I have a few custom ones (a post-trained flux 2 checkpoint for web design and a image->metalness map generator that the build step can call for more advanced lighting situations), but gpt-image-2 is better than my own for design, so it's weighted much more heavily in outputs my tool generates. I think gpt-image-2 currently generates 99%+ of the design outputs on diffui
Yes but it generally expected documents to be in a certain layout and the user had to fill in a lot of the gaps. It's more able to identify elements on an incoming PO, match a supplier invoice, or create a receipt based on a packing list with inference.
EDIT: The other activities, yes I think it made it cheaper and easier. They both required human intervention to make an accurate forecast or execution plan. Forecasting requires humans to do things like apply smoothing factors, identify abormal demand over huge datasets, or identify seasonal demand cycles, etc.
Constraint scheduling requires a human to intervene in dispatching and resource allocation as well. AI can do 95% of the work with a proposal, or even control the process. Supposedly much better than historical algorithms. I haven't implemented this part of it, but I've seen it done.
The documentation part is much more robust. You can ask it complex questions, like "How do I setup a phantom blow-through part in a MS level 1 BoM?" and it walks you through the entire thing.
With the sales suggestions, it can identify things like "The customer is buying spaghetti and pancetta" AI: "They are probably making carbonara, how about some garlic bread?"
Then take a photo of a document and throw it into ChatGPT, or Gemini, along with the word "Transcribe".
ChatGPT wins, by a landslide. BUT AI's advantage doesn't stop there.
"Take this picture of some idiot filling in form 49, I've also attached the PDF, fill in the PDF fields, provide a database record according to the schema attached and flag if there are any obvious problems with the entries".
That works too. "Produce a latex document of this kid's math homework and flag any problems" - works. "Produce a MS word document of this letter" - works. "Read this bill and produce a JSON version following the schema from this example" ... and so on and so forth.
More than that, it is starting to work pretty well with Gemma 31B local model (will still do 30 document analyses on cpu only, at Q4 on a DDR4 or higher machine. Yes it's mostly memory speed that matters) at this point. I mean, that's GPT 5.0 or so quality (95% correct with the occasional problem), but hey, Qwen 3.8 27B may be coming out next week ...
Actually 30 is sequential on an hosted Xeon server. No GPU, about 2 minutes for a ~2000 token answer. I should try parallel since you're quite right, I should be able to do faster that way.
reply