In the absence of stated benchmarks it's hard for me to understand what Nathan Lambert means by "superhuman." A calculator can significantly outperform me if I have to do math in my head and even if I have access to pencil and paper. How should I interpret phrases like "superhuman" and "peak performance."
I still get all my statements and tax forms in the mail. They are constantly asking me to switch to paperless, but so far haven't forced me to. I expect they will start forcing me to before much longer...
We have the same mental model for the chain-of-thought AIs. That is, when you ask a question, a basic AI (say GPT4) runs a whole bunch of times down several sequences, evaluating the end of each sequence, ending the failures and running farther down the successes.
In effect, instead of a question using one query of the AI, now a question can use 400 queries, or 5000 queries or as many queries as you have the budget for to run down the forking paths.
In this world, there are no 'more powerful' AIs that need $100B data centers. Rather, there are more efficient normal AIs that deliver by running dozens, hundreds or thousands of times on a given task.
This is also similar to how the human brain seems to function. 70% of the human brain is neocortex. Neo cortex is billions (trillions?) of identical neocortical columns, each a tightly wired stack of neurons less tightly connected to other neocortex stacks.
You mention 2FA, which raises an interesting security problem that many AI companies and organizations (such as my own) which want to implement AI tech have already started to run into and for which there is, as I understand it, not yet any satisfying answer (indeed, perhaps as the inherent trade-off between speed and ease of use on the one hand and security on the other, no possible 'answer'.)
The problem is that you are going to want to have the AI do a lot of things for you, and have access to most of your personal info, and be able to do a large number of these accesses and actions in rapid sequence on your behalf quickly and seamlessly. But that, current annoying best practice - which has come about by combination of regulation and lessons in major security failures learned the hard way - is for -each- of these steps to introduce as much friction, time-consumption, and redundancy in verification as the user will tolerate, and specifically "anti-bot" techniques like CAPTCHA and so forth.
But now instead of fearing an adversary or impersonator bot, you are paying for a bot to work for you. If that bot agent has to stop every 3 seconds to get some permission or verification or chip-card insertion or text message code or smartphone app login password or answer to a security question or RSA token code or CAPTCHA solution from it's human master, all it's really useful use-cases will be neutralized in attractiveness by all that annoyance and delay.
You can't both require proof some activity is being done by a human, but also have the human be able to get the activity done for them by a non-human.
So, we already know what people (users and developers alike) are going to do in the face of all that. They will just immediately try to permanently disable as many of those security measures as possible and give the keys to the castle to their AI agents. And the opportunities for all manner of mischief will explode. Or else they will be forced to put up with it, in which case they won't bother, which could very plausibly create a serious impediment and obstacle to widespread adoption of the most valuable potential uses of cutting edge AI agent tech, which might slow profit growth in that field to disincentivize rapid innovation, or at the extreme scenario, to nip these innovations in the bud.
Siri is gonna evolve into a useful local human interface & verification with phone number so that you tell Siri what you want, and Siri contacts the working agents, with Siri doing the waiting and verifying.
The work is being done under management by the human phone owner, but phone & ai-Agents do the work. Everything one can now do remotely, digitally. All that work.
We see AI everywhere but in the productivity statistics.
In the absence of stated benchmarks it's hard for me to understand what Nathan Lambert means by "superhuman." A calculator can significantly outperform me if I have to do math in my head and even if I have access to pencil and paper. How should I interpret phrases like "superhuman" and "peak performance."
I still get all my statements and tax forms in the mail. They are constantly asking me to switch to paperless, but so far haven't forced me to. I expect they will start forcing me to before much longer...
If only there was an organization that had all of that information at hand and they could just tell us how much we owed ...
We have the same mental model for the chain-of-thought AIs. That is, when you ask a question, a basic AI (say GPT4) runs a whole bunch of times down several sequences, evaluating the end of each sequence, ending the failures and running farther down the successes.
In effect, instead of a question using one query of the AI, now a question can use 400 queries, or 5000 queries or as many queries as you have the budget for to run down the forking paths.
In this world, there are no 'more powerful' AIs that need $100B data centers. Rather, there are more efficient normal AIs that deliver by running dozens, hundreds or thousands of times on a given task.
This is also similar to how the human brain seems to function. 70% of the human brain is neocortex. Neo cortex is billions (trillions?) of identical neocortical columns, each a tightly wired stack of neurons less tightly connected to other neocortex stacks.
You mention 2FA, which raises an interesting security problem that many AI companies and organizations (such as my own) which want to implement AI tech have already started to run into and for which there is, as I understand it, not yet any satisfying answer (indeed, perhaps as the inherent trade-off between speed and ease of use on the one hand and security on the other, no possible 'answer'.)
The problem is that you are going to want to have the AI do a lot of things for you, and have access to most of your personal info, and be able to do a large number of these accesses and actions in rapid sequence on your behalf quickly and seamlessly. But that, current annoying best practice - which has come about by combination of regulation and lessons in major security failures learned the hard way - is for -each- of these steps to introduce as much friction, time-consumption, and redundancy in verification as the user will tolerate, and specifically "anti-bot" techniques like CAPTCHA and so forth.
But now instead of fearing an adversary or impersonator bot, you are paying for a bot to work for you. If that bot agent has to stop every 3 seconds to get some permission or verification or chip-card insertion or text message code or smartphone app login password or answer to a security question or RSA token code or CAPTCHA solution from it's human master, all it's really useful use-cases will be neutralized in attractiveness by all that annoyance and delay.
You can't both require proof some activity is being done by a human, but also have the human be able to get the activity done for them by a non-human.
So, we already know what people (users and developers alike) are going to do in the face of all that. They will just immediately try to permanently disable as many of those security measures as possible and give the keys to the castle to their AI agents. And the opportunities for all manner of mischief will explode. Or else they will be forced to put up with it, in which case they won't bother, which could very plausibly create a serious impediment and obstacle to widespread adoption of the most valuable potential uses of cutting edge AI agent tech, which might slow profit growth in that field to disincentivize rapid innovation, or at the extreme scenario, to nip these innovations in the bud.
Siri is gonna evolve into a useful local human interface & verification with phone number so that you tell Siri what you want, and Siri contacts the working agents, with Siri doing the waiting and verifying.
The work is being done under management by the human phone owner, but phone & ai-Agents do the work. Everything one can now do remotely, digitally. All that work.
This automation sounds swell. Until it goes wrong and no one has a clue what happened or how to fix it.