What I suggested was a transfer of goals in a GPT-2, just like its synonym understanding of meanings for words. You teach it food/sex/immortalityOfTheSpecies are the strongest goals, then it spreads to tall buildings, diamonds, cars, animals etc. Then, when it thinks of plans, it questions the effect and if it harms any values. It's just one big GPT-2 that equalizes connections and does plan generating.
I am impressed with the arm and Hide&Seek OpenAI made though, the arm just adapts as listens to a plan....but the other achievement is strange, they learn tool use without planning anything. I really think that won't scale to building real rockets. Thinking of plans scales. Humans think, THEN test. Repeat. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tb23b29b1f508a00f-M93ced6438bdac62f50866cac Delivery options: https://agi.topicbox.com/groups/agi/subscription
