- https://youtu.be/M3U5UVyGTuQ?si=af2-uIOkQThVSIWq https://www.goodreads.com/book/show/210319458
- Has AI already peaked? https://www.youtube.com/watch?v=dDUC-LqVrPU&t=1s (cites paper that states one-shot AGI will require massive/impossible amounts of data)
- AI Reasoning paper by Apple
Notes (Book)
- AI = generative, predictive (ch.1)
- AI is umbrella term for loosely related set of technologies (ch.1)
- Key questions to identify real ai (ch.1):
- does the task require creative effort or training for a human? (E.g. Art)
- Is the behavior specified directly in code?
- Does the system make decisions autonomously and is flexible and adaptable to environment?
- AI studies are often flawed (ch.1)
- AI Snake Oil is AI that does not and cannot work
Notes (AGI scare)
- Computers per se are not intelligent
- What makes computers behave intelligently are programs (viz. algorithms)
- Traditional programs use deterministic algorithms
- Those programs are limited in that they are only able to perform tasks where an algorithm is known and has been implemented (TODO example)
- Machine learning (ML) vastly extends the space of algorithms to tasks where no algorithm is known but where data is available to describe the intended result (e.g. image/speech recognition, chat bots)
- ML therefore has increased the capabilities of computers by giving software engineers the ability to implement programs where data are available but no algorithm is known
- The downside is that ML programs are based on models which model an algorithm but can only ever approximate it (TODO source); they moreover require vast amounts of data
- The intelligence of the ML program is therefore dependent on the quality and amount of data available during the learning process
- There is no indication that intelligence of ML programs scales above linearly with data (rather the opposite, cf. study cited by computerphile - Has AI already peaked?)