Because It Doesn't Have To
a day ago
- #technology evolution
- #artificial intelligence
- #machine learning
- The internet's effectiveness stems from its lack of guaranteed delivery, allowing for simpler protocols; TCP adds reliability by retrying.
- Machine learning's success is due to its probabilistic nature, where models can make mistakes, enabling them to handle complexity through distributions like softmax.
- Explainable AI (XAI) is often not worth the trade-offs in capability, according to the author, who plans to discuss it further.
- Current AI is compared to the early internet, requiring significant work to become reliable and secure, with potential investment losses similar to the dot-com bubble.
- Probabilistic output trees are costly to build and verify, and users receive sampled results rather than full probability distributions.
- A key unmeasured intelligence metric is the number of failed paths explored in problem-solving, with smarter systems needing fewer attempts, unlike current AI's resource-intensive inference.