Arsalan
@bum_py
Followers
74
Following
2K
Statuses
650
HMI Enthusiast @IIM_Calcutta @iitkgp
Bangalore
Joined October 2013
1. Humans have implicit agency. (Currently) an LLM is “guided” by prompts & RL to look for patterns & connections. It doesn’t do *anything* by itself. Every prompt/fine-tuning is conditioned modelling on an expected set of outcomes. So it can’t make a new connection by itself. 2. The way attention works, retrieval happens top-down. You look for a broad search criteria, and then drill down, like searching for a needle in a haystack. What you’re saying is, the model should find two needles, establish a connection between them, and create a new needle, all by itself.
0
0
0
1. Humans have implicit agency. (Currently) an LLM is “guided” by prompts & RL to look for patterns & connections. It doesn’t do *anything* by itself. Every prompt/fine-tuning is conditioned modelling on an expected set of outcomes. So it can’t make a new connection by itself. 2. The way attention works, retrieval happens top-down. You look for a broad search criteria, and then drill down, like searching for a needle in a haystack. What you’re saying is, the model should find two needles, establish a connection between them, and create a new needle, all by itself.
0
0
0
@waitin4agi_ Agreed. But Twitter helps a lot
Twitter really shows your ideas/thoughts aren’t weird, just locally unpopular.
0
0
4
@pranitbauva1997 Ahh I tried to connect its API to some LLM but the list of things that could be done was pretty limited. Hope they launch a chat native interface
0
0
1
@cursor_ai AND Devin, but native.
Today, we are infusing the power of agentic AI into the GitHub Copilot experience, elevating Copilot from pair to peer programmer 🤖 (1/4)
0
0
0
I think about this Avengers: Age of Ultron scene a lot. That’s how I envision AI safety to look like, a core set of protocols at the heart of the model.
0
0
0
0
0
1