
CEO @bitreel; big fan of pirates
Boston, MA
Joined November 2010
- Tweets341
- Following429
- Followers119
- Likes1.1K
Alignment question: should we replace LLM training data with a corpus of information for what we want the world to be? And then optimize parameters to maximize the likelihood of this training data?
In a world of super intelligence, is the new “adaptation tactic” for humans to stay relevant (via-a-vis other humans… not super intelligence) the ability to change?
Statement: the only thing that separates a sentient being from a non-sentient being is the ability to set goals
Hypothesis: LLM don’t need to be able to predict like a world model would, if your goal is to focus on achieving goals in modern society.
This is because humans are not rational, and so the ground truth to strive for is not “predict what will happen in the real world” but how would a human react.
“Short-term memory” and “long-term memory” (for both biological and artificial/computational systems) is a misnomer. Time is not the variable we should be focused on. Memory isn’t time-bound in the way most of us think. Instead, there is just working memory.
Everything else (what we’d refer to as “long-term memory) is just a database. Storing information in a database is a solved issue.
