Agree with the point on AI Alignment being insufficient. Both Human-AI Alignment and Human-Human Alignment will need to be successful for a positive future.
If Mechanize succeeds in automating the economy, one unintended consequence might be that people start to take the notion of powerful AI seriously (much like how tariffs have been a sobering for some).
Yes, I agree. I think that one of the big 'warning shots' for powerful AI might come from signs of AI having a major economic impact, so companies like Mechanize could generate that evidence. (I think it is a genuinely difficult question as to whether this is the best way to create the warning shot, but it's certainly an option on the table.)
It’s very odd that mechanize comes from epoch, and from epochs public communication, some of them have AGI timelines out to 2045, so mechanize seems like a very early idea relative to simply waiting for better AI by their own lights. I’m confused that they are the group starting a full automation company
(note, I haven't spoken directly with the Mechanize founders, so could be wrong) - one inference might be that the Mechanize people have shorter timelines than others at Epoch, or think that some types of automation don't require quite as sophisticated capabilities as other people think.
Experience-based learning is plausibly why humans can learn so well from so little data. By interacting with your environment you in effect choose your own training set. This is in marked contrast to the big models of today where we have to throw "everything" into training and end up dramatically over-sampling.
I see this as most applicable to domains like motion and sensory processing where existing AIs are pretty bad and an experience-based approach could deliver immediate value.
Wonder about your thoughts regarding the economic and policy implications on electricity use for training and inference. There are clear economic benefits for AI use, and I wonder if you're studying consumer substitution between different models (for different tasks) when considering the per token cost and underlying electricity costs (dynamic pricing and routing API calls based on electricity prices and fuel mix possible?). I'm a econ postdoc over at Vanderbilt working trying to research this topic, considering demand side and supply side optimization, if you're interested in discussing more.
Agree with the point on AI Alignment being insufficient. Both Human-AI Alignment and Human-Human Alignment will need to be successful for a positive future.
Yes, figuring out how we as humans relate to aligned AI systems is one of the huge challenges ahead of us.
If Mechanize succeeds in automating the economy, one unintended consequence might be that people start to take the notion of powerful AI seriously (much like how tariffs have been a sobering for some).
Yes, I agree. I think that one of the big 'warning shots' for powerful AI might come from signs of AI having a major economic impact, so companies like Mechanize could generate that evidence. (I think it is a genuinely difficult question as to whether this is the best way to create the warning shot, but it's certainly an option on the table.)
It’s very odd that mechanize comes from epoch, and from epochs public communication, some of them have AGI timelines out to 2045, so mechanize seems like a very early idea relative to simply waiting for better AI by their own lights. I’m confused that they are the group starting a full automation company
(note, I haven't spoken directly with the Mechanize founders, so could be wrong) - one inference might be that the Mechanize people have shorter timelines than others at Epoch, or think that some types of automation don't require quite as sophisticated capabilities as other people think.
Experience-based learning is plausibly why humans can learn so well from so little data. By interacting with your environment you in effect choose your own training set. This is in marked contrast to the big models of today where we have to throw "everything" into training and end up dramatically over-sampling.
I see this as most applicable to domains like motion and sensory processing where existing AIs are pretty bad and an experience-based approach could deliver immediate value.
Wonder about your thoughts regarding the economic and policy implications on electricity use for training and inference. There are clear economic benefits for AI use, and I wonder if you're studying consumer substitution between different models (for different tasks) when considering the per token cost and underlying electricity costs (dynamic pricing and routing API calls based on electricity prices and fuel mix possible?). I'm a econ postdoc over at Vanderbilt working trying to research this topic, considering demand side and supply side optimization, if you're interested in discussing more.