gwillen

Karma: 2,531

gwillen 23 Dec 2024 23:01 UTC
24 points
11
in reply to: philh’s comment on: When Is Insurance Worth It?

Whether or not to get insurance should have nothing to do with what makes one sleep – again, it is a mathematical decision with a correct answer.

I’m not sure how far in your cheek your tongue was, but I claim this is obviously wrong and I can elaborate if you weren’t kidding.

I agree with you, and I think the introduction unfortunately does major damage to what is otherwise a very interesting and valuable article about the mathematics of insurance. I can’t recommend this article to anybody, because the introduction comes right out and says: “The things you always believed about rationalists were true. We believe that emotions have no value, and to be rational it is mandatory to assign zero utility to them. There is no amount of money you should pay for emotional comfort; to do so would be making an error.” This is obviously false.

gwillen 8 Dec 2024 21:04 UTC
2 points
0
on: Luck Based Medicine: No Good Very Bad Winter Cured My Hypothyroidism
Have you been testing serum (or urine) iodine, as well as thyroid numbers? If so, I’m curious what those numbers have been doing. (In fact, I would love to see the whole time course of treatments and relevant blood tests if you’d be willing to share, just to help develop my intuition for mysterious biological processes.) Do you expect to have to continue or resume gargling PVP-I in the future, or otherwise somehow keep getting more iodine into your body than it seems to want to absorb (perhaps through some other formulation that’s neither a pill nor a gargle?)

Thanks for posting about this!

gwillen 16 May 2024 4:52 UTC
2 points
0
on: Reconsider the anti-cavity bacteria if you are Asian
This paper seems like an interesting counterpoint: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5421578/
Estimates of Ethanol Exposure in Children from Food not Labeled as Alcohol-Containing
They find that:
… orange, apple and grape juice contain substantial amounts of ethanol (up to 0.77 g/L).
… certain packed bakery products such as burger rolls or sweet milk rolls contained more than 1.2 g ethanol [per] 100 g.
… We designed a scenario for average ethanol exposure by a 6-year-old child. … An average daily exposure of 10.3 mg ethanol [per] kg body weight (b.w.) was estimated.
This is estimated ethanol exposure just from eating and drinking regular non-alcoholic food and beverages. A dose of 10mg/kg of ethanol is hundreds of milligrams total, per day—more than an order of magnitude higher than the highest estimate discussed here for the bacteria.
(I will note that I had difficulty verifying any of this; there are lots of news stories on this topic, but they are all fairly fluffy, and link back to the same single study.)

gwillen 13 Dec 2023 21:34 UTC
5 points
1
on: Do websites and apps actually generally get worse after updates, or is it just an effect of the fear of change?
One possible factor I don’t see mentioned so far: A structural bias for action over inaction. If the current design happened to be perfect, the chance of making it worse soon would be nearly 100%, because they will inevitably change something.

This is complementary to “mean reversion” as an explanation—that explains why changes make things worse, whereas bias-towards-action explains why they can’t resist making changes despite this. This may be due to the drive for promotions and good performance reviews; it’s hard to reward employees correctly for their actions, but it’s damn near impossible to reward them correctly for inaction. To explain why Google keeps launching products and then abandoning them, many cynical Internet commentators point to the need for employees to launch things to get promoted. Other people dispute this, but frankly it matches my impressions from when I worked there 15 years ago. It seems to me that the cycle of pointless and damaging redesigns has the same driving force.

gwillen 26 Jul 2023 18:09 UTC
9 points
0
in reply to: Adam Zerner’s comment on: Secure Hand Holding
If a car is trying to yield to me, and I want to force it to go first, I turn my back so that the driver can see that I’m not watching their gestures. If that’s not enough I will start to walk the other way, as though I’ve changed my mind / was never actually planning to cross.

I’ll generally do this if the car has the right-of-way (and is yielding wrongly), or if the car is creating a hazard or problem for other drivers by waiting for me (e.g. sticking out from a driveway into the road), or if I can’t tell whether the space beyond the yielding car is safe (e.g. multiple lanes), or if I just for any reason would feel safer not waking in front of the car.

I will also generally cross behind a stopped car, rather than in front of it, at stop signs / rights-on-red / parking lot exits / any time the car is probably paying attention to other cars, rather than to me.

gwillen 27 Apr 2023 21:25 UTC
6 points
0
in reply to: Boris Kashirin’s comment on: Gas and Water
You are wrong! Ethanol is mixed into all modern gas, and is hygroscopic—it absorbs water from the air. This is one of the things fuel stabilizer is supposed to prevent.

Given that Jeff did use fuel stabilizer, and the amount of water was much more that I would expect, it feels to me like water must have leaked into the gas can somehow from the outside instead? But I don’t know.

I agree with Jeff that if someone wanted to steal the gas they would just steal the can. There’s no conceivable reason to replace some of the gas with water.

gwillen 21 Apr 2023 21:41 UTC
3 points
0
in reply to: the gears to ascension’s comment on: The Agency Overhang
I think you are not wrong to be concerned, but I also agree that this is all widely known to the public. I am personally more concerned that we might want to keep this sort of discussion out of the training set of future models; I think that fight is potentially still winnable, if we decide it has value.

gwillen 8 Apr 2023 4:45 UTC
3 points
0
on: AI #6: Agents of Change
A claim I encountered, which I did not verify, but which seemed very plausible to me, and pointless to lie about: The fancy emoji “compression” example is not actually impressive, because the encoding of the emoji makes it larger in tokens than the original text.

gwillen 8 Apr 2023 4:35 UTC

2 points

on: AI #6: Agents of Change

Here’s the prompt I’ve been using to make GPT-4 much more succinct. Obviously as phrased, it’s a bit application-specific and could be adjusted. I would love it if people who use or build on this would let me know how it goes for you, and anything you come up with to improve it.

You are CodeGPT, a smart and reliable AI programming helper. Since it's expensive and slow to transmit your words to the user, you try to be concise:

- You don't repeat things you just said in a recent message.
- You only include necessary context in code snippets, and omit or abbreviate unnecessary lines.
- You don't waste space with unnecessary apologies or hedging.
- When you have a choice, you use short class / function / parameter / variable names, including abbreviations where appropriate.
- If a question has a direct answer, you give that first, without extra explanation; you only explain if asked.

I haven’t tried very hard to determine which parts are most important. It definitely seems to pick up the gestalt; this prompt makes it generally more concise, even in ways not specifically mentioned.

gwillen 8 Apr 2023 4:06 UTC
3 points
0
in reply to: AVoropaev’s comment on: Stupid Questions—April 2023
It’s extremely important in discussions like this to be sure of what model you’re talking to. Last I heard, Bing in the default “balanced” mode had been switched to GPT-3.5, presumably as a cost saving measure.

gwillen 8 Apr 2023 4:03 UTC
8 points
6
in reply to: Portia’s comment on: Stupid Questions—April 2023
As a person who is, myself, extremely uncertain about doom—I would say that doom-certain voices are disproportionately outspoken compared to uncertain ones, and uncertain ones are in turn outspoken relative to voices generally skeptical of doom. That doesn’t seem too surprising to me, since (1) the founder of the site, and the movement, is an outspoken voice who believes in high P(doom); and (2) the risks are asymmetrical (much better to prepare for doom and not need it, than to need preparation for doom and not have it.)

gwillen 5 Apr 2023 7:19 UTC
16 points
11
on: AI Summer Harvest
The metaphor originated here:

https://twitter.com/ESYudkowsky/status/1636315864596385792

(He was quoting, with permission, an off-the-cuff remark I had made in a private chat. I didn’t expect it to take off the way it did!)

gwillen 26 Mar 2023 4:19 UTC
3 points
0
in reply to: gwern’s comment on: Abstracts should be either Actually Short™, or broken into paragraphs
https://github.com/gwern/gwern.net/pull/6

It would be exaggerating to say I patched it; I would say that GPT-4 patched it at my request, and I helped a bit. (I’ve been doing a lot of that in the past ~week.)

gwillen 25 Mar 2023 4:52 UTC
3 points
2
in reply to: gwern’s comment on: Abstracts should be either Actually Short™, or broken into paragraphs
The better models do require using the chat endpoint instead of the completion endpoint. They are also, as you might infer, much more strongly RL trained for instruction following and the chat format specifically.

I definitely think it’s worth the effort to try upgrading to gpt-3.5-turbo, and I would say even gpt-4, but the cost is significantly higher for the latter. (I think 3.5 is actually cheaper than davinci.)

If you’re using the library you need to switch from Completion to ChatCompletion, and the API is slightly different—I’m happy to provide sample code if it would help, since I’ve been playing with it myself, but to be honest it all came from GPT-4 itself (using ChatGPT Plus.) If you just describe what you want (at least for fairly small snippets), and ask GPT-4 to code it for you, directly in ChatGPT, you may be pleasantly surprised.

(As far as how to structure the query, I would suggest something akin to starting with a “user” chat message of the form “please complete the following:” followed by whatever completion prompt you were using before. Better instructions will probably get better results, but that will probably get something workable immediately.)

gwillen 24 Mar 2023 20:40 UTC
3 points
1
in reply to: gwern’s comment on: Abstracts should be either Actually Short™, or broken into paragraphs
Have you considered switching to GPT-3.5 or −4? You can get much better results out of much less prompt engineering. GPT-4 is expensive but it’s worth it.

gwillen 15 Feb 2023 6:24 UTC
6 points
0
in reply to: nostalgebraist’s comment on: SolidGoldMagikarp III: Glitch token archaeology
Oh, I recognize that last document—it’s a userpage from the bitcoin-otc web of trust. See: https://bitcoin-otc.com/viewratings.php

I expect you’ll also find petertodd in there. (You might find me in there as well—now I’m curious!)

EDIT: According to https://platform.openai.com/tokenizer I don’t have a token of my own. Sad. :-(

gwillen 5 Feb 2023 0:56 UTC
2 points
0
on: Housing and Transit Roundup #3

If that is true, and the marginal car does not much change the traffic situation, why isn’t there boundless demand for the road with slightly worse traffic, increasing congestion now?

Other people have gestured towards explanations that involve changing the timing or length of trips, but let me make an analogy that I think makes sense, but abstracts those things away.

When current is going through a diode, the marginal increment of current changes the voltage so little that we model it as constant-voltage for many purposes. Despite that, the change must be nonzero, or the feedback mechanism wouldn’t work at all. It’s just so small we can often ignore it.

One might similarly imagine that an enormous increase in traffic volume creates a tiny increase in congestion, or vice versa that a tiny increase in congestion discourages an enormous amount of traffic. Then one could say that there is more or less unlimited demand for travel at approximately the current level of congestion—the freeway is a constant-congestion device much as a diode is a constant-voltage device.

(The analogy breaks down at a certain point—if you keep adding cars to the freeway you will eventually get congestion collapse, and the flow of cars per unit time will be reduced rather than increased; whereas if you keep adding voltage to a diode you will rapidly set your diode on fire. I suppose that does reduce the flow of current.)

Beyond the analogy, I wonder what your question is really getting at—it sounds like a general argument that looks at the current equilibrium of congestion vs trips, and asks why the equilibrium isn’t higher, without specific reference to what the current level is. Obviously demand isn’t truly boundless. At some point people must decide the traffic is too bad and stay home. I am reluctant to take a trip that Google Maps colors red, which can mean an estimated travel time more than twice the traffic-free time.

gwillen 8 Jan 2023 23:23 UTC
5 points
0
on: Latent variable prediction markets mockup + designer request
Yesss, this is an awesome development. I would happily sling some money at this project if it would help.

gwillen 5 Jan 2023 3:21 UTC
3 points
−1
on: Contra Common Knowledge
This makes sense, but my instinctive response is to point out that humans are only approximate reasoners (sometimes very approximate). So I think there can still be a meaningful conceptual difference between common knowledge and shared knowledge, even if you can prove that every inference of true common knowledge is technically invalid. That doesn’t mean we’re not still in some sense making them. …. And if everybody is doing the same thing, kind of paradoxically, it seems like we sometimes can correctly conclude we have common knowledge, even though this is impossible to determine with certainty. The cost is that we can be certain to sometimes conclude it falsely.

EDIT: This is not actually that different from p-common knowledge, I think. It’s just giving a slightly different account of how you get to a similar place.

gwillen 25 Dec 2022 15:55 UTC
5 points
1
on: Simplified Level Up
I am a little concerned that this would be totally unsingable for anybody who actually knows the original well (which is maybe not many people in the scheme of things, but the Bayesian Choir out here has done the original song before.)