I have taught Python to high schoolers before and this is frustratingly common. Students will name their files `string.py` or similar and get crashes with esoteric error messages. It's a difficult stumbling block for someone who's just learning the basics
I don’t have any kind of ML background but I have always thought of sample efficiency as the great unsolved problem of AI. We humans have unbelievably good sample efficiency; often we can durably learn something on just a single example or two. This is the main area in which LLMs are vastly, vastly behind us.
the caveat is that we are not learning those small number of samples from scratch, since we're coming in with a large amount of training already, much of it from before we were even born
I've seen a few videos about people having their girlfriend/dad/etc play very simple video games for the very first time. It's a really great example of this. The most basic, simple, mechanics are difficult for them, and take surprising repetition to remember, because they have zero framework around any of it. It's really similar to watching a genetic algorithm learn to play Mario, but on fast forward.
No, the unsolved problem of AI is continuous learning. We never stop learning, we don't have a "training phase". You are always updating your world model even when you sleep. Also more quality training data does lead to greater learning efficiency as you have more priors to work with.
They are not well structured. They are syntactically correct. But, if a person wrote them, you would say they are badly structured. I have seen tripple negatives and sentences that should be split into four.
I agree. The afterimage test was a major red flag for me. To see something in your “mind’s eye” has nothing to do with physical vision, yet the afterimage works by exhausting some of the eye’s physical cones. It’s a confusing non sequitur and totally unrelated to the subject at hand. I have always suspected that people with “aphantasia” simply misunderstand the idea of a “mind’s eye,” believing that others can somehow superimpose something onto their actual field of vision at will. Of course, this is not possible.
Edit: it’s also telling that this website sells an “Aphantasia Membership” for a monthly fee.
> To see something in your “mind’s eye” has nothing to do with physical vision, yet the afterimage works by exhausting some of the eye’s physical cones. It’s a confusing non sequitur and totally unrelated to the subject at hand.
Which they go to the correct length to explain in the link associated with that demonstration. I found that exercise quite apt. They aren't saying "This is what non-aphants see" but rather "This is an approximation of what someone who can see images in their mind sees using a separate process so that you get an idea".
For me, the afterimage effect really DOES look similar to aspects of my own internal image perception (colors that are sort of there, but not quite; hazy quality; transparency).
Maybe you're right, but there are people on this thread claiming that they can actually see things in their field of vision, not "akin to seeing things" but see-see them. I can't do this and I don't think I have aphantasia; I am interested in people who say that they can do this. https://news.ycombinator.com/item?id=49469365
Do the apple test (closed eye imagination and then pick the most-similar image on the scale). It's a validated test. And yes, apparently some people actually see an apple (not me, as a total aphantasic). Whether that means in their eyes or in their brain is pretty irrelevant. The people who claim to see an apple have physically observable differences in reactions to other mental imagery tests.
I am on the other end, high five inverse brother. When I see an apple in my mind, I can take it, rotate it, pull the stem off, look at the blossom scar on the bottom, cut it open, count the seeds, and peel the skin off, crush it into a cup and have juice. I can see the blemishes, if there are any, which side faced the sun, and how shaded it was, and I can tell you what the worm had for breakfast (it was apple. that part isn't fancy)
It uses the visual center of my brain (fMRI would show it according to the literature), but it doesn't use my eyes. I do, however, tend to pay much less attention to what I'm seeing when I'm visualizing intensely, so there's some competitive inhibition.
I used to be very frustrated with people who just couldn't see how the furniture in a room could be rearranged, or the color a wall could be, or where the bolt is on the underside of the engine now that we've installed it. Typical mind fallacy, that was me.
Interestingly, a lot of aphantasics are pretty good at spatial reasoning. It seems surprisingly non-visual.
In retrospect, the earliest sign that my cognition was different was in being asked to recall some event. I can virtually never remember being at a moment in the past. People ask me if I remembered doing some very-important-thing and the honest answer is consistently: No, but I remember the fact that it happened and the fact that I was there.
But there's literally no first-personness to any of my memories.
Completely opposite on that front, here, as well. I can't even really associate, I have visual memory all the way back to my earliest childhood memories (specifically, 3-4 years old, stepping on the track of a sliding door in San Diego, I believe, it hurt, I wanted to go see the pool outside), and my memory is so visual that sometimes I can see people's face, and what they were wearing and where we were standing, in the conversations we had but not necessarily hear what they're saying. I also have some semantic memory, something like "we talked about x for a while, and then I think we did y", but it's much hazier.
Honestly I love the fact that humans are so cognitively diverse and that we all can share our experiences.
After reading your comment I went to see the test. The scale shows the same picture in four different opacities, which isn't useful. The textual descriptions were more useful.
> I have always suspected that people with “aphantasia” simply misunderstand the idea of a “mind’s eye,” believing that others can somehow visualize something in their actual field of vision at will.
Many people can do precisely this, though the projected image is not usually 100% clear in some form, i.e. it may have a translucence or impermanence, and it can indeed take some practice to develop this skill. But it isn't uncommon at all, and comes up regularly in some spiritual and meditation practices.
Tell me you can still tell it's two separate layers though. You aren't actually getting visual feedback, you are combining the actual visual feedback from your eyeballs with the pseudo-visual feedback from your imagination.
I would say when I do this, it's more like underlaying my imagination with reality, than overlaying reality with imagination - the combination is still clearly imaginative. Right? (I think I'm a 3 or a fraction higher, on the apple test scale)
I can almost always tell the projection and reality are different, but in some very rare and extreme cases (psychedelics, sleep deprivation, extremely protracted and deliberate meditative focus / practice, or e.g. in situations where the projection is properly aligned with the actual stimuli, like seeing a face in the clouds, or in vivid / lucid dreams), the differences can be so small that what allows you to distinguish the two can simply be knowledge of the will / effort involved or e.g. knowledge from laws of physics that what is visualized cannot be present, or, in the case of dreams, an inconsistency with autobiographical memory.
Otherwise, yes, in almost all other cases, I suspect our experiences aren't that much different.
For me, when I'm really concentrating, the imagined becomes more dominant.
I like to do mechanical/visual designs in my head, intentionally avoiding whiteboards and whatnot, so I don't lose the ability! When I do this, I've had people come up to me and wave their hands right in my face to get my attention, because I don't see them. I "spaced out" alll the time as a kid.
Those people here (not you) telling all of us what is and what is not possible have their view based on how they see things. That's it. I hear they don't see things in mind, visually like a phone or a screen. Or do overlays. I remember finding out and being incredulous about whether others did something I don't.
My mother can take herself anywhere in the world she has been. However, she doesn't render people. This is full immersive. She said I disappear when she's doing it, if sitting there with her. This is simply daydreaming, a well explored topic.
All one has to do is read up on the idea. Anyone arguing against "seeing in mind" likely doesn't and will have some degree of difficulty understanding they are that way. I did. You may as well, or not.
In closing, what exactly is the Mind's Eye. History tells us a lot about what people describe it doing, when it works or is not blown out.
> I have always suspected that people with “aphantasia” simply misunderstand the idea of a “mind’s eye,” believing that others can somehow superimpose something onto their actual field of vision at will. Of course, this is not possible.
If "at will" is the hinging point then maybe not because the scenario to trigger the behaviour isn't clear to me still, but this is definitely possible.
I have had the superimposed belief-image confuse my spatial brain until I moved my point of view (in which the real world obviously reorients while the brain conjured image doesn't) to clear it before. This happened far more when I was younger than now as an adult, but it very much still happens.
The very last time I recall this happening was at my former home during a strenuous house move. I was on my back for a breather between boxes and a delay getting the car around gave me some time for my mind to wander, which it did on that occasion to thoughts about how very old buildings with column supports used to look. By the time I needed to move the next box I had to shake my head to de-conjure the little stone plinths that had appeared in my vision from my imagination trip.
If you have working vision, the brain is doing a ton of post-processing that you're unaware of. The eyes are part of the brain, after all. Closing them with your eyelids doesn't magically turn off those post-processing circuits that do most of the heavy lifting.
There is a full visual memory, just not any sense of an image or sound nor smell in my case.
I am very sure there is a difference after interrrogating multiple people about it.
I keep hearing conflicting answers on this and its a debate I've had with several people.
I have aphantasia based on the description here because I don't see jack.
Some people I talk to though say that they can visualize at the same fidelity as they dream at.
When I dream, I do "see" things in a way that is much more "real" than when I "imagine" them - to the point where I would basically describe it as effectively identical to seeing them.
So now I'm confused about if these few people are the exceptions or the norm, and to what extent I'm lacking something that some others have.
I think when I dream it feels real, but upon closer analysis, it's about as unstable as when I imagine things. It's just overlaid with the additional semantic vector of "yes this is the real world and everything is normal" or something.
In which direction are you supposed to read these charts? The arrows seem to point from conclusion/solution to reasoning, but the reasoning is at the top which makes me feel like I should read it first. It would be less confusing for there to be a single convention (eg “reasoning/evidence at the top with arrows flowing mostly down towards the conclusion”).
Similarly, I was confused by the use of both active and passive: eg sometimes “creates” and sometimes “created by,” with different arrow directions for each.
Thanks for the feedback. I had intended problems to be seen first to convey motivation, so they're at the top, but I wanted the arrows to point in the direction of causation, so solutions point upward. I agree it's not great, and knowing where to look is a common problem. My latest plan has been to have a more-linear view be primary and the diagram view be secondary (split panes on desktop) in attempts to provide more context for the diagram. Also better diagram views so that it's easier to keep only 5-15 nodes showing at a time, then direction wouldn't matter quite as much.
I did notice after the fingerprinting update a marked uptake in strange language in responses. Specifically if I ask it to do something sometimes it will replace some of my request language with synonyms that don't actually make any sense. Like my request was fed through google translate twice
The qualifier is there because it changes the outputs, so it’s necessarily true that some outputs will be worse.
But it’s just as likely to make an output better.
Take the example from the article. He complains that watermarking might sometimes, for example, choose to say “bananas” over “pineapples” because only the former is on the green list, potentially making an output less precise. But 1. It could do that regardless of watermarking since the model is probabilistic, and 2. The more accurate word choice of “pineapples” is equally likely to be on the green list instead, further increasing its likelihood!
Overall, the article is pretty silly because he’s complaining about the possibility of Claude not always choosing the most “optimal” token, even though LLMs are probabilistic so that will happen anyways.
> But it’s just as likely to make an output better.
No, for any particular output token the model's true logits are definitionally the 'best' that the model can achieve.
This is inherently probabilistic. The model's top-1 guess is not guaranteed to be optimal, but it should be so a proportionate fraction of the time. Same with the top-2, top-3, etc.
Watermarking necessarily alters the output distribution away from the model-set distribution, and that alteration is inherently 'worse' in expectation.
You can liken this to a weather forecast. If there's a 25% chance of rain, the forecast should say so (or a 'sampled' deterministic forecast should predict rain 25% of the time). If the forecast is 'watermarked' and predicts rain 27% of the time under identical circumstances, it's a worse forecast.
That being said, this is a case of hiding a message in a noisy channel. Watermarking only needs to communicate one bit ('yes watermark'), so the effects can be arbitrarily small provided one is willing to tolerate an increase to the text size needed for reliable detection.
I think you can also just use the random number generator (seeded with a secret key) as the watermark. Then the probability distribution is exactly the same.
This assumes "writing quality" were somehow an individual property of each word, and the writing quality of the entire text would just be that property for each word summed together.
Which is obviously not how it works.
> He complains that watermarking might sometimes, for example, choose to say “bananas” over “pineapples” because only the former is on the green list, potentially making an output less precise.
If I replace "pineapples" with "bananas" in any kind of meaningful text, I've not made the text "less precise", I made it plain wrong. An incorrect statement. And if the words next to it are still correct or even somehow replaced with even more correct words, the text in its entirety will still be wrong.
> It could do that regardless of watermarking since the model is probabilistic
No, because the probability distribution of the model is generated by its training data and contains semantic information. So the model choosing a completely incorrect word is unlikely. However the probability distribution of the red/green lists is not guided by semantic information.
It's not good for writing code either, despite the many claims to the contrary. At best you come out even on speed as you have to review everything it does. At worst it actually slows you down as you clean up its mess.
> Assuming it's real, I'm not going to put much faith in the opinions of the American public.
You’re talking about Amazon’s actual customers here. Is Amazon supposed to ignore the opinion of their customers for the opinion of some commenter on HN?
In this case, the callout about increasing Merchant Fees is the right thing to do. This isn't great for payments at all. Reduce the friction, improve ease of business.
reply