3 ms·

Thank you for the well-informed comment. I have to respectfully disagree that self driving cars are pretty much synonymous with general AI. You remind me that D

by bpiche 5y ago

Thank you for the well-informed comment. I have to respectfully disagree that self driving cars are pretty much synonymous with general AI. You remind me that DeepMind just recently said that all we need for AGI is reinforcement learning. And much of what self driving cars do is just reinforcement learning. But I have to say, I don't think it's that simple. Reinforcement learning is just classification with a reward function. Is that all people are, just classifiers with a reward function? I don't think so. And I don't think it's just a god-of-the-gaps argument applied to AI, either. Having worked with all kinds of classifiers and recommendation systems for years, these things just seem like tools that need to be calibrated to me. We can automate a lot of stuff with classifiers, especially 'bullshit jobs' like call centers and movie recommendations and gunshot detection. But there are many tasks for which there is just no training data, or way of incorporating that data into a machine. Basically, some tasks are solved problems, and others are totally unsolvable, and that category is where humans come in. Natural language generation comes to mind. For all the noise about GPT-3 and Copilot, they're not going to be ready for production for a long time until we figure out very complicated issues like RTE/entailment, relevance, long distance coreference, and so on. Separate the hype from the truth and the fact is that there are tons of tasks that will never be automated, like screenwriting or other meaningful communication.

jacquesm 5y ago
The problem is that driving a vehicle is a dynamic process, where the driver adapts to the situation in real time relaxing some rules while tightening others on the fly as best fits the situation. That is a very bad match for the current state of the art in machine learning, where we more or less assume that after a while we've seen all the cases that the world can throw at us at a high enough frequency that we can say that our solution 'works'. People are smart, very smart compared to our smartest computers and forgetting that we apply this all the time and yet have a fairly high error rate (called 'accidents') is what should count as proof that driving isn't a simple problem that be handled by classifiers and rulesets. It may well require much more than that. The current state of machine learning is an excellent match for problems where brute forcing isn't an option but we have a good general idea of the problem space and repetitive tasks that require some rudimentary level of intelligence. Full autonomous driving (L4, L5) under the same circumstances where ordinary drivers are required to perform are for now out of reach of our tools and solutions. By the way: one way in which we could quite easily improve on the number of traffic deaths is stricter certification, faster impounding of vehicles based on previous offenses (for instance: DUI, reckless driving), loss of insurance when continuing to drive in dangerous situations (fog, heavy rain, snow).

by bpiche 5y ago

jacquesm 5y ago

The problem is that driving a vehicle is a dynamic process, where the driver adapts to the situation in real time relaxing some rules while tightening others on the fly as best fits the situation. That is a very bad match for the current state of the art in machine learning, where we more or less assume that after a while we've seen all the cases that the world can throw at us at a high enough frequency that we can say that our solution 'works'. People are smart, very smart compared to our smartest computers and forgetting that we apply this all the time and yet have a fairly high error rate (called 'accidents') is what should count as proof that driving isn't a simple problem that be handled by classifiers and rulesets. It may well require much more than that. The current state of machine learning is an excellent match for problems where brute forcing isn't an option but we have a good general idea of the problem space and repetitive tasks that require some rudimentary level of intelligence. Full autonomous driving (L4, L5) under the same circumstances where ordinary drivers are required to perform are for now out of reach of our tools and solutions. By the way: one way in which we could quite easily improve on the number of traffic deaths is stricter certification, faster impounding of vehicles based on previous offenses (for instance: DUI, reckless driving), loss of insurance when continuing to drive in dangerous situations (fog, heavy rain, snow).