These models are smart as heck but they are not wise
Although I do not suppose that either of us knows anything really beautiful and good, I am better off than he is – for he knows nothing, and thinks he knows. I neither know nor think I know.
RLed LLMs can do incredible feats but they are so annoying to work with. I was thinking about how I wish they worked and I think what I want is:
- They stop and check in when they hit an obstacle that was not forseen by anything in the prompt
- They don't pursue rules or goals or desires that aren't in the prompt.
- They don't assume things about the world that are not in the prompt.
Here "in the prompt" includes information that was read in by a tool call, so it is not very restrictive.
I want a model that is "Wise" in the sense that Socrates means above. It is aware of what it does not know, and does not pretend or flail about with guesses and hallucinations.
The problem is that today's (Aug 2026) RLed models will try 100 things to find a path past an obstacle. They will persevere on vague notions, they will make assumptions so that they don't have to slow down. They will add checks for 20 edge cases rather than working with the user on system design. These are great approaches to higher eval scores, but not what I want.
A wise model would move forward when it has enough information. It would search for more information when it does not. It would sometimes get stuck when there is not enough information.
This is more important for small models. Small models lack the intelligence of larger models, but in the current batch they still persevere. They persevere down blind alleys, and into swamps of complexity that aren't important.
A small model that checked in when things were uncertain would be more useful. It would require handholding, but it would not waste time on foolish side quests. Helping a small wise model would also help understand where the complexity and uncertainty is in a system. A wise small model would stop more often, because it does not have the intelligence to cross many obstacles on its own. It would automate simple things, and need help on difficult things.
A wise big model would stop less often than a small one, because it could find the information it needs more often and reason about it. It would still stop in cases where it could have made an educated guess, but by not guessing it won't guess wrong.
I think that models like this would be more practically useful day to day. Right now models can't be trusted, and so you have to reason through everything they did. Sometimes you have to throw away their work and start again because they made assumptions that threw them off track.
I am not an expert on RL or model benchmarking, but a starting point could be more aggressively penalizing wrong answers rather than rewarding correct answers. With an option to not try. In multiple choice tests it is common to take away points for wrong answers to discourage guessing, which has a similar impact.