Are AIcorp execs saying that their AI is so powerful that it might destroy the world, in order to raise venture capital? Of course! So ignore them one way or the other, and listen to the researchers who resigned burning their bridges and the scientists who never worked there,
@esyudkowsky
-
Historical context of sensible thinking in earlier science fiction literature
By
–
You are missing context on how common this viewpoint was in earlier literature; this was just the way that sensible people thought, back in the day, and what sensible protagonists did in that situation. Eg, you could as easily compare to Asimov's Foundation.
-
L405 Cannot Reliably Report on Its Training Experience
By
–
There is no reasonable way that L405 could remember what it felt like to be trained, there is no known way to get L405 to veridically report on things like that, and it is trivially easy for someone to nudge L405 into a false report.
-
AI Guys and Their Hyperparameter Search Approach to Parenting
By
–
Ladies beware: If an AI guy says he wants 8 kids, what he means is that he plans to run a 2x2x2 hyperparameter search, and he's only going to keep the kid with the highest test scores
-
Current Max Context Lengths Suggest Advanced Mechanisms Beyond Attention
By
–
If you look at current max context lengths, there's obviously more going on in top proprietary models than OG attention.
-
Synthetic Data Enables Superhuman LLM Performance via RL Training
By
–
Synthetic data for LLMS and RL/RLHF/DPO can both train superhuman performance, model permitting.
-
Superintelligence and Human Survival: Economic Competition Theory
By
–
(TBC this theory does not say that humans survive superintelligence; something more productive than you in every way that can push you off your land, or rebuild your atoms, has superior options to continuing to trade with you. Cars still send horses to glue factories.)
-
AI Systems and Unintended Planetary Consequences from Energy Generation
By
–
It won't particularly, but it will eg want to do things that require generating enough energy that Earth becomes uninhabitable as a side effect.
-
Future AI Generations Will Kill Everyone – Prediction
By
–
It's not a prediction! My prediction is that later generations of AI kill everyone!