In symbolic AI humans control too much, and in neural networks too little.
@pmddomingos
-
Why LLMs Need So Much Help
By
–
If LLMs are so smart, why do they need all these prompts, harnesses, post-training, scaffolding, etc.?
-
AI brings previously unreachable goals within range of ambitious people
By
–
AI is a gift to ambitious people, because it brings within range things that previously weren't.
-
Reinforcement Learning as Time Travel for Rewards
By
–
Reinforcement learning is time travel for rewards.
-
LLM as a search engine that never fails
By
–
An LLM is a search engine that never fails, because when it can't find a document it makes up.
-
AGI Requires Novel Task Mastery
By
–
AGI hasn't been achieved until AI beats or ties humans at tasks neither has done before.
-
AI Is Not Solved Yet
By
–
No part of AI is even close to being solved. This includes vision, language, reasoning, and agents. Massive datasets + benchmarks = illusion of intelligence. For example:
-
AGI testbeds must be i.i.d. samples
By
–
If your testbeds aren't i.i.d. samples from the space of AGI-relevant testbeds, performance on them is not an indication of AGI capability.
-
AI Coding Assistants Starting with ‘C’
By
–
The one thing we can safely predict for AI in 2026 is that the dominant coding assistant will start with a C.
2023: Copilot
2024: Cursor
2025: Claude Code