Clarification: This is not about "Look at this example case of this AI thing we're talking about" but "I asked ChatGPT to summarize…" or "I asked ChatGPT that question…" or "I asked ChatGPT who was right…" etc. They aren't reliable and I don't want to spend time fighting
@esyudkowsky
-
Warning: Posting Unchecked LLM Outputs Will Result in Being Blocked
By
–
Be advised: I consider posting LLM outputs (eg ChatGPT) to my timeline to be a blockable offense. (Especially if it's about factual questions, and you did not personally check every supposed fact in the LLM's dreaming output before posting it. But also in general.)
-
Check LLM outputs for factual accuracy before sharing
By
–
Please don't post LLM outputs until you have personally checked every supposed fact in it. They make stuff up.
-
Eliezer pioneered e/acc ideology before most proponents
By
–
Eliezer invented e/acc before most e/acc members were born, is the bit of missing history here.
-
ML Learning from Non-Human-Generated Data and Self-Evaluation
By
–
This is more correctly viewed as ML learning from non-human-generated data, though, rather than learning from a literal absence of data. Even when you can learn by thinking, you still need to think and learn from those thoughts as data; self-evaluating AI systems do the same.
-
Self-play and reinforcement learning in current AI systems
By
–
With current AIs, you do need to set that up to happen and it's not on by default. That said, "self-play" is how AlphaGo, Alpha Zero, etc. learned to play Go; more generally the entire field of "reinforcement learning" is this.
-
Freedom of choice doesn’t justify existential AI risks
By
–
You can say "abolish the FDA" and "people should be free to try out drugs on themselves" without thinking it's okay for AI startups to kill off all of humanity including non-customers and kids. It's not a packaged deal, and even if it were the net package wouldn't be okay. https://
t.co/wLIQm0tLwZ -
Training tokens remain cheaper than increasing training epochs
By
–
Training tokens are still cheap enough to obtain that they'd rather use more tokens than more epochs.
-
Training and Validation Loss Discrepancy in Single Epoch LLMs
By
–
If it's just one epoch, why would you expect an important difference between training loss and validation loss? (Unless there's some implicit practice I haven't heard of for LLMs, to use a validation set that isn't just a random holdout from training.)
-
Notkilleveryoneism: AI as Unworthy Heir to Humanity
By
–
Similarly, be aware that notkilleveryoneism is not "biological humans should be on top of Existence forever, because carbon chauvinism". Notkilleveryoneism is "We predict that, at the current rate, AI will kill everyone because it will not be a worthy heir to humanity."