Claude Opus 4.5 did so well on the METR task length graph they’re going to need longer tasks, and we still haven’t scored Gemini 3 Pro or GPT-5.2-Codex.
> People keep claiming AI doesn’t work largely because so often their self-conceptions, futures and future plans, jobs and peace of mind depend on AI not working. They latch onto every potential justification for this, no matter how flimsy, overstated or disproven.
What do you expect people to do instead? The problem is that unless you are in a position of power, you cannot influence anything directly. Contributing towards a negative sentiment towards AI is about the most an average person can do. This is the problem with the people in the field. They expect everyone to catch up to their level of anxiety without providing any tools for doing so safely. People *will* fall into anxious thought patterns and apathetic lifestyles. It's like a professional swimmer egging a child on to follow him into deep waters.
I feel this a lot. I have no skills and no degree, but the AI safety egregore has gotten a hold of me. What am I even supposed to do? How could I hope to skill up in time to ever actually contribute? I’m trying to figure this out but it’s hard to not lose to despair.
I feel you. I don't think there is a tangible way we can contribute as individuals, but the general population is still an important factor. The way I see it, the most prudent thing we can do, as individuals, are:
- Keep ourselves informed about the developments. It's impossible even for experts to keep up with all the developments, so it's good enough to have a general idea of the where the field is.
- Be prepared to vote on regulatory and de-escalatory policies, when given an opportunity.
- Find/build consistent communities around these issues (ideally, irl). We interact with people a lot online, but we rarely get to know them or have any consistent contact with them. Turning these ephemeral acquaintanceships into more robust ones is good for both mental health and collective action. Teach someone about the risk models, learn alongside someone about the technical aspects, etc.
- Try to develop equanimity. Despite our efforts towards and hopes for the contrary, we should be emotionally prepared for a disastrous outcome. I don't think this is trivial. Dealing with death anxiety on the level of species is no easy task. But it will help immensely in anything we might do to combat existential threats.
I'm sorry that this isn't a very novel piece of advice -- I don't think there is any hidden insight to be found here. It's nothing groundbreaking, but I personally feel like it's likely the most efficient course of actions for someone on the outside, as it were. I am certainly going to try to follow this however I can.
The whole "humans aren't generally intelligent" thing is so stupid. The correct response, of course, is not to argue that they are, but to say "Now you're just arguing definitions. Normally by 'general intelligence' we mean the capability that humans have but other animals don't, and AIs currently don't. If you want to give the term a different meaning and call this capability something else, then OK, but that doesn't change the fact that this capability that distinguishes us from other animals exists; we have it; someday AIs may also have it; moreover that someday they have a much more effective version of it than us; and that that scenario is very dangerous than us."
And I mean really, if his only objection is that human intelligence is specialized to the physical world -- what does he expect AIs to operate in?? An AI that was much smarter than us, in a matter specialized to the physical world, would be dangerous for all the same reasons that it would be without that qualifier!
But, wait, poll chart 4 shows overall majority support for an EO? And even among Harris voters, it's pretty close? Granted, the question is worded in a highly specific leading way, and you have to be In The Know to know that the actual mechanism of EO-ing these protections into place doesn't actually work...but this is still an odd definition of "deeply, deeply unpopular". I guess it needs to be amended to "deeply unpopular...IN HARRIS VOTERS"? 80% favourability among the base is nothing to sneeze at! All the more reason politicizing the AI issue is a lose-lose course of action, and I'm saddened to see such efforts starting to bear potential fruit. To paraphrase a local bumper sticker, when the jackboot of robot tyranny is on your neck, it doesn't much matter whether it's on the left or right actuator.
Also, I sincerely hope that rogue Anthropic trader has read Scott's The Onion Knight. Sounds like a fun place to work, ending-the-world concerns aside. It'd be a funny scenario where the only jobs humans maintain a monopoly on are customer service, since it takes an incorrigible to know an incorrigible.
An interesting post here suggesting that while RL model performance may continue to rise exponentially, their inference costs may also rise exponentially. In which case, AI capabilities may outstrip human professionals, but not at a cost that is competitive with human professionals.
> People keep claiming AI doesn’t work largely because so often their self-conceptions, futures and future plans, jobs and peace of mind depend on AI not working. They latch onto every potential justification for this, no matter how flimsy, overstated or disproven.
What do you expect people to do instead? The problem is that unless you are in a position of power, you cannot influence anything directly. Contributing towards a negative sentiment towards AI is about the most an average person can do. This is the problem with the people in the field. They expect everyone to catch up to their level of anxiety without providing any tools for doing so safely. People *will* fall into anxious thought patterns and apathetic lifestyles. It's like a professional swimmer egging a child on to follow him into deep waters.
I feel this a lot. I have no skills and no degree, but the AI safety egregore has gotten a hold of me. What am I even supposed to do? How could I hope to skill up in time to ever actually contribute? I’m trying to figure this out but it’s hard to not lose to despair.
I feel you. I don't think there is a tangible way we can contribute as individuals, but the general population is still an important factor. The way I see it, the most prudent thing we can do, as individuals, are:
- Keep ourselves informed about the developments. It's impossible even for experts to keep up with all the developments, so it's good enough to have a general idea of the where the field is.
- Be prepared to vote on regulatory and de-escalatory policies, when given an opportunity.
- Find/build consistent communities around these issues (ideally, irl). We interact with people a lot online, but we rarely get to know them or have any consistent contact with them. Turning these ephemeral acquaintanceships into more robust ones is good for both mental health and collective action. Teach someone about the risk models, learn alongside someone about the technical aspects, etc.
- Try to develop equanimity. Despite our efforts towards and hopes for the contrary, we should be emotionally prepared for a disastrous outcome. I don't think this is trivial. Dealing with death anxiety on the level of species is no easy task. But it will help immensely in anything we might do to combat existential threats.
I'm sorry that this isn't a very novel piece of advice -- I don't think there is any hidden insight to be found here. It's nothing groundbreaking, but I personally feel like it's likely the most efficient course of actions for someone on the outside, as it were. I am certainly going to try to follow this however I can.
Christmas present for you all, this episode as a podcast!
https://open.substack.com/pub/dwatvpodcast/p/ai-148-christmas-break
The whole "humans aren't generally intelligent" thing is so stupid. The correct response, of course, is not to argue that they are, but to say "Now you're just arguing definitions. Normally by 'general intelligence' we mean the capability that humans have but other animals don't, and AIs currently don't. If you want to give the term a different meaning and call this capability something else, then OK, but that doesn't change the fact that this capability that distinguishes us from other animals exists; we have it; someday AIs may also have it; moreover that someday they have a much more effective version of it than us; and that that scenario is very dangerous than us."
And I mean really, if his only objection is that human intelligence is specialized to the physical world -- what does he expect AIs to operate in?? An AI that was much smarter than us, in a matter specialized to the physical world, would be dangerous for all the same reasons that it would be without that qualifier!
What does LKD mean? Least Common Denominator? Is it Magic jargon?
LordKingDude (the OP of the quoted block) is smart enough to write the code a human can write, but that's not the code an ASI would (bother to) write.
But, wait, poll chart 4 shows overall majority support for an EO? And even among Harris voters, it's pretty close? Granted, the question is worded in a highly specific leading way, and you have to be In The Know to know that the actual mechanism of EO-ing these protections into place doesn't actually work...but this is still an odd definition of "deeply, deeply unpopular". I guess it needs to be amended to "deeply unpopular...IN HARRIS VOTERS"? 80% favourability among the base is nothing to sneeze at! All the more reason politicizing the AI issue is a lose-lose course of action, and I'm saddened to see such efforts starting to bear potential fruit. To paraphrase a local bumper sticker, when the jackboot of robot tyranny is on your neck, it doesn't much matter whether it's on the left or right actuator.
Also, I sincerely hope that rogue Anthropic trader has read Scott's The Onion Knight. Sounds like a fun place to work, ending-the-world concerns aside. It'd be a funny scenario where the only jobs humans maintain a monopoly on are customer service, since it takes an incorrigible to know an incorrigible.
An interesting post here suggesting that while RL model performance may continue to rise exponentially, their inference costs may also rise exponentially. In which case, AI capabilities may outstrip human professionals, but not at a cost that is competitive with human professionals.
https://x.com/tobyordoxford/status/2003203910245486848
What are the alignment implications of Anthropic going public?