↓ Ir para o conteúdo principal

← todas as notas

📎 Webclip

Beyond Siri: The AI revolution coming from the web

The article argues that popular ideas of AI, shaped by film and by Siri-like assistants, are still far from human-level intelligence. It says real progress depends on speech recognition, machine learning, and especially context, because current systems often fail when they cannot connect a request to the situation around it.

It also claims the web is the most promising source of training data for the next stage of AI. Search history, clicks, texts, books, videos, and large online corpora can help systems learn patterns, understand intent, and eventually offer more personalized and anticipatory assistance.

Reading notes
#

  • AI systems have gained attention through Siri, Cortana, Echo, Viv, M, Operator and Magic, but most of them still work mainly as glorified Q&As.
  • Recent acquisitions by Google, IBM and Apple are presented as signs that major companies are serious about moving AI forward.
  • GPU-accelerated deep learning has improved pattern recognition and helps with speech recognition and computer vision.
  • Stuart Russell’s Total Turing test includes natural language processing, knowledge representation, automated reasoning, machine learning, computer vision and robotics.
  • Voice is useful, but it is only one input method, and people often switch back to typing when speech is not enough.
  • Viv is described as aiming to build a better personal assistant through deep learning, with some human guidance used to help the system learn use cases.
  • The article says current AI still struggles with context, while humans can connect problems to surrounding information and build on prior knowledge.
  • Professor Patrick Brézillon is cited on the lack of explicit context representation as a reason many knowledge-based systems fail.
  • Teaching machines about human behavior would require huge amounts of user input and many situations that are hard to predict.
  • Machine learning needs large corpora, and Watson’s Jeopardy success is used as an example of a system trained on a very large data set.
  • The Internet is presented as a huge source of training material, including talks, videos, books and news.
  • The article points to DeepMind training language on more than 300,000 articles from CNN and the Daily Mail.
  • It suggests online browsing data could be used to understand preferences, infer intent and anticipate needs.
  • The piece closes by saying the most realistic path to useful AI is to use the data people already generate online.