| Aspect of Shen's Work | Description | | -------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------- | | Challenging AI Paradigms | Shen's work challenges the traditional computer science paradigm by creating AI systems that infer what we want rather than being explicitly told | | Technical Development | Built planning compilers and policy programming infrastructure as tools to create AI systems that better understand human intentions | | Interdisciplinary Approach | Deeply linked reverse engineering of human theory of mind with engineering more capable AI systems | | New Models of Human Intelligence | Challenged assumptions in computational cognitive science by creating models of fallible human rationality that are more accurate and realistic | courage of interdiscplinary - ai ssytem what we want () -> challengs cs - planning compilers prob.prog (inference subgoals) - ai faculties (understanidng minds of toher agent) - diverse (reverse engineering) - helped pragramatic - cognistiv (fallable human rationality) ambition and playfulness (ai system learn the rules of a game; can't tell the rules of the game) construction cooperative intelligence (infer andd understand others' goals) collaborative robots (infer subtask), - reliably, safely, theory of mind as invesrse planning sequential invers plan search (real time open-eneded bayesina goal inference) : prob.prog (prob.ai)+ model based planning cooperative communication + coooperative decision making -> C:O{S} ⭐️refinement 🙋‍♀️connect boltzman rationality (inver rl) with percept odeule 🙋‍♀️ assign nonzero probability to action that is not optimal (implemnting!!!!) goal inference = refinement planning more tractable goal -> plan 🙋‍♀️first make second harder -> (by introducing planning) hundreds of goals ## SIPS partial pla for the goal (1/9) (total particle for each action) 🙋‍♀️downweighted when it didn't predict the observation - initialize - resample wehn ESS is lower than N/2 - agents start backtracking (yellow gems are downweighed) much faster than irl etc # not only optimal but also robust red gem allows for the (once you unlock the door, it should go to 0) set of ' particle mangement, programmable particle filtering pf)initialize, update, less particle = human like (approximate bayes inf p) modeling utterances with LLM likelihoods infer commands from ratinal language - infer goal sumbolic, proabilistic, only action or only utterance we don't know the intention but combining we can 🙋‍♀️-> connecting this with resource use and uncertainty ⭐️lesely -> decision theoretic pioneer