30 comments

[ 2.7 ms ] story [ 58.8 ms ] thread
This is cool, Id love to be able to fine tune on this architecture. Is this something on the roadmap ever?
Smaller models really arent great at structured output. If this works it would be great for a local model that might not be as good but as long as it respects structured output will be vastly more useful.
> These are deep neural network architectures that are task-specific for things like OCR, translation, or GUI detection. The way they consume and see data is trained to be task specific, which makes them up to 100x more accurate at their specific task. They also produce useful metadata like bounding boxes and confidence scores, letting developers build predictable workflows they can rely on.

Does code extraction and manipulation fit in that? Would interfaze be the agent that a coding agent uses?

Potentially stupid question: Does that mean we can chain them together line UNIX command line programs ? That would be so, so intuitive.
This is very cool, though I don't understand exactly what they've done here. Is it some kind of LLM with convolutional layers added?

The graph doesn't exactly make it clear but it describes a pipeline that goes beyond the LLM, so the CNN could be a separate model there.

Here’s the academic paper behind it: https://arxiv.org/abs/2602.04101
Thanks. Well this is fascinating.

>Instead of a single transformer, we combine (i) a stack of heterogeneous DNNs paired with small language models as perception modules

It seems that we're reinventing the brain's organs one by one from first principles. (Though Transformer + Common Crawl unintentionally builds a whole bunch of them we don't even understand yet.)

I found some broader context and the whole thing is indeed very harness-shaped:

>Using Interfaze as a Tool Inside Your Agent

https://interfaze.ai/blog/using-interfaze-as-a-tool-inside-y...

Well, Harness is the wrong word here... "environment/tools the LLM interacts with" definitely fits though. Or "other organoid" to use the previous metaphor.

Similar to a large action model?
Not directly, LAMs tend to be focused a lot on tool calling or trained for a set of specific action for example in the robotics field. Good tool calling might be a good by product of Interfaze but wasn't specifically trained for that use case.

The focus has been for deterministic outputs that require high accuracy. In situations where there is "one right answer"

So is this basically a task-specific MoA transformer arch with a DNN that helps make routing decisions? Trying to understand this.
The other way round, task specific DNNs adapted to share the same vector space as omni-transformers with generalized vision, audio encoders.

E.g. For an OCR task, the first pass will be handled by the CNN, converted to shared tokens which the transformer can consume, correct any issues if needed and a decoder that can handle both the DNN and transformer output.

What I want are precise and tight bounding boxes. Why is this so difficult?
Amazing!

I just tried the OCR capabilities with a photo of a DIN A4 page which was written with a typewriter. The image isn't the easiest to interpret. The text perspective is distorted because the page is part of a book and the page margin toward the spine of the book is very small. There are also many inline corrections due to typing errors while the page was written (backspace couldn't erase characters back then, and arrow keys couldn't be used to add text in between existing words). Over the past months I've tried to use several LLMs on this very same image already (1 out of 200 pages that seek digitization). The result is by far the most accurate so far. Only some very minor errors (which are also non-trivial for human translators) were made.

This page induced costs of about 25 cent. I assume I could tweak the input image a little more to consume less input tokens. OCR-ing all 200 pages would otherwise cost a juicy 50$ - although there is a generous 20$ of free credits.

Induced cost: 108.8k Input tokens => 16,32 cent 24.5k Output tokens => 8,58 cent

// Edit: I just re-tried the same task utilizing a capability of the API to only run a specific part of the model (e.g. _only_ OCR). This cuts cost by 3x (to ~8c/page) but significantly worsens the result. The result is missing entire lines of the original document. There are also many error in the text that was recognized.

Have you tried this task using an actual OCR model like Google Cloud Vision AI? I am not sure if this is what Gemini uses under the hood but multi-modal LLMs are not designed to extract text like this so it should be no surprise it's not good at it?
I don't think I've tried Google Cloud Vision on that particular image, no. In my experience, based on some tests from a year ago or so, Azure Document Intelligence impressed me the most in terms of OCR - out of the big three players: GCP, AWS and Azure.

I should retry the experiment because there has been a lot of progress since then and I could imagine that GCP improved there vision models since then.

Yup run task mode runs a much smaller part of the model when can drop quality of scans. The issue with run task we have to figure out is how much of the model is needed just for OCR and how to activate the right parts. A lot more improvements coming here with the same cost reduction.

I'd be happy to test it against your sample and see how we can get good results at a lower per page cost. Feel free to email me yoeven@interfaze.ai

New account created ~5 hours after this post, with a single comment specifically praising the model / product. I want to believe, but this sort of astroturfing isn't very encouraging.
I totally understand, and I can't blame you for that. I wouldn't think otherwise. I am a long-time follower of YC but never posted any comments. I wanted to share that experience which is the reason I created the account. I don't know how I can proof to you that I am a legitimate person who has _no_ affiliation whatsoever with Interfaze. I can only ask to try it out for yourself. I was genuinely impressed by the results.
Can this run locally or is this a service?
Ok that's...just cheating. You can't take a benchmark like MMLU designed to test the performance of a single general language model and compare it to performance of a small specialized model designed to do well on MMLU.
does it handle source code extraction from images?

how do I run it locally?

Interesting approach! One question though: can the model do column detection?

The first OCR example returns output that does not detect the article columns - the bounding box is the entire first line.

Gave it a try for structured data extraction. Tested returning a JSON object from images.

The output was correct, and seemed deterministic, although I ran it only 2-3 times on the same image.

Main problem is response time: it took about 20-25 seconds for a simple structure of 5 fields. As such unusable at scale, let alone "real time" processing.

Other problem is cost, it is considerably more expensive than more established models for the same document, like flash-light.

Shame, the architecture is very interesting.

Thanks for the feedback!

We're working a lot more on speed in the coming few weeks :) More GPUs and more optimizations.

Our has been focus on quality of output first and we'll make optimizations as we grow :)

The lite models are great for simple use cases but won't don well in more complex OCR use cases.