Get out of our houses before we are driven to expend effort killing them, and similarly for all the other places ants conflict with humans (stinging, eating crops, ..)
Ant mafia: “Lovely house you’ve got there, wouldn’t it be a shame if it got all filled up with ants?”
I can’t tell you whether this is right or not but this is what ChatGPT thinks:
“Is it possible to put a latex table in a footnote?”
“Yes, it is possible to include a LaTeX table in a footnote. You can do this by using the footnote or footnotetext commands provided by the footmisc package.
Here is an example of how to use these commands to include a table in a footnote:
\usepackage{footmisc}
...
\begin{table}[h]
\centering
\begin{tabular}{c c c}
A & B & C \\
1 & 2 & 3 \\
4 & 5 & 6
\end{tabular}
\caption{Table caption}
\end{table}
...
\footnote{
\begin{table}[h]
\centering
\begin{tabular}{c c c}
D & E & F \\
7 & 8 & 9 \\
10 & 11 & 12
\end{tabular}
\caption{Table caption in footnote}
\end{table}
}
Note that when including a table in a footnote, you should use the table environment as you would normally, but place it inside the footnote or footnotetext command. You should also make sure to use the [h] option for the table environment to prevent the table from floating.
I hope this helps! Let me know if you have any other questions.”
I think the article undersells the problems of ChatGPT’s hallucinations. One example from the article where ChatGPT is said to win is a recipe for risotto. However, I wouldn’t follow a risotto recipe for ChatGPT just because I can’t be confident it hasn’t hallucinated some portion of the recipe but would happily follow one from Google, even if the format is a bit more annoying. Same issue with calculating load bearing capacity for a beam only more serious!
Having said that, it does seem like there are definitely specific areas where ChatGPT will be more useful. Coding is a good example as verifying the code is usually straightforward and/or would need to be done anyway. In many cases ChatGPT for an overview followed by Google for more detail/verification is probably a good way to go—I think this would be a good idea for the load bearing wall example.
One thing I’ve found useful is to make sure I identify to the supplier what specifically I need about the product I’m ordering—sometimes they have something similar in stock which meets my requirements.
One thing I think makes a big difference to me is whether I feel like the provider is taking a collaborative or adversarial stance.
I don’t usually skip ads on Youtube content but if the channel is often clickbaity/misrepresenting content then I will
The printer/ink thing feels very out to get me. The alternative model of printer subscription (e.g. hp) feels alot more collaborative so I don’t feel the need to ensure that every page I print is as filled with ink as possible so as to get the “best” deal.
If the premium charged on foods in an amusement park/movie theatre feels excessive then I will feel less bad about eating my own food.
For the six/man thing my first association was six pack. Obviously the prototypical image would be topless but my guess is topless images aren’t in the training set (or Dall-E is otherwise prevented from producing them)
I realised something a few weeks back which I feel like I should have realised a long time ago.
The size of the human brain isn’t the thing which makes us smart, rather it is an indicator that we are smart.
A trebling of brain size vs a chimp is impressive but trebling a neural network’s size doesn’t give that much of an improvement in performance.
A more sensible story is that humans started using their brains more usefully (evolutionarily speaking) so it made sense for us to devote more of our resources to bigger brains for the marginal gains that would give.
As I said, I feel like I should have known this for ages. I had a cached thought that human’s big brains (and other things) cause us to be smarter and had never re-examined the thought. Now I think that the “and other things” is doing almost all of the heavy lifting and the size is more incidental to the process.
Thanks for publishing this. I’ve been around the rationality community for a few years and heard TAPs mentioned positively a lot without knowing much about them. This roughly matches my best guess as to what they were but the extra detail is super useful, especially in the implementation.
This suggests a different question. For non-participants who are given the program which creates the data, what probability/timeframe to assign to success.
On this one I think that I would have put a high probability to be solved but would have anticipated a longer timeframe.
I think the resulting program has lower length (so whatever string it generates has lower KC)
I don’t think this follows—your code is shorter in python but it includes 3 new built in functions which is hidden complexity.
I do agree with the general point that KC isn’t a great measure of difficulty for humans—we are not exactly arbitrary encoders.
What were the noise levels on the Corsi-Rosenthal?
Humans are very reliable agents for tasks which humans are very reliable for.
For most of these examples (arguably all of them) if humans were not reliable at them then the tasks would not exist or would exist in a less stringent form.
Curious as to what the get under the desks alarm was supposed to help with and how long ago this was? I’m having trouble fitting it into my world model.
I see that the standard Playground Q&A prompt on OpenAI uses a similar technique (although boringly uses “Unknown” instead of “Yo be real”).
I think the thing which throws people off is that when GPT-3 goes wrong it goes wrong in ways that are weird to humans.
I wondered if humans sometimes fail at riddles that GPT-3 would think of as weird. I tried a few that I thought would be promising candidates (no prompt other than the questions itself)
Q: If a red house is made with red bricks, a blue house is made with blue bricks, a pink house is made with pink bricks, a black house is made with black bricks, what is a greenhouse made with?
A: Greenhouses are made with glass.
Say silk 4 times
Silk, silk, silk, silk.
What does a cow drink?
A cow drinks water.
You put a roast in a roaster, what do you put in a toaster?
Bread.
These are very cherry picked—I deliberately chose trick riddles that I thought GPT-3 would be good at. I think there’s possibly something here about how humans vs GPT-3 rely on context.
I imagine a society of GPT-4s discovering a human and discussing how hollow their intelligence is.
I think the natural/manmade comparison between COVID and Three Mile has alot of merit but there are other differences which might explain the difference. Some of them would imply that there would be a strong response to an AI , others less so.
Local vs global
To prevent nuclear meltdowns you only need to ban them in the US—it doesn’t matter what you do elsewhere. This is more complicated for pandemic preparedness.
Active spending vs loss of growth
Its easier to pass a law putting in nuclear regulations which limit growth as this isn’t as obvious a loss as spending money from the public purse on measures for pandemics.
Activity of lobbying groups
I get the impression that the anti-nuclear lobby was alot bigger than any pro-pandemic-preparedness lobby. Possibly this is partly caused by the natural vs manmade thing so might be kind of a subpoint.
Tractability of problem
Preventing nuclear disasters seems more tractable than pandemic preparedness
1979 vs 2020
Were our institutions stronger back then?
FWIW I agree that a large AI disaster would cause some strong regulation and international agreements, my concern is more that a small one would not and small ones from weaker AIs seem more likely to happen.
Assuming this is the best an AGI can do, I find this alot less comforting than you appear to. I assume “a very moderate chance” means something like 5-10%?
Having a 5% chance of such a plan working out is insufficient to prevent an AGI from attempting it if the potential reward is large enough and/or they expect they might get turned off anyway.
Given sufficient number of AGIs (something we presumably will have in the world that none have taken over) I would expect multiple attempts so the chance of one of them working becomes high.
There’s a theory of humor called benign violation theory.
The BVT claims that humor occurs when three conditions are satisfied: 1) something threatens one’s sense of how the world “ought to be”, 2) the threatening situation seems benign, and 3) a person sees both interpretations at the same time.
I think your description of pranks etc. fits in nicely with this—you even chose the same words to describe it so maybe you’re already aware?
It’s worth noting that the while number of courses at Berkeley almost doubled in the period shown, the number of courses per student has increased at a lower rate due to an increase in students.
Eyeballing the graph and looking at Berkeley’s enrollment numbers I think the number of courses per student has increased by around 50%. Smaller but still a big effect.
Example:
I have a couple of positions I need to fill at my work. I’ve been off on holiday this week and it occurred to me that I should change one of the roles quite a lot and redistribute work.
I’ve had this issue for a few months and while in work I’ve been a bit overworked to actually take a step back and see this opportunity.
FWIW this matches my own experience with one as a company car pretty exactly.
(On mine (UK, right hand drive) the wipers can be activated manually by pressing the button on the end of the left stalk. This also brings up an on-screen menu for selecting constant wiping)