There is a tradeoff:
100% chance that I, the writer, will have to look up an archive link
some chance that some unknown number of future readers will have to look up an archive link
The tradeoff depends on the numbers. My guess is it’s worth it for evergreen articles and not worth it for comments.
I agree, you shouldn’t pretend not to be excited if you are.
Adrenaline junkies should not be involved in building AGI, any more than they should be commercial pilots or bus drivers. (Less, even.)
Is the only way to get o1 Pro, which was the highest-compute reasoning variant. Currently I think it’s considered superseded in capabilities by o3, available at $20/month.
It sounds like you’re saying o3 is better than o1 Pro, so there’s no reason to pay extra for o1 Pro?
(I am not the first person to observe that OpenAI’s naming scheme is terrible)
So, pushing back 1 year is only worth it if we reduce P(doom) by 1%.
Only if you don’t care at all about people who aren’t yet born. I’m assuming that’s your position, but you didn’t state it as one of your two assumptions and I think it’s an important one.
The answer also changes if you believe nonhumans are moral patients, but it’s not clear which direction it changes.
Do $100–200/month LLM plan get access to smarter models than $10–20/month plans? Or do they only get higher query limits + faster generation?
AI companies’ marketing materials almost seem to be optimized for being as confusing as possible. I’ve read through them and I cannot tell whether the expensive tiers give access to better models.
I’m trying to decide if it’s worth it for me to buy an expensive plan. I like Deep Research but I don’t do enough queries to hit the limit on a cheap plan.
Paul Graham called it schlep blindness. He was talking in the context of startups but it applies more broadly.
In the version of this mental motion I’m proposing here, you keep your eye out for ways that everyone’s being totally inept and incompetent, ways that maybe you could just do the job correctly if you reached in there and mucked around yourself.
I would go even further than that—you don’t have to think you can do the job correctly. Some of my most fruitful projects started with me thinking something like
Everyone is dropping the ball on this. I really don’t think I can do the job correctly, but someone should at least give it a shot, so I might as well do it.
An untested hypothesis:
LLMs are fundamentally text predictors. There are many high-probability replies to “Tell me a funny joke”, so you wouldn’t necessarily expect them all to tell the same one. But perhaps, somewhere in the training data, someone published their conversation with an LLM in which they said “Tell me a funny joke” and it replied with the joke about atoms. Next-gen LLMs learn from this training data that if an LLM is asked to tell a joke, the probability-maximizing answer is that particular joke. So now they all start telling the same joke.
What can ordinary people do to reduce AI risk? People who don’t have expertise in AI research / decision theory / policy / etc.
Some ideas:
Donate to orgs that are working to AI risk (which ones, though?)
Write letters to policy-makers expressing your concerns
Be public about your concerns. Normalize caring about x-risk
I was just thinking about this, and it seems to imply something about AI consciousness so I want to hear if you have any thoughts on this:
If LLM output is the LLM roleplaying an AI assistant, that suggests that anything it says about its own consciousness is not evidence about its consciousness. Because any statement the LLM produces isn’t actually a statement about its own consciousness, it’s a statement about the AI assistant that it’s roleplaying as.
Counterpoint: The LLM is, in a way, roleplaying as itself, so statements about its consciousness might be self-describing.
Thank you! Good to hear your perspective
I was wondering why this comment had a disproportionately large disagree-score but then I saw it was only one vote. Still, that leaves me wondering why that one person disagreed so strongly.
Obviously you’re allowed to vote without commenting, it’s just nice to understand how I might be making a mistake when I make a significant decision. Perhaps there’s a perspective I’m missing.
Update: Having slept on it, I just donated $10,000.
Update: Having now slept on it, I just donated $10,000.
I’m considering donating $10,000 to CAIP. I am not going to decide anything until tomorrow because I want to sleep on it, but I’ll write up my thoughts now.
I wrote about CAIP in 2024 and it was one of my top donation candidates. I haven’t re-assessed the AI safety landscape since then, but my current best guess is that CAIP is the single best use of marginal funds.
My basic thinking:
I believe international treaties + US regulations are the best strategies for keeping AI safe.
Even if they’re not the best thing to do, they’re good and we should be trying to make them happen.
I agree with OP that having good US regulations increases the chance of getting international treaties.
A critical step in getting legislation passed is writing and advocating for model legislation.
To my knowledge, the only orgs that have written legislation are CAIS and CAIP. It’s important that these two orgs continue to exist.
There are a handful of other orgs that do political outreach on AI x-risk, and I like most of them too, but none of them are as funding-constrained as CAIP.
My list of concerns is more wordy than my list of positives, Lest that gives readers the wrong idea, I would like to note that I think the positives are much stronger than the negatives.
Why aren’t more institutional donors funding CAIP?
My best guess is they’re (overly) averse to supporting political action, especially because CAIP is open about its concern for x-risk, and I think many donors are timid about pushing the Overton window. That’s just a guess, I don’t want to claim to be able to read donors’ minds.
There could be some externally-illegible reason why CAIP is actually ineffective or counterproductive. CAIP looks quite competent as far as I can tell, but I know little about politics.
Why didn’t CAIP write this post until they had 30 days of runway left?
But I’d expect they’d try to fundraise from institutional donors first and only write this post after that failed.
If I donate, CAIP might shut down anyway, and then my donation didn’t achieve anything.
This is less of a concern if I donate through Manifund because the donation only goes through if the fundraiser reaches its minimum goal.
The fact that CAIP can’t downscale means it might have continual difficulty raising funding.
I think donations still have high expected value because there’s a small chance that my donation will buy CAIP just enough time to get the fundraising momentum it needs, and then in some sense I can claim 100% credit for everything CAIP does after that.
My current thinking is that I should donate regardless, but I would feel better about it if there were some mechanism (like what Manifund has) where I only pay if CAIP gets enough donation commitments to sustain itself.
I wasn’t planning on donating for another 4–6 months, but this is a special situation so I may adjust my plans.
I may donate more than $10,000, but I’ll need to review my financial situation first.
I can donate $10,000 immediately out of my bank account. I can donate more than that from my DAF, but the recipient must be a 501(c)(3). (So I could donate via Manifund.)
As an aside, I’ve noticed you routinely go out of your way to help charities get funding when they’re in difficult situations, which I think is great for the ecosystem and I’m glad you do it.
How do Manifund donations interact with your 501(c)(4) status? Are donations through Manifund just as good as unrestricted donations, or are you more constrained on how you can use that money?
As I understand it, PauseAI Global aka PauseAI supports protests in most regions, whereas US-based protests are run by PauseAI US which is a separate group of people.
PauseAI US is a separate entity from PauseAI so I believe it should also be listed.
Not OP but I think Functional Threshold Power is fine. I don’t know of any literature directly comparing it to VO2max, but much of the literature on VO2max didn’t actually measure VO2max, it used proxies like “maximum gradient at which a participant can walk for 3 minutes” (called the Balke treadmill test. When meta-analyses report that VO2max strongly predicts health outcomes, what they usually* mean is “VO2max, and also various proxies for VO2max, when thrown together into a meta-analysis, strongly predict health outcomes”. So as far as I can tell from what (little) research I’ve looked at, there are a lot of metrics that work and it’s not clear which ones work better than others. And FTP seems like as good a measure as any.
For example, have a look at Table 2 in Impact of Cardiorespiratory Fitness on All-Cause and Disease-Specific Mortality: Advances Since 2009, which gives a list of studies and what measure each study used. You can see that they used a variety of fitness metrics.
*I’ve only actually looked at two meta-analyses
Your experience may vary but I’ve done 12-week weight loss cycles where I ate no sweets and I never lost my desire to eat sweets. I’m on week 6 of a 6-week weight loss cycle right now, I had pretty strong cravings on week 2–3 and they significantly subsided by week 4 but they’re still there.
I do still eat fruit, which may be enough to maintain my sugar cravings, but if your goal is to improve health then I think it’s a bad idea to cut out fruit. And anyway I don’t get cravings for fruit, I get cravings for artificially-sweetened foods.
I’ve heard at least one person report that they entirely lost their sugar cravings when they stopped eating sugar. So it works for some people it just doesn’t work for me.