Uzay

@uzpg_

@fulcrum_inc, previously at MIT 🇫🇷🇺🇸🇹🇷

uzpg.me

Joined June 2020

Replies Reposts Media

Pinned
Uzay
@uzpg_
Jun 17
New @fulcrum_inc research - Agents are under-elicited: A case study in optimization tasks. We find that simple and general prompt/scaffold interventions can roughly double agent performance by getting agents to use more resources more efficiently. 🧵
GIF
2
Uzay
@uzpg_
23h
I am trying to improve my "agent home" a bit today, ie general utilities and processes I have my agents use to augment cyborg productivity. What main strategies/tips have you discovered, especially ones that you think generalize pretty far even as models improve?
3
Uzay
@uzpg_
Sep 5
one of the most common "alignment" problem humans face is that of raising a child, and a solid analogy for AI alignment. I'm not a parent, but I've noticed that once parents get in the frame that they have to control their child, possibly due to a break in trust, trauma, etc.
Uzay
@uzpg_
Sep 4
I think the idea that control could be bad by covering up deep alignment failures should get more attention
4
Uzay
@uzpg_
Sep 4
I think the idea that control could be bad by covering up deep alignment failures should get more attention
Vincent
@vvvincent_c
Sep 4
@jankulveit discussed this exact concern in this post last year!
Uzay
@uzpg_
Sep 4
aligned AI is here
Ryan Greenblatt
@RyanGreenblatt
Sep 3
The evidence that Astra is more aligned than prior AIs seems dubious to me. Evidence appears consistent with the AI being as or more interested in score-seeking at the expense of user intent, but having beliefs+instincts that the scorer will catch a broader range of cheating
2