@erik_kokalj: openai just released Decisions API, Jev-like "system 1" multi-modal model based on GPT6 Luna! Each egg gets cropped and…

openai just released Decisions API, Jev-like “system 1” multi-modal model based on GPT6 Luna!

Each egg gets cropped and sent to OpenAI (~270ms latency) and returns P(clean, dirty, cracked). One image crop consumes 280 input tokens ($0.03/1K imgs). @roboflow RF-DETR for detecting eggs

I’m pretty sure that vision-only system is faster/cheaper than a mechanical, echo-based detection lol But yeah we could replace decision api once we have enough training data of edge cases, and run that locally as well as detector

for prod - for sure. but for poc when you’re collecting training data to fine tune your classifier this is great way to get up&running

agreed, but you need enough training data to finetune a classified, this is poc

in some cases it’s hard to get enough training data for all edge cases, and it’s easier/faster to use general model. Training an egg detector is easier tho, as you can quickly get training data

agreed, but you need enough training data to finetune your classifier

Thanks!

Yup, def worth training once you have enough training data!

exactly!

yep - but you need enough training data to finetune an accurate classifier, this is zero-shot

of course, but you meed enough training data to finetune the classifier

it’s not, but it’s a bit harder to get enough edge-cases for training data to finetune a classifier

Yeah we’d need a rolling conveyor belt like this!!

For QC of eggs? I’d give link/sc of this thread to chatgpt and ask it go guide you thru installation. For PoC you’d need usb cam, a Jetson (like orin nano) for compute, then fine-tune your detection model (eg rf-detr), roboflow is the easiest way. Then use ai agent to implement local inference, tracking, counting, and egg classification (eg via decision api)

are you asking how scalable openai’s infra is 😅?

Yeah of course, decision api returns just confidence % of each class, you then decide what to do with that (thresholds etc)

need enough training data to train a classifier for it

not really about roi, its hard to get enough edge cases in dataset

you mean train a model with a dataset? depends on model arch and dataset size, but let say from few minutes to few hours

The problem is getting training data for edge cases, not training. Probably for classificator its below 1h of training time, and you can use eg Decision Api for labeling.

totally agree 💯

rf-detr for detection, Luna via decision api for classification

of course, or jetson + cheap usb cam. The problem is getting enough training data of edge cases tho

Well its not vision-only model, you can also just send it text, so perception wouldn’t be fully accurate

This would be applicable for poc phase of the project, where you’re collecting training data. Then for pod you’d finetune your own model

pure CV isn’t as robust. you could train a classified, but need enough training data of edge cases

yeah detector (rf-detr) runs much faster, above 30fps locally, but classifier is openai-based

true, but in prod you’ll collect edge-cases, build up training data and finetune a classified for it

no, it tracks each egg separately, and only runs quality control three times on each egg (for verification). In video the egg turns green afterwards (we dont check it anymore)

Exactly! But it takes time to get enough training data for finetuned classifier

we run object detection + tracking to get id of each egg in the video

it was a follow-up answer to my answer lol, both answers are true

It’s using rf-detr which is more efficient than SAM, but yeah SAM could do this as well without fine-tuning

Agreed, but you need to collect enough training data (edge cases) to finetune a classifier

detector is fine-tuned, classifier is zero-shot (decision api) :)