15 karma · joined June 6, 2024
With opus 5.5, it's quite good. Normally I just prompt that it will be 3d printed and it avoids floating objects, from time to time I see on the slicer that there are and it checks all geometry and fixes it.
Most of the issues come from bad measurements or interactions with real world objects (margin for impurities, improving ergonomy, etc.).
For easy changes I could be faster if I knew how to fix the STL file directly. For more complex designs, I think opus is faster. I would say that knowing what you actually need to create and the limitations and characteristics of 3d printing is more important than knowing how to handle the 3d design tool.
For example, I use LLMs to create 3d models. I don't tell them to recreate the 3d vertexes every time, they have a python script that programmatically creates the shapes, and with a freecad mcp, they get views in images, also they can create other scripts to check for collisions and other issues.
They can also get a histogram, discard noisy values, check outliers, etc, etc, using scripts or by running a traditional ML value to get later the refined data.
Opus 5.5 is quite decent currently, but expensive and not enough for many challenges, but I see the progress happening.
You could also pay for a ML product, but not so easily and cheap. This enables PoCs and small products.
It is not, it doesn't matter if a brain's neurons or a prediction algorithm managed to get the solution if it has done so in a (sort of) reliable way.
https://en.wikipedia.org/wiki/Duck_test
Errors are made by both machines and humans, so that's not a differenting factor.
What is clear is that the human element will be more valued on art as time goes on.
Still, if LLMs are capable of writing working code, art and solving math problems, they qre definitely capable of doing suboptimal routing (with some algorithmic aid)
The LLM would learn the heiristics (example: data lines first, power lines later, etc.) and would request the autorouter to do that routing, then take the image output and request a different part (depth-first). If later no routing is posible with that configuration, after some retries it could try another way.
It wouldn't solve complex boards, and engineers are always needed for short comings (and even if not, for research), but I wouldn't say this is something different than poetry, music or pixel art, LLMs can imitate although most of the times in a uncanny way.
The big AI labs won't do that unless they are forced to, as they want you to spend more money on the big, expensive, frontier models (so they can live up to their valuation), so it's more likely that you will see this on smaller open weights models.
Companies have to pay monthly for the harness app (codex, claude code) and the tokens are priced separately based on standard API pricing.
Then they have a price table that is the official public API prices for the models, but converted to their AI credits.
Currently, I'm averaging 15 credits (0.15 $) per request for some QnA questions with claude sonnet 4.6. It should be similar with gpt 5.4 (same output token price, similar input token price).
I think with the current pricing you can do light work with normal models (10-20 code completions/day), but vibe coding, heavy usage or using bigger models is not viable anymore.
It's hard to measure since I did not average my token expenditure per day.
Unfortunately, nobody saved it on the wayback machine (I just did with today's page), but you can search "github copilot pricing multiplier" and still see on google images the cost increase x6 or x9 on some models.
I am on an Enterprise plan so it's different for me: I pay the subscription and I get credits, but the token pricing is essentially the same as the public API pricing.
I understand they needed to change it, but I don't like how they try to erase the old prices from the record.