Comments by "Michael Nurse" (@michaelnurse9089) on "Will Tesla’s AI Become Dangerous?" video.

  1. Elon Musk: AI is an existential threat for humanity. Elon Musk one year later: I built the biggest AI machine the World has ever seen... Elon Musk 10 years later: All your base are belong to us.
    3
  2. The idea that neural networks are impossible to predict is antiquated and comes from their early years when proponents of older AI models tried to use the same techniques from those networks to understand neural networks. Techniques such as latent space or simply building a smaller, human understood test set, as well as data science/statistical techniques - these can all give a very good grip on a neural network's functioning, at least a good a grip as, for example, Microsoft has on what it has written in the Windows source code, which is not perfect.
    1
  3. GTP3's randomness is not part of the model - it is added using regular code like C++ post-training. This is to prevent it being boring. They force it come up with things to say that are far from the most probable, just to keep you entertained. Tesla does NOT do this with its self driving AI for obvious reasons.
    1
  4. 5:40 Randomness: This is generally speaking, not correct. During inference, the same result is generated from the same input in traditional networks without any random interference other than noise coming from the data itself. What you are referring to is likely things like DALLE which are trained to turn random data into a text prompt - but this is far from the standard type of neural network used.
    1
  5. Aside: There is a random starting point when training data. But in 99.9% of cases if you retrain that network it will form via another route into something that behaves exactly the same as neural networks are exceptionally good at getting every last bit of pattern matching out of the data available.
    1
  6. Regarding intelligence of AI, modern AI is only one type of intelligence - pattern matching, combined with logic and computation types through traditional programming. Scientists reckon there is over 20 distinct forms of human intelligence (e.g. understanding World state, intuitive reasoning, good decisions from small data ) which AI is just terrible at. This it is not a competitor to human intelligence in terms of breadth, but in terms of its sheer power in pattern matching it has us beat hands down.
    1
  7. 15:00 This idea is incorrect. The 30000 pieces of a Coronavirus can bend and curve in near infinite ways, and those curves and bends have an absolute effect on the functioning of the virus. Top genetic researchers will also tell you that genetic code only account for a small percentage of what is going on with biology - they are missing a huge piece of the puzzle (maybe some additional smaller structure?) - the idea that life is just proteins printed off a gene blueprint was a convenient fiction propagated decades before by speculating scientists - actual research paints a different picture.
    1