Knowledge Vault 5 /83 - CVPR 2023
An AI Odyssey: the Dark Matter of Intelligence
Yejin Choi
< Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4o | Llama 3:

graph LR classDef ai fill:#f9d4d4, font-weight:bold, font-size:14px classDef challenges fill:#d4f9d4, font-weight:bold, font-size:14px classDef paradoxes fill:#d4d4f9, font-weight:bold, font-size:14px classDef limitations fill:#f9f9d4, font-weight:bold, font-size:14px classDef personal fill:#f9d4f9, font-weight:bold, font-size:14px A[An AI Odyssey:
the Dark Matter
of Intelligence] --> B[AI advancement: hyperbolic, unpredictable 1] A --> C[Transformers cant master compositionality 2] C --> D[Impossible possible with engineering 3] A --> E[Plasma generates better
procedural knowledge 4] A --> F[AI field full of paradoxes 5] F --> G[Dark matter of AI:
common sense 6] F --> H[GPT-4 struggles with
theory of mind 7] G --> I[Common sense trivial for humans,
hard for AI 8] F --> J[Moravecs paradox: reasoning easy,
sensorimotor hard 9] A --> K[Computer vision less
impacted by LLMs 10] A --> L[AI: generation easier
than understanding 11] A --> M[GPT-4 doesnt consistently
outperform others 12] A --> N[Speaker: late bloomer in AI 13] N --> O[Risks, strange problems
led to success 14] A --> P[Inclusive environment
encourages innovative work 15] class A,B ai class C,D challenges class F,G,H,I,J paradoxes class E,K,L,M limitations class N,O,P personal

Resume:

1.- AI advancement is hyperbolic and unpredictable. CVPR 2050 might be in the metaverse or on Mars.

2.- Transformers cannot truly master compositionality like multi-digit multiplication, despite appearances. Scrutiny in evaluation is important.

3.- Impossible things may be possible with the right engineering, like distilling good small models from weak large models.

4.- Plasma, a small model, can generate better procedural knowledge and plans compared to GPT-3.

5.- The AI field is full of paradoxes - GPT-4 passing the bar exam while having major limitations.

6.- "Dark matter" of AI is the unspoken common sense humans use to interpret language and images.

7.- GPT-4 struggles with theory of mind tasks involving multiple people, locations and objects. Solutions are unreliable.

8.- Common sense is trivial for humans but hard for AI. Yet common sense varies considerably even among humans.

9.- Moravec's paradox - high-level reasoning is easy for computers but sensorimotor skills of a toddler are hard.

10.- Computer vision is less impacted by large language models, possibly due to data quality issues. More work is needed.

11.- For AI, generation seems easier than understanding. For humans, it's the opposite - we can critique but not generate.

12.- Even for common sense reasoning, GPT-4 does not consistently outperform other large language models off-the-shelf.

13.- The speaker considers herself a "late bloomer" in AI, showing talent is made, not inborn.

14.- Taking risks and working on strange problems helped the speaker eventually find success over a decade.

15.- An inclusive environment that encourages diversity is important for people to gain confidence and do innovative work.

Knowledge Vault built byDavid Vivancos 2024