This is just a very rough basic game skeleton. Companies believing AI is smart enough to spit out almost ready product will learn a hard way that it will will give them just a starting point that still needs a lot of work. This is a nature of all modern LLMs - they easily lose context: the bigger the context the more losses and distortions are, especially in the middle. So, giving AI the complete book does not mean it will follow everything in it - quite the contrary.
Karpathy’s Pelican
231–240 of 461 posts
Re: Karpathy’s Pelican
#232I don't think it's a bad way to benchmark new models, I just find it concerning that the author implies that "pelican on a bicycle" has been exhausted. At the risk of making overly broad, unfalsifiable claims I think multi-year exposure to AI content has dramatically raised our expectations for speed and volume but lowered them for quality. We see a very janky pelican and declare the problem solved.
I think it's touching the limit of what one can reasonably expect any intelligent thing to produce with the only direction being "produce an svg of a pelican riding a bicycle". When you aren't sure if an LLM can write an svg well, or that it will be able to form a pelican shape, or animate a bicycle, it's a good test. After that, it's all judgement: how detailed should the pelican be? pelicans are the wrong shape for…
Could you share any examples that come close to those limits? I haven’t seen any that don’t have obvious flaws in proportions, layering, composition, color palette, visual clutter, or stylistic consistency.
Re: Karpathy’s Pelican
#233This is just a very rough basic game skeleton. Companies believing AI is smart enough to spit out almost ready product will learn a hard way that it will will give them just a starting point that still needs a lot of work. This is a nature of all modern LLMs - they easily lose context: the bigger the context the more losses and distortions are, especially in the middle. So, giving AI the complete book does not mean i…
No, it cost 1M tokens, or about $10
EDIT: The parent comment originally claimed they spent $1M on the demo -- they seem to have edited it after I replied.
Re: Karpathy’s Pelican
#234I’d like to see a human one shot a pelican on a bicycle in raw svg.
But the LLMs are also not one-shotting it, or are they? I assume they have some ways to verify it, e.g. to visualize it (convert to PNG, then feed as vision tokens back to the LLM), or other ways, maybe also pure text LLMs have some ways to verify the result at least somewhat? And with such feedback loop they can iterate.
Re: Karpathy’s Pelican
#235A lot of people are posting here about how bad the end product is, but that is kind of the point. Models have moved beyond generating images to a new kind of benchmark that better exposes understanding of the physical world, and we can use benchmarks like this to measure future progress. (Of course, it will have to be a qualitative/subjective measurement.)
Aren't they still bad at understanding how bicycle frame works? Especially the steering part?
I don’t know what datasets are available to these LLMs, but I’d imagine if there was training on CAD code, text, and images, a prompt steered towards that probably could get it pretty good.
I am not a mechanical engineer, so even prompting well with ME lingo probably will take some effort.
Re: Karpathy’s Pelican
#236It would be interesting to see the models work on a book it hasn't been trained on yet. I guess sadly that means any book released very recently. Definitely impressive demo, I do wonder though if the countless artwork, films, images etc produced over many decades around Lord of the Rings somewhat influenced the outcome of this though.
Yeah that seems like an enormous problem with this example.
Re: Karpathy’s Pelican
#237Earlier quoted context omitted.
What do you mean by "oneshot"? The term applied to genai makes sense, but it doesn't make a whole lot of sense to apply the term to human art.
I understand it as “write SVG, say when you are done and only then you are allowed to see the rendered result”. Multiple shots would be either “you get more than one try, we'll pick the best” or “you can see the rendered result and iterate” (or maybe that would be “model + harness with tools”?)
[dead]
Re: Karpathy’s Pelican
#238I don't think it's a bad way to benchmark new models, I just find it concerning that the author implies that "pelican on a bicycle" has been exhausted. At the risk of making overly broad, unfalsifiable claims I think multi-year exposure to AI content has dramatically raised our expectations for speed and volume but lowered them for quality. We see a very janky pelican and declare the problem solved.
I don't think he's claiming it's been exhausted. It's just that things have progressed to a point where people are arguing over the finer points of which pelican looks better -- which is often a matter of taste, and an indication that we've hit the knee in benchmark where models are no longer failing in obviously awful ways.
I haven't really seen evidence that any ai can reliably draw a pelican riding a bicycle. Not if you look at the image long enough to take it in. Even the best ones have something wrong with them. Not a matter of taste but a matter of having both legs peddling on the viewer's side of the bicycle or having two beaks.
I'm actually beginning to wonder if some people who ignore these things have a different, somewhat lesser ability to percieve image details than I do.
I mean I guess its fine to go on to another test despite never actually passing the pelican bike test, but there's a sense that we have to use another test because AI is now good at pelicans on bikes, which is just not true.
Re: Karpathy’s Pelican
#239Re: Karpathy’s Pelican
#240Earlier quoted context omitted.
Agree. The pelican benchmark was interesting a year ago when most models struggled and a good pelican indicated an unusually capable model. Now it’s saturated and uninteresting. A good new benchmark should have awful performance to start and there should be a lot of headroom for improvement. This benchmark is also intentionally difficult and requires the LLM to develop the animation through spatial reasoning and firs…
Have you seen pelicans in Simon Willison’s tests? It is still not a pelican on bike I would like to publish :) Imho we don’t need to make benchmarks that draw the whole 3D world. Pelican’s drawing is really nice in its simplicity and complexity at the same time. It seems to be an obscene waste of compute time to generate useless 3D worlds that are just a bragging - 3D is really heavy discipline to make it right, see…
Do you expect the SVG to emulate a hand-drawn picture, become more realistic, or just a more detailed illustration?
As for the often quoted issues with the bike's frame or problem with the steering column, I can't really tell, I am no bike expert.
I can instead judge how poor of a job it is doing with a LOTR rendition in Three.js, so that seems like a better benchmark.