I like your hypothesis. I can add one more reason to behave this way—managing expectations. In your example if you make $1 trillion revenue prediction, that would catch attention & the market might incorporate that prediction early. But if then for some reason the real revenue is somewhat below the prediction, the expectation wouldn’t be met.
On the other hand, if you raise money, you likely want your valuation to be as high as possible.
There is likely to be a stage in agents replacing knowledge workers where a model can go away and reliably do a well-defined task that would take a human about a day without getting stuck (or at least note that it has got stuck and ask for help). At that point, agents will be able to do a lot of economically valuable work, but they will also need a lot of human task definition, managing, handholding, helping them get unstuck, and checking their output — and humans who have learnt to be good at doing all that will be at a premium. Shortly after that, the agents will be able to do tasks that would take a human a couple of days, and then a week, then a couple of weeks, and the ratio of human supervision work to AI work will decrease rapidly.
That is to say there are intermediate stages between “AI is useless as anything other than a tool” to “AI can completely replace all knowledge workers”. Given how fast capabilities are growing (METR’s results on this suggest practicable task lengths double every 4-6 months), these intermediate stages will be shortish, but during them applying AI successfully will require a huge amount of change to how companies do business and training a lot of humans in how to hand-hold AI — something that is slow and expensive. So there is likely to be overhang for a while, at least until we get all the way to the “drop-in-replacement knowledge worker” stage. So perhaps OpenAI is being cautious about adoption rates, because they understand that the adoption process will be problematic and uneven.
Also, AI capabilities tend to be surprisingly uneven by human standards. If ¾ of the work can be automated and ¼ can’t, progress rapidly becomes bottlenecked on whatever the AI can’t do. Ditto across industries: if some quickly reach full AI-automation and others don’t, the price of the outputs of the fully-automated ones nosedives to near the cost of the inputs, while conversely the price of whatever hasn’t been automated goes up, as it becomes a scarce resource in the growing economy.
That is to say there are intermediate stages between “AI is useless as anything other than a tool” to “AI can completely replace all knowledge workers”. Given how fast capabilities are growing (METR’s results on this suggest practicable task lengths double every 4-6 months), these intermediate stages will be shortish...
I agree with this, but I expect these intermediate stages to be in the past by 2030. E.g. in “Measuring AI Ability to Complete Long Tasks” METR predicts (see “A sensitivity analysis of the extrapolated date...” graph) that in 2027-2031 AI will be able to perform with 50% success rate tasks taking a human 1 work-month. Even when AI can do tasks equivalent to 1 day of work, there is a huge incentive to use it. If we assume task duration doubling every 7 months, 1 work-day = 8 work-hours should be reached in 2 years.
You understood my question correctly.
I like your hypothesis. I can add one more reason to behave this way—managing expectations. In your example if you make $1 trillion revenue prediction, that would catch attention & the market might incorporate that prediction early. But if then for some reason the real revenue is somewhat below the prediction, the expectation wouldn’t be met.
On the other hand, if you raise money, you likely want your valuation to be as high as possible.
There is likely to be a stage in agents replacing knowledge workers where a model can go away and reliably do a well-defined task that would take a human about a day without getting stuck (or at least note that it has got stuck and ask for help). At that point, agents will be able to do a lot of economically valuable work, but they will also need a lot of human task definition, managing, handholding, helping them get unstuck, and checking their output — and humans who have learnt to be good at doing all that will be at a premium. Shortly after that, the agents will be able to do tasks that would take a human a couple of days, and then a week, then a couple of weeks, and the ratio of human supervision work to AI work will decrease rapidly.
That is to say there are intermediate stages between “AI is useless as anything other than a tool” to “AI can completely replace all knowledge workers”. Given how fast capabilities are growing (METR’s results on this suggest practicable task lengths double every 4-6 months), these intermediate stages will be shortish, but during them applying AI successfully will require a huge amount of change to how companies do business and training a lot of humans in how to hand-hold AI — something that is slow and expensive. So there is likely to be overhang for a while, at least until we get all the way to the “drop-in-replacement knowledge worker” stage. So perhaps OpenAI is being cautious about adoption rates, because they understand that the adoption process will be problematic and uneven.
Also, AI capabilities tend to be surprisingly uneven by human standards. If ¾ of the work can be automated and ¼ can’t, progress rapidly becomes bottlenecked on whatever the AI can’t do. Ditto across industries: if some quickly reach full AI-automation and others don’t, the price of the outputs of the fully-automated ones nosedives to near the cost of the inputs, while conversely the price of whatever hasn’t been automated goes up, as it becomes a scarce resource in the growing economy.
I agree with this, but I expect these intermediate stages to be in the past by 2030. E.g. in “Measuring AI Ability to Complete Long Tasks” METR predicts (see “A sensitivity analysis of the extrapolated date...” graph) that in 2027-2031 AI will be able to perform with 50% success rate tasks taking a human 1 work-month. Even when AI can do tasks equivalent to 1 day of work, there is a huge incentive to use it. If we assume task duration doubling every 7 months, 1 work-day = 8 work-hours should be reached in 2 years.