This question is quite out there, but does anyone have any idea how fast a model specific chip would be? Like now we have TPU's, which are application specific. But how much faster would a chip be that was purely made to process a layer of GPT-3, compared to running such a layer on a TPU?
Generally dedicated hardware logic will be a couple of orders of magnitude faster. But obviously if you encode the model into hardware then it won’t be learning any more.
Just some thoughts on the security issue. In my opinion the problem is not really the "black box" aspect but the possibility that a general AI becomes significantly (and increasingly) more intelligent than the most intelligent humans and in consequence even if it could be retracable how the AI came to a certain result it would not help, if no human can understand it, as we would not be intelligent enough. A suggestion for the foreseeable future would be to have "multilayered AI systems" in a sense that you have different AI models solving the same problem and than evaluate the outcomes, if they are very similar you have at least a certain amount of risk reduction. Another aspect which I find very interesting is that the average human / society might reject decisions which would be evidently the best decision for humanity but would collide with our average habits, biasses or interests of people in power. Furthermore the problem is not always to answer a question but to put the right question. If you would ask a highly developed AI without moral instances implemented "how can I make the highest profit with an insurance policy without violating the law", the outcome would most likely not be beneficial for the majority. And finally, if we implement safety measures the only hold as long as AI is not a multiple more intelligent as we humans and is not widely implemented and importantly interconnected. An AI with an intelligence of a multiple of us humans would for sure find a way to "outsmart" us and implement in hardware would have the tools to accomplish its targets. Of course we are far away form such scenarios, but nonetheless it is worth thinking abut it.
Why are the comments sections of these videos always filled with empty, generic, sycophantic, fawning praise? E.g. "Really amazing podcast. I wish more people can hear this." It's as if the creators of the content paid a handful of fake fans to post fake comments of praise. Does this signal a healthy, vibrant, innovative ecosystem? In a healthy ecosystem, one would expect to see deep discussions, questions & answers, exploration, teaching & learning, etc. There is none of this here. Just the same bland, generic praise, again and again. As if chanting some superstitious mantra in hopes that it'll bring good luck or ward off evil. So stupid.
You are doing just the same what you criticize, but on the negative side. What about starting a discussion on an interesting topic? If you want an interesting discussion give it at least a try and start on yourselves......