LogFAQs > #932696976

Lol no

The breakthroughs in AI nowadays are largely thanks to the computational resources we now have and even then the greatest achievement is defeating humans in games with clearly defined rules, which has absolutely zero applications to anything related to the real world where you're not given an exhaustive list of rules. Reinforcement learning as it currently stands is also not the solution given how long it takes for an agent to learn how to not be useless and what they learn doesn't transfer to a different task.

If it weren't for GPUs being able to crunch an assload of numbers faster than CPUs, AI/ML would still be stuck in the late 80s and early 90s. And throwing more hardware at the problem isn't going to lead to something that is self aware. Figuring out how ML algorithms can leverage prior knowledge quickly instead of spending 100+ hours training to be useful in the loosest definition of the word is the critical thing to address.

Take a look at anything impressive in AI/ML/related fields in recent years like AlphaGo, YOLO, and Watson and realize that success there came from having a gigantic dataset that was already labeled by hand and/or a ton of computational resources that cost millions of dollars to acquire. And the results of this are being able to more quickly identify objects, play one game, or serve as a more sophisticated search engine

Last edit: An article about how Watson failed to even come close to delivering on its hype and a reminder that the Jeopardy appearance was already 9 years ago

https://spectrum.ieee.org/biomedical/diagnostics/how-ibm-watson-overpromised-and-underdelivered-on-ai-health-care

---

Roll Tide & Go Irish