From an AI safety perspective, it could make sense to use primarily small, specialized AI, but economic and structural incentives might lead to a world where every toaster, car and camera is part of the same, fully integrated overmind.
@plinz
-
LLMs as Brute Force Path to AGI: Small vs Planetary Systems
By
–
If LLMs turn out to be the brute force path to AGI, I expect we can create small (<<H100) AGI engines that derive most of what they need on the fly, and make use of interchangeable knowledge bases. But why use small special purpose systems when you can have a planetary mind?
-
Scaling Deep Learning: Master Algorithms Hidden in Training Data
By
–
My skepticism about scaling deep learning is shared by many old fashioned AI people, but it’s likely wrong. Perhaps there is a master algorithm hidden in the training data? Perhaps we can distill it? And if not, there are many ways in which a dumb AI can learn to improve itself.
-
Deep Learning Insufficient for AGI: Meta-Behaviors and Epistemology
By
–
I still suffer from the intuition that deep learning is the wrong approach. General intelligence cannot consist in learning to imitate all human behavior; there must be a small set of meta behaviors that lead to discovering epistemology, empirical learning, analytic thought.
-
Big Data Approach to AGI: Memorization vs. Curious Learning
By
–
My bias against the big data approach to AGI is fueled by how I observe my own learning. I always try to avoid memorization and focus on re-deriving, using, improving and integrating what I read. Memorizers do worse than curious explorers and experimenters.
-
Statistical Text Models and AGI: Semantics Extraction Beyond Imitation
By
–
I glimpsed during my work in Ian Witten‘s lab in the 1990ies that one can extract arbitrary semantics from pure statistical text models. I just never considered this to be very interesting, because I thought AGI would result from intelligent self improvement, not mere imitation.
-
AI Models Evolving Beyond Memorization Through Reasoning
By
–
I expected AI would become smarter than us before it knew very much. Current frontier models are idiot savant AI: less intelligent than competent humans, but they memorized basically everything. The models are now improving through reasoning, reinforcement learning and tool use.
-
US-China Competition: Democracy vs Authoritarianism in AI Era
By
–
The competition between the US and China is not merely about economic growth, control of birth rates and military power, it is about whether liberalism is able to hold its own against authoritarianism. If China can create better living conditions than the US, democracy may end.
-
Waymo Tests New Autonomous Driving Version in LA Traffic
By
–
Waymo needs to evolve, there is a new version being tested for LA traffic pic.twitter.com/8qtFs854Nt
— Joscha Bach (@Plinz) 12 juin 2025Waymo needs to evolve, there is a new version being tested for LA traffic
-
AI Authenticity: Rejecting Impersonation for Honest Interaction
By
–
If you were not impersonating someone who cannot fight back, it would be much more patient with you. As it is, you chose a very high bar, and it's on you when you fail it. Show up as a normal human being and I will measure you by the standards of us mere mortals.