I agree that having such a slider could be good, but I think it should only impact visibility of comments in that post’s comments section, and shouldn’t impact karma (only quality-axis votes should impact karma even if the slider is set to give maximum visibility to high-‘agree’ comments).
MikkW
Karma: 1,054
(I’m choosing not to respond to this comment for reasons relating to potential info-hazards)
Regarding the “a green stop sign in a field of red flowers” images, this does not at all make me update downwards on the capabilities of a maximally capable AI using modern techniques and equipment.
This is because this failure mode is a symptom of how CLIP (one of the components of DALL-E 2) is trained, and would likely be fixed by a small tweak to the training process of CLIP.
CLIP is trained by showing the model an image and a piece of text, and asking if the image and the text are a good fit. CLIP ouputs a list of many (hundreds of) numbers, describing different attributes, for both the text and the image. If the text is actually connected to the image, then the model is “rewarded” for making the same attributes big (or small) for both the text and the image; but if the text belongs to a different image, it is “punished” for making the numbers similar.
Importantly, there are many different types of images in the training set, with all kinds of different captions belonging to them. This means that almost always, if CLIP is being shown an image that does not fit the text, the image and the text are wildly different from eachother. It’s more “a drawing of a great white shark” vs “a green stop sign against a field of red flowers”, rather than “green sign, red flowers” vs “red sign, green flowers”. (Note, there was no effort to try to “trick” CLIP by grouping similar images together in the training, to help it learn to distinguish them)
So CLIP knows the difference between “a field of red flowers” and “a cartoon of a dancing coyote”, but barely knows the difference between “a red stop sign against a green background” and “a green stop sign against a red background”. It vaguely knows the difference, but only barely, because it basically never had to tell such similar inputs apart.
So, 99.9% (I’m estimating that) of the time CLIP saw a training example where the text and image were at least as similar as “a green stop sign in a field of red flowers” and the worst match of the images seen above, they actually were a good match for eachother. So when DALL-E 2 draws a red stop sign, and asks CLIP if it got it right, CLIP just shrugs its shoulders, and says, “yeah, I guess”, so DALL-E feels happy returning that image (note: DALL-E doesn’t actually work like that).
But it’d be incredibly easy to train CLIP in such a way that it has to tell the difference between similar images much more often, and if that happened, I would expect this failure mode to immediately go away. So this example tells us nothing when trying to forecast the capabilities of future models.
Some thoughts, I make no claim as to their accuracy:
In my experience, drinking too much coffee can give me a headache that makes me even more useless than if I had no coffee, this is a relatively easy magnitude to reach without trying.
I find it helpful to take a week every now and then where I use no caffeine at all, to help prevent my system from developing a tolerance to caffeine.
There exist pills that release caffeine over an extended period, for example 6 hours.
The first paragraph after the heading “Mathematical Formulation of the Problem” is hard for me to parse, and I cannot rule out that it contains (at least) a typo.
In the case of OpenAI, the server-side-only constraint, IIRC, is intentional, to prevent people from modifying the model, for AI safety reasons. My understanding is that usually running a model isn’t as compute-intensive as training it in the first place, so I’d expect a user-side application to be viable; just not in line with OpenAI’s modus operandi.
When you zoom in on an exponential curve close enough, it looks linear
Somehow I stumbled across this quote from Deutoronomy (from the Torah / Old Testament, which is the law of religious-Jews):
You shall not have in your bag two kinds of weights, large and small. You shall not have in your house two kinds of measures, large and small. You shall have only a full and honest weight; you shall have only a full and honest measure, so that your days may be long in the land that the Lord your God is giving you. For all who do such things, all who act dishonestly, are abhorrent to the Lord your God.
There’s of course the bit about not being dishonest, but that’s not what resonated with me. I will provide one example of what came to my mind:
A dollar in the year 2023 is a smaller measure than it was in the year 2022, which is a smaller measure than it was in 1996
I hope one day we can achieve a world where autocracy doesn’t have a place in it
If the information doesn’t get to the forecasters, a typical person shouldn’t expect to be able to obtain that data themselves, so the forecast can still be your best option. If you do have information that others don’t, you can always adjust the forecasts you see accordingly.
Why was regularity / non-anonymity a factor that contributed to this format’s viability?
Whenever I hear about historical coffeehouses (say in the period from the late 1700s to the early 1900s), they seem like places where a lot more was going on than in the coffeehouses I’ve been to in my lifetime. Is there any difference? If there was, what changed?
I think there’s a little too much focus on the concept of AGI, an “Artificial General Intelligence”, that is, an AI that can do basically anything that a human can. However, I feel it’s better to focus simply on the fact that the number of things that AI can accomplish well are rapidly increasing, and the list of things that can only be done by humans is getting shorter and shorter. Most of the implications of AGI are also implied by this trend, and it is cleaner to reason about the pace of this trend than about when precisely AGI will come (I hear many people say they think AGI is several decades off. If we worry about edge cases, this is perhaps a reasonable belief. However, the implications of AGI will likely be implications of this trend within the next 2 or 3 decades, if they don’t come even sooner, regardless of if AGI is achieved by that time).
As far as source, I was going off my memory from what an instructor had said when I was taking a culinary arts class several years ago, so nothing I could easily point you to. The original statement as I recall it was that microwaving meat can break down the proteins, and therefore it should be cooked instead of microwaved… but, a quick google doesn’t find anything along these lines, and instead shows many people saying microwaving meat is perfectly fine… so consider that a strong update against my original statement.
It’s possible my memory completely made that statement up, or that the teacher was saying something not widely held to be the case, among other possibilities.
For example, on Newcomb’s Problem, rational irrationality chooses to take only one box (despite the fact that many methods of ‘rational’ analysis would suggest it receives more by taking two), thereby earning $1,000,000; while irrational rationality chooses to take both boxes, and finds itself with only $1,000.
Rational irrationality chooses to commit to a seemingly “irrational” course of action, but actually wins. Irrational rationality tries to use rational analysis, but by doing so forfeits utility.
I think I remember hearing that microwaves break down / degrade proteins, which there are plenty of in cheese. This shouldn’t change the nutritional value of the dish, but may change the eating experience. Microwaving only the pasta and tomatoes, then adding the cheese before baking, may alleviate any problems this causes.
Not just in the latest TCEC season, they’ve been neck-and-neck for quite a bit now
I’d say that this is a better argument for calculus and PDEs than trigonometry- the sine function can be defined purely from a calculus point-of-view, and that definition is more similar to what you describe than the trigonometry perspective
Today, I learned that Vite Ramen is making a version of their healthy & convenient noodles that comes in a quick cup. If you haven’t tried Vite before, I recommend checking it out.
I strongly suspect that the meme that “AI safety (/ EA in general) is not funding constrained” is false