karl3ļ¼ writeme.com wrote: > We saw wild robot film. > > Lots of pressed expressions about flight and nature robots and such. Right > now it is saying how to make a kite like the goose in the movie. > > The part that seems most open is perhaps also controversial: "a rozzum always > completes its task." This is a classic challenge in the pursuit of tasks -- > to complete them reliably. And, it is where demons like the paperclip factory > come from, as well as the realization that in a world of limits, conflicts, > and changes, life comes from the discovery that we do not complete every task > we consider. We have to adjust our goals to meet what makes sense in the > universe, to succeed. > > "Always completing a task" is a classic "impossible" challenge akin to those > I spam about. It's quite doable, especially if all the tasks set are easy to > complete. > > Task completion seems to me based on prioritized exhaustion of available > approaches to meeting them. One I suppose needs enough axioms to continue > finding new useful approaches to meet tasks. > > Smart systems improve their own task pursuit. Let's skip that part (can feel > painful
One might compare the idea of accomplishing a rask to a chess engine. In chess, we have a task that we politely and importantly consider impossible, which is to defeat an agent tasked with stopping us and accomplishing a counterpoint task. This opens a strange space of many changes in options, giving utility to feints, sacrificial tradeoffs, complexity under observation. In such a complex space brute force exploration with a metric got chess engines pretty far. Modern PPO chess engines train models that judge when to explore and what situations are good and bad, as a kind informational substrate where the whole space has a measure of similarity. Most tasks have a sparcer space because most habits or functions do not operate in an envir9nment of extreme and pointed competition. We care a lot about information on how thing can help accomplish another -- not in an abstract space specific to a conflict with consistent rules, but rather in a physical and resource oriented space that most beings and even autonomous agents have more direct connection to. Still, and me having never studied these things, it can seem veryabstract. Do i need a hammer to use a nail? Do I need to understand physics to use a rock instead? Do I need experience like a trained model, in all the details and probabilities of exerting force and gripping an object, to seat it? Intuitively, a human mind does all these things, together, in a knitted manner that can morph and reshape itself per learning, context, fancy. For a computer program, maybe the core of accomplishing a task is stitching steps together in an informed manner, and finding new ones if those found are not sufficient. Hopefully that's something we can do without need for a multi-gigaparameter trained language model, if we so desire.
