The first source link is empty.
One bottle of water per 100-word AI prompt seems to be extremely high. I don’t know what the source is. Claude says it is https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/, but it is paywalled. I am curious to see the precise claim.
This is an interesting idea, which I haven’t considered.
Do you allow the agent to access the Internet from the container? If so, isn’t there a risk that it will get prompt injected and leaks your code?
I do enjoy the agent having access to some of my files or search stuff on my computer for me. I suspect I would need to give the agent access to them from inside the container, but then I am worried that it would leak it via (1).
This is an interesting idea, which I haven’t considered.
Do you allow the agent to access the Internet from the container? If so, isn’t there a risk that it will get prompt injected and leaks your code?
I do enjoy the agent having access to some of my files or search stuff on my computer for me. I suspect I would need to give the agent access to them from inside the container, but then I am worried that it would leak it via (1).
I would be enthusiastic about it for a couple of hours, and then, come evening, I would start to forget. Tomorrow morning I would see a reminder on my calendar and feel another rush of excitement – but only for a short while.
I am wondering whether the workout itself is susceptible to this loss of enthusiasm too. How long have you been doing the workout for? How do you decide when to trigger it?
I am trying to resolve the paradox around Claude Code. On the one hand it definitely speeds up many smallish tasks—it often implements stuff in 1-2 minutes, which would normally take me 30-60 min. On the other hand, I don’t see this speed up in the macro sense. So far my best hypothesis is Jevons Paradox. Without Claude Code I knew that work takes me substantial time, so I was incentivized to prioritize. As a result I was doing just the core required work. With Claude Code any small idea I have feels cheap—I can just tell Claude to do it, but it ends up not being cheap. First, it often does require multiple iterations and I ultimately have to review the result. Second, it adds complexity to the system, making AI agents less efficient long-term; it also increases the surface area of something going wrong. Third, it increases context switching.
How does one distinguish “idea is a flop and should be killed” from “idea is good, under-executed and more effort would make it work”?
I really enjoyed this and read it all in one go.
Reading the mechanism behind the contract verification made me realize that there is no way I can imagine what a sufficiently advanced AI will be doing.
It is also interesting to see analogy to humans—we are like these agents, but with the upper bound for number of credits (max lifespan) and, if you don’t bring value, you die earlier.
I also suspect in real life scenario humans would have penalized scams in such a system, since that does not advance the goal.
What is the sample size?
Thank you for sharing!
I cannot understand how exactly the game was played. Would players just write down what AI accomplishes into the spreadsheet to simulate AI working on a task? If so, how do you simulate 200 hour time horizon? E.g. how do you decide that this piece of work done by AI can be done by 200 hour time horizon AI and this one cannot?
Regarding predictions to loosen the bottleneck:
Agent’s best-guess about what comments you’d get from Beth, Hjalmar, Ajeya. Agent’s best-guess about survey results, if you launched the survey. Agent’s best-guess about how this will be received on Twitter.
These just feel like basic sanity checking. I doubt AI would be very good at such predictions. Is this the intention? Or am I missing something?
Your quote is actually about a specific set of products. The pdf mentions:
L’Oreal Visible Lift Serum Absolute—liquid foundation, 1 fl oz, ~$13
Maybelline Superstay Better Skin Skin-Transforming Foundation—liquid foundation, 1 fl oz, ~$13
L’Oreal Age Perfect Eye Renewal Eye Cream—eye cream, 0.5 fl oz, ~$15-20
L’Oreal Revitalift Bright Reveal Brightening Day Moisturizer—day
moisturizer, 1 fl oz
They all are very small volume & have somewhat unusual packaging. E.g. this seems to be #3 - https://www.amazon.com/LOreal-Paris-Anti-Aging-Antioxidants-Dermatologist/dp/B098QZL96Z
I don’t think 12 oz lotion packing is so inefficient.
Walmart does not even use uniform units in the denominators across products of the same kind—it is a mixture of oz, lbs, fl oz, sometimes l, sometimes count. So it would be a great start to just require uniform units to make it all easily comparable at a glance.
Is this tiredness periodic? I experience something similar and my very weak suspicion is that it is non-24-hour sleep–wake disorder. Basically I start feeling tired even in the morning, but I am able to stay up all night without any problem. Going to bed late for a couple days seems to resolve it. Unfortunately I don’t know a good way to verify my hypothesis—I thought about measuring my base temperature to see how the peak shifts, but the device costs $300 and I have such tiredness only a couple times a year, so I didn’t bother.
Domain: Other Lists like This
Link: Map of Reddit (warning: pressing enter does not work in the search box, you have to click on a suggested subreddit in the dropdown)
Author(s): Andriy Kashcha
Type: Interactive Chart
Why: Groups Reddit’s subreddits into categories & shows subreddits related to a given one.
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 was a discussion of this on Reddit one week ago.
They have more examples.
The weak hypothesis was that there is a mismatch between request/response, i.e. people get responses to other people requests. This definitely disagrees with you being able to reproduce this, it sounds implausible that so many people ask about terrorist organizations all the time and overall would have been a huge security risk for OpenAI users.
If you were to ask a software engineer doing something for help and ask 10 technical questions on topic and then start asking off-topic, that would piss them off too, because that’s wasting their time. I am wondering whether Claude picked this up from the training data.