Adapting to Change: How Machine Intelligences Adapt to a Changing World
summary
Argues that neural adaptation centers on changing cortical reference frames (about 150,000 cortical columns) and that machine models freeze their world view until retrained; cites Thousand Brains theory, place/grid neurons, and embedding spaces. Shows supervised learning limits (need labeled data), gives gecko vs chameleon example (including a satanic leaf‑tailed gecko), notes GPT‑3 scale (~175B parameters), and describes production tooling: TFX pipelines, automated continuous training, and MLOps Level 1 for pipeline automation to keep models current.
AI-written summary · about
chapters (26)
- 0:00Intro
- 0:55Change can be stressful
- 1:55Do I need to adapt?
- 2:51Mental adaptation
- 4:01Neocortex
- 5:47Cortical columns
- 6:45Reference frames
- 9:54Machine intelligence
- 10:05Artificial neural networks
- 11:00Supervised learning
- 12:32Prediction aka inference
- 14:26Expectations
- 15:27Machine expectations
- 16:45Labeling data
- 17:50Consequences
- 18:17MLOps
- 19:41Production ML
- 21:35Tales from the trenches
- 21:53Production ML research
- 22:29Process
- 22:40What is MLOps?
- 23:10CI, deployment & training
- 24:35Training & deploying models
- 25:29TFX production components
- 26:12MLOps Level 0
- 29:46Outro
related talks
59mIntelligence and Machines: Creating Intelligent Machines By Modeling the Brain
32mHow Neuroscience Makes AI Intelligent
39mChanging Realities - The Final Frontier for Machine Learning?
51mComputing like the Brain: the Path to Machine Intelligence
29mDeep Learning, Neuroscience and the future of AI
09mLightning Talk: Intellectual Cognitive Development in Artificial Intelligence
36mScaling Laws, Emergent Behaviors, and AI Democratization
48mKeynote – Sameer Singh