrickas responded with some examples: https://news.ycombinator.com/item?id=35966307 https://news.ycombinator.com/item?id=35966307
- patrickas 3y agoIt is not about what the model tells you. This paper shows an emergent world model in an LLM that was taught to play otello moves https://ar5iv.labs.arxiv.org/html/2210.13382 https://ar5iv.labs.arxiv.org/html/2210.13382 https://arxiv.org/pdf/2303.12712.pdf https://arxiv.org/pdf/2303.12712.pdf This paper discusses (among other things) how a GPT4 model navigated between rooms in a text adventure game and was able to create a map afterward. Literally building a model of the world as it was navigating and drawing a map of that afterwards