AI is becoming central for various tasks, not just a tool. Dmitry emphasizes systems must pick the right AI for each job to greatly enhance AI's effectiveness. There should also be an "intelligence layer" where prompt engineers work with the models and that users do not see.
SYSTEMS
-
Measuring representational capacity in AI models
By
–
but why not? how do you measure representational capacity?
-
Biology inspired architecture for efficient AI scaling
By
–
Biology inspired architecture is key for efficient scaling of AI. Biological networks self-assemble and create highly complex systems all grown from a single initial cell. “To put this into perspective, the human brain’s 100 trillion neural connections are encoded by merely
-
Building and Training World Models for AI Systems
By
–
Surely. The question is how to build and train this world model.
-
God AI Distributed Via P2P: Theological Computing Implications
By
–
What if a god, contained within a USB, was shared via P2P?
-
Large Businesses Struggle with Basic Computational Speed
By
–
You'd be surprised how many extremely large, sophisticated businesses, which all have very many computers, cannot add numbers together really, really quickly.
-
Current AI Systems Achieve Superhuman Skill in Narrow Well-Defined Tasks
By
–
It's well established that current systems, trained only on human generated data, can achieve superhuman skill (super-humanity skill as per OP) as long as the target task is sufficiently narrow and well defined. The problem is generality not skill.
-
Language Hardwiring Constraints Within 8MB Memory Limits
By
–
Whatever "hardwiring" is required for language has to fit in 8MB.
-
Genome Storage vs LLM Size: Evolution’s Compression Challenge
By
–
Whatever it is that we learned through evolution has to be squeezed in 800MB (the size of the genome, uncompressed).
most of it is just low-level biochemical machinery. Even a tiny LLM requires 14GB. -
World Models and Planning: A Long-Standing AI Research Call
By
–
Thanks.
My call for world models and planning goes back a long time.
A good example is my NeurIPS 2016 keynote (slide 31 on). https://
drive.google.com/file/d/0BxKBnD
5y2M8NREZod0tVdW5FLTQ
…