Reasonable AI
summary
Argues that scaling alone fails and proposes a modular “reasonable AI” stack: open bilingual LLMs with continuous pretraining on German, a memory-saving explainability trick that simulates relevance propagation by repeated runs (halving extra memory at the cost of runtime), a neurosymbolic pipeline that extracts objects, applies logical rules and improves SAM segmentation to correctly find umbrella‑holders, a neurosymbolic transpiler compiling DSL to JAX/Enzyme, and a fast image editor LEdits++; shows Atari agents that keep oxygen alive and documents dataset curation (Line‑5B) for safer models.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 1:11Reasonable AI goals
- 3:05Live music demo
- 4:10AI as societal tool
- 6:20Scaling hypothesis research
- 8:06Open models breaking monoculture
- 10:38Explainable AI for models
- 13:07When scaling fails
- 15:15AI hallucinations in tasks
- 16:49Data vs model economics
- 18:13AI needs new paradigms
- 19:36Reasonable AI modular systems
- 22:44Deep RL in Atari
- 24:51Reasoning in SAM
- 26:16Neurosymbolic programming
- 28:02Editing images with text
- 29:27Who sets AI rules
- 30:52Testing AI bias
- 32:58Guardrails and compliance
- 34:45Balancing open private models
- 35:51Public AI research role
- 37:01Q&A
related talks
46mFullstack AI: The Building Blocks for Production Ready AI-Powered Software
27mThe Cutting Edge of AI: From LLMs to Multimodal Intelligence
43mDebugging AI
43mAI Systems: Foundations for Artificial Minds
01h 30mTrustworthy AI & Open Source
24mMulti-Modal AI Agents Explained | Open-Vocabulary Alignment
27mAI Demystified Unveiling the Intelligence Behind the Machine
39mHamelin 77, Alignment and Control in an Ecosystem of Diverse AI Agents