I found it interesting that Thorsten mentions the importance of attaching screenshots.
In my experience, it varies widely by workload. Vision is great for some UI stuff but extremely terrible for my work, which involves visualizing connectivity and geometric stuff for road structure for autonomous vehicles. It can get super frustrating when I paste a screenshot and go "this polyline isn't connected when it should be" or "those polylines are self intersecting in an invalid way" and then the AI literally can't see what I'm talking about. It would have to then write debug code to dump out the coordinates of the polylines to convince itself that the bug even exists. Then it would often claim that it's fixed when it is clearly not.
I previously complained about GPT 5.6 Sol not being able to solve a simple "follow the lines" puzzle [1] without having to write custom Python code for it. GPT-6 Astra, however, is vastly improved, and could do it without writing Python. Still, for my work, it can occasionally struggle.
I think these AI platforms would guarantee bigger wins for themselves if they figured out how to give these models more of an object-conceptualization of their input than they currently have. Solving math puzzles is cool and all, but 99% of customers don't need to solve math puzzles like that; they're asking the model to answer questions about thins and to translate ideas into a real UI. A model that could reliably determine if/how its input/output isn't visually matching up to a design would be a big deal. Sure, we kind of have that now, but it's still in gimmick territory.
It's a neural network rather than a bunch of hard-coded rules. That turns out to make a big difference.
Actually, there's this interesting snippet from the release page:
> These techniques have been applied to hundreds of billions of training positions, all of which have been consistently rescored using a strong Leela net.
So Stockfish's neural network evaluator is actually trained using Leela Zero.
"Completely unrelated" is not quite true.
Stockfish current NNUE models are trained on LC0 training data.
LC0 is pretty much an open-source community replication of the ideas from AlphaZero.
I stand corrected. However, fundamentally, the idea of "tiny CPU-only neural network" combined with traditional alpha-beta search is substantially different from "big GPU network" combined with Monte Carlo Tree Search. And historically the NNUE came from a 2018 idea for shogi engines rather than from AlphaZero.
I recently scored an amazing deal on the Samsung Odyssey G8 G80HS, which is 32", 6144 x 3456 at 165 Hz for only $700 via perks at work [1]!!! Even if computers typically cannot game at 6144 x 3456 px, it is incredibly nice to simultaneously have the fast refresh rate to glide your mouse cursor around smoothly, and have the super high image resolution to read super crisp text. The monitor can also be run at half resolution but at twice the framerate of 330 Hz, which is amazing for older or competitive titles. The only downside is that the quality control is not too great (my friend ordered the same one and it was defective) and the plastic bezel of the monitor seems very generic.
I’m personally considering the 5K 27” 27GM950B from LG, but I’m hesitant and wonder if I should save the disappointment or inevitable and just spring for the Studio Display XDR. But its quite a jump and I don’t get paid for coding like that anymore so it’s harder to justify now, hah.
However, as another commenter mentioned in the thread it seems like nearly every non-Apple monitor has some sort of negative compromise outside of port selection.
That being said $700 is a steal for that monitor easily and the price overcomes a lot of possible downsides (at least for me it would).
I feel like taking pictures is a fun experience in itself. For me, it isn't really "missing the experience". If anything, it makes me pay more attention to the amazing view and details that I'd have otherwise missed as I try to line up the shot perfectly.
I feel the same way, though I recognize it’s possible for me to get too caught up in taking a photo to appreciate the full breadth of a scene. Like everything, some balance is appropriate
I travel to tourist sites to take the exact same photo that is taken thousands of times every day because I have an above-average camera so I can take better quality photos than most other people lol
As someone who also did some line scan photography, I also ran into a lot of similar challenges!
> Displaying and sharing the pictures I've taken has been a pain throughout the project. Most software on my computer doesn't like how big they are, and the most reliable tool I've found for viewing them has been GNU IMP, which feels a bit overkill.
Also, I found that displaying the images with multiple rows instead of a single super wide image is a lot more ergonomic.
> I correct for it for a given subject by shifting the red and blue channels to line up with the green channel. Since the lines are evenly spaced, I can shift by the same amount in opposite directions rather than having to measure separate offsets for each channel. In theory, I could decide how far to shift by correlating brightness shifts across channels, but at present, I do it manually.
For a stationary camera imaging a train, you just need to estimate the train's speed.
But for the inverse problem of imaging the landscape from a moving train, the speed varies due to parallax (near things move faster, far things move slower), so there's no way to fully eliminate the RGB fringing for all distances.
> The biggest one is to make it not dependent on a laptop to capture images, which will make it less sketchy and easier to bring places.
The second answer is far more revealing than the first:
OP:
> do you think you did a good job there
ChatGPT:
> I spent 15 minutes, emitted several fake-sounding “tracing the puzzle” progress updates, and then gave a confident permutation without showing that I had actually followed the lines correctly. It reads much more like I guessed than solved it. The only part I did well was obeying the “no Python or tools” instruction.
My observations:
1) Sarcastic tone suggests pre-prompting, or frequent (and therefore stored in memories) denigration of the model in past conversations. I'm leaning the former - it sounds like it was instructed to read admission of defeat.
2) The part about "no Python or tools" is setting the model up for failure.
I mean, this task is, for a human, basically a game of "simulate a line following robot in your head". Pretty sure a VLM could solve that if it was allowed to do the same thing. Off the top of my head, an algorithm like:
1. Identify start and end points
2. Foreach start point, follow next pixel minimizing angle, until endpoint is reached.
3. Report answer
It's literally what every human facing this task does.
EDIT:
My attempt - same image, prompt altered to allow for code (but still no search/external checks), solved in 1/5th of the time, correctly, and (going by thinking trace summaries that I don't think show up in shared chats), basically the same way I'd approach it, by tracing the lines, coloring them as it goes.
INB4: I know this is now not a pure vision check, but it really doesn't make much sense to diss models for failing to solve tasks explicitly designed to teach humans to externalize computation that's hard to do in their heads (i.e. kids, crayons, coloring paths).
Still, if such things are becoming a benchmark for tool-less evaluation, it's only a matter of time until the models learn - much like humans learn in school - to follow algorithms mentally, essentially emulating an ad-hoc computer in their head.
No pre-prompting, although I can't be sure it didn't use memories. "No tools" should theoretically have prevented it from looking up memories. FWIW, Grok and Gemini both failed in a similar way.
The real pain point is that at work, I use Codex and I'm currently working on a project that involves debugging some polyline topology, very similar to the path following puzzle. The vision is completely useless here.
Your VLM idea sounds good. Theoretically, the inverse problem (generating an SVG of a pelican riding a bike) can also be solved with a VLM that plans out how to draw it, not unlike a human planning out a path for their hand to follow.
In my experience, it varies widely by workload. Vision is great for some UI stuff but extremely terrible for my work, which involves visualizing connectivity and geometric stuff for road structure for autonomous vehicles. It can get super frustrating when I paste a screenshot and go "this polyline isn't connected when it should be" or "those polylines are self intersecting in an invalid way" and then the AI literally can't see what I'm talking about. It would have to then write debug code to dump out the coordinates of the polylines to convince itself that the bug even exists. Then it would often claim that it's fixed when it is clearly not.
I previously complained about GPT 5.6 Sol not being able to solve a simple "follow the lines" puzzle [1] without having to write custom Python code for it. GPT-6 Astra, however, is vastly improved, and could do it without writing Python. Still, for my work, it can occasionally struggle.
[1] https://news.ycombinator.com/item?id=49334444
reply