On Monday, December 16, 2024, at 1:17 AM, YKY (Yan King Yin, 甄景贤) wrote: > Hello, long time no see 😁 > > There are 2 levels: we humans understand the objective of the game right > from the start. A simple reinforcement-learning (RL) agent does not "know" > what's going on, it just learns from rewards. I guess RL can learn > multi-agent, cooperative games. By observing play moves and outcomes, the RL > agent could build a "model" of the game, which is equivalent to understanding > the game. > > On another level, an AGI with human-like understanding can read the game > description and start playing with meaningful moves right away. Yes, indeed! Hello again!
Yes, I do not disagree. I have provided a benchmark concept for AI (or AGI) systems that is relative to other AI/AGI systems that involves detecting patterns and biases in arbitrary data, in a very general sense. This is central to any AI/AGI system, regardless of how it understands the game. If this is not relevant to your concept/system can you explain why, and or provide alternative benchmarks that are more relevant to your system? I'm not asking rhetorically, I am genuinely curious. In fact, I think I would rather see the benchmarks you plan to create for your system before looking at the concepts/architecture etc. That way I (and others) can see what goals you are trying to achieve, and how to test how far you are in terms of achieving those goals, plus other highly relevant aspects such as what the inputs/outputs are. One example: my goal is to achieve better than X % accuracy at solving these sets of CAPTCHAS (and provide the corpus of CAPTCHAS). I'm not saying the benchmarks have to be perfect, and of course they can evolve over time, but I think you have to start moving towards something tangible, and hopefully something relevant. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T8685950780e86bd5-Mb270d5b222a8c532a8ce4a11 Delivery options: https://agi.topicbox.com/groups/agi/subscription
