Well congrats.

Here are some thoughts not facts:

On AA Decision Index.

Decider v1.3 and Decider v.1.2 seem to have same benchmarks. However, V1.3 flash beats V1.2 flash by 4 points.

I am not particulary excited about this pattern/trend of having flash versions for decision models. The whole point of Decision models was to deliver faster ouput for high volume requests at much cheaper cost than Reasoning models. Since most reasoning models can and probably do outperform decision models already in accuracy but they are a lot slower and are pricier. Reasoning Models can already return structured ouput.

So your flagship model must be fastest and most accurate.

Unless it can handle some different form of input other than just image and text, I don't see the point of having both flagship and flash version for the same model.

> So your flagship model must be fastest and most accurate.

That's rarely the way anything in life works though. Do you prioritize fastest OR most accurate? That's a question that depends on context, task difficulty, etc.

You can use a 17M parameter NLI model for simple stuff and get orders of magnitude faster decisions. But the intelligence of that model will be quite a bit weaker.

In some tasks you might be talking about very small differences in accuracy between that model and a 27B parameter LLM decision model for several orders of magnitude in speed. In other tasks, the 17M might be about as good as a random guess.

But isn't there still a meaningful spread of difficulty, risk tolerance, etc. in decision model relevant tasks? Especially at high volumes I think there would still be value in being able to make intelligence/cost/speed/calibration tradeoffs between models

This blog post acted as a rabbit hole into some interesting stuff:

https://www.perplexity.ai/hub/blog/pii-trace-detecting-perso...

https://www.perplexity.ai/hub/blog/a-local-first-agent-for-p...

It performs well on my local benchmark, and is priced competitively. Good timing too, considering that inception Labs is discontinuing their Mercury models API, which I was previously using. It seems plausible that the bidirectional attention performs better for decision models.

Any ideas what decider models are actually good? There does not seem to be a strong competitive benchmark culture (as with LLMs) for whatever reason?

I've found this fairly accurate: https://huggingface.co/spaces/multimodalart/jev-decision-ind...

That's good but I would definitely appreciate one with cost to run the benchmark and cost vs accuracy breakdowns.

Artificial Analysis has an Index which also includes Cost per Task, Time per Task, accuracy, calibration etc. here: https://artificialanalysis.ai/evaluations/artificial-analysi...

(caveat, I work at AA and helped build this Index)

just found this today and used it as one of the links I sent a coworker who hadn't heard of decision models yet!

Complete lack of creativity in the field of AI...

Here's new product, which is just our shitty version of this easily reproducible technology

What is this?

Same for Google and their decisions api

You seem to be confusing market capture for novelty. This almost certainly is not intended to be taken as some greenfield thing from perplexity, same goes for google

Perhaps, but if you're already inside GCP it's nice to have something easy to turn on.

Not just that, if your company approved and gets billed by X cloud, it's quite difficult to request another product from Y cloud

> version of this easily reproducible technology

so you'd prefer indefinite first-mover monopolies?

And all the people involved are making millions to billions. Maybe we're the schmucks.

I don't know how we're going to convince the best and brightest minds of the future to do real work when grifting, corruption, and trend chasing pays the real money. Why try to cure cancer when you can make a guaranteed bag building drones that kill people, gambling apps, or ponzi scheme financial engineering?

Does anyone really use such things for security?

I highly doubt it. Even if a company does internally, its highly unlikely they will reveal it publically.

One use case is live content moderation or streaming where reasoning models could be a lot slower and its hard to hardcode logic.