Hacker News Reader

Comment by irthomasthomas | Hacker News Reader

This is a perfect illustration of something I noticed with llm progress. Ask them to improve an svg like this, and it never fixes the missing crossbar or disconnected limbs, it just adds more stuff. In this example they have obviously improved greatly, and it contains a ridiculous amount of detail, but they still to get the basic shape of the frame wrong. It's weird. And the pattern shows up everywhere, try it with a webpage and it will add more buttons and stuff. I've even experimented with feeding the broken pelican svgs to an image model to look for flaws, and they still fail to spot the broken elements.

edit: fixed human hallucination

View on HN

When you say "improve an svg like this", how are you imagining setting that workflow up? Are you just feeding them the SVG to iterate on; or are you giving them access to a browser to look at the rendering of the SVG?

I ask because:

Insofar as the original pelican test is zero-shot, it effectively serves as a way to test for the presence of a kind of "visual imagination" component within the layers of the model, that the model would internally "paint" an SVG [or PostScript, etc] encoding of an image onto, to then extract effective features from, analyze for fitness as a solution to a stated request, etc.

But if you're trying to do a multi-shot pelican, then just feeding back in the SVG produced in the previous attempt, really doesn't correspond to any interesting human capability. Humans can't take an SVG of a pelican and iteratively improve upon it just based on our imagined version of how that SVG renders, either! Rather, a human, given the pelican, would simply load the pelican SVG in a browser; look at the browser's rendering of the pelican; note the things wrong with that rendering; and then edit the SVG to hopefully fix those flaws (and repeat.)

I imagine current (mult-modal and/or computer-use) LLMs would actually be very good at such an "iterative rendered pelican" test.

I'm talking about two type of improvement, model improving, and prompt based improving. I am noticing that the baseline output has a lot more going on, the model has improved, yet it still makes those obvious looking mistakes with the shape of the frame or disconnected limbs etc.

And I am saying that if you take one of these SVGs and ask an LLM to look for flaws, it rarely spots those obvious flaws and instead suggests adding a sunset and fish in the birds mouth.

This is also my gripe with a lot of this stuff, always evaluating models on what they can literally oneshot is completely pointless; it's not how anything works, neither for humans nor for scaffolded AIs. I guess it's neat if you want to argue that a certain level of intelligence can "never be achieved" in a single forward pass, but like, so what. No one cares about that, except people who have already decided to be anti AI.

(not that I am in any sense pro AI, but it's just a weird lack of intellectual rigor)

Asking a model to improve its output is not one-shotting tho? My observation was that asking an llm to iterate and improve a response causes it to add more stuff, rather tha repair the broken stuff. And that model progress in general has the same pattern. This new model adds more details to its responses but continues to make mistakes at about the same rate.

The question was whether you were giving it the rendered image and using the model's visual modal capability, or feeding back in the textual SVG.

It's hard to "imagine" what the rendered SVG looks like, for both humans and LLMs, so just iterating on text won't really be as useful of a test. But if you show it what it rendered, it might observe the bad-looking bicycle and be able to fix the text that way.

"I've even experimented with feeding the broken pelican svgs to an image model to look for flaws, and they still fail to spot the broken elements."

Same pattern for opus 4.7>4.8 and a 3D astronaut https://www.reddit.com/r/ClaudeAI/comments/1tt3a8h/differenc...

To a certain extent, it feels like a Sonnet 3.7 moment. Slightly overeager - you ask for a button color change, you see layout changes, new package dependencies, and the README rewritten from scratch - and not necessarily correctly.

When I ask for a pelican on a bike, I want the Platonic ideal of a pelican on a bike, not a vision of an alternative reality in which pelicans created bikes. Though, thinking about it again, maybe I should.

What is “Sonnet 3.7 moment”?

Slightly overeager - you ask for a button color change, you see layout changes, new package dependencies, and the README rewritten from scratch - and not necessarily correctly.

Sonnet 3.7 tried its damnedest but it was just kinda "off".

It's because LLMs are fundamentally generative (creative), not truth-seeking or logic-seeking. Simple logic has always been incredibly expensive to impossible for LLMs.

So we have to train llms on debugging too, not just how to make things (which you easily train by feeding the outputs).

This matches my experience with human too FWIW.

Why is there always an identical reply like this when anyone criticizes LLMs?

Their ability is best described as "spiky". To steal from aphyr: think kiki, more than bouba. Whats interesting is that a lot of the models seem to have similar spikes and "troughs", though there are differences.

Reply on news.ycombinator.com