Cannot for the life of me work out how they're pricing this as their API page says it's "billed at the chosen language model's input and output rates." But really, why isn't OpenAI building their own agents if their tech is so powerful? - ThreadSky

edzitron.com • 3 days ago

Cannot for the life of me work out how they're pricing this as their API page says it's "billed at the chosen language model's input and output rates."

But really, why isn't OpenAI building their own agents if their tech is so powerful?

Reposted from The Verge

OpenAI launches new tools to help developers build AI agents

Comments

darklock.com•3 days ago

They're crowdsourcing the problem of figuring out how to make the agent actually fucking work for something, so they can steal it and pretend it's their own product.

t0mmy44.dev•3 days ago

And like many kickstarters...

aetherlev.bsky.social•2 days ago

Oh boy, they’re going to be disappointed when the results are as mediocre as Elon Musk’s sexual performance

psynautic.bsky.social•2 days ago

like when apple did this for vision pro lol

darklock.com•2 days ago

Once upon a time companies would publish an API and when you got successful they would offer to buy your company and you could make bank. Now they keep the SDK in the cloud, you have to put your app in the cloud, and the EULA says all the apps technically belong to them so you don't get shit.

vought.bsky.social•3 days ago

I suspect it is because building agents requires human intervention and instruction.

It’s bullshit all the way down from there.

t0mmy44.dev•3 days ago

Because it's not. I've tried to use it for software development tasks. It kinda sucks, but it looks convincing while it does it. I'm told I was doing it wrong by not telling it explicitly to not make shit up, but hey, I'm just a software engineer what the fuck would I know.

t0mmy44.dev•3 days ago

The good-ish AI agents being made are being built by human engineers. They're narrowly scoped, pretty handy for what they actually do, but they aren't what people imagine. And also, look at the effectiveness of rage bait bots. They ruin things without even being high quality.

automationbias.bsky.social•3 days ago

Right? I also wonder if they have people astroturfing forums, because I see a lot of claims that don’t bear any resemblance to what I’ve experienced for programming tasks.

nathanharper.bsky.social•2 days ago

you’re misunderstanding what an “agent” is. If these things could just freely use everything on your computer, that would be terrifying. The point of this is to allow developers to integrate the AI in a sane and safe way with their products. OpenAI sucks, but this is unremarkable beyond pricing

nathanharper.bsky.social•2 days ago

Also often the model needs to be trained on data for niche use cases

edzitron.com•2 days ago

1. What're you talking about? Like what is the way it would be integrated
2. What is "the way it is priced" and how does it work

nathanharper.bsky.social•2 days ago

I feel like my explanation was terrible so I’ll try again: the agent is the middleware that allows an ai model to safely and consistently use proprietary softwares. It does not make sense for OpenAI to make these because it does not make the target software, it makes the operator.

nathanharper.bsky.social•2 days ago

You COULD make an agent that has free reign to use all your software the way a real person would, but that is abjectly terrifying whether the operator is OpenAI or SmarterChild

edzitron.com•2 days ago

You absolutely cannot make an agent that does that

nathanharper.bsky.social•2 days ago

I could write an AppleScript right now to control the GUI interface of applications on my computer. I could also right now generate that AppleScript with an ai. It is not a big leap to execution

nathanharper.bsky.social•2 days ago

Why?

nathanharper.bsky.social•2 days ago

By mentioning price I was saying that that’s what the interesting development is going to be here. The fact that they’re releasing a developer toolkit at all is not surprising. I don’t know what “correct” pricing would be and I’m not buying it. 1/2

nathanharper.bsky.social•2 days ago

This is hard to explain if you don’t work in software, but most applications kind of have their own language that other software can use to interact with it. You CAN probably just rig this stuff up to simulate button presses like a person using the OS, but like I said, kind of scary.

nathanharper.bsky.social•2 days ago

FWIW I do think they oversell how “smart” this stuff is and a lot of what they’re doing is wildly unethical, but the tech IS scary enough to be dangerous without guardrails

doomscrollmvp.bsky.social•2 days ago

edzitron.com•3 days ago

"They ARE building them Ed!" Where are they then lmao

edzitron.com•3 days ago

here's the thing: this is basically OpenAI's entire plan for their products, except nobody seems to be able to make anything good
https://bsky.app/profile/darklock.com/post/3lk535qsi6s2v

langoustine.bsky.social•3 days ago

Concerning

mux-north.bsky.social•2 days ago

seems to underline how good those early-dqys fart apps were for iphone

xlem.bsky.social•3 days ago

This feels a little bit like what Amazon tried to do with Alexa Skills back when voice assistants were supposed to be The Next Big Thing

dsquareddigest.bsky.social•3 days ago

they're concentrating on literary metafiction, where the real money is

eschatonblog.com•3 days ago

[guys who don't read fiction, ever] this fiction machine is going to make me rich

dustbowl-clickfarm.bsky.social•3 days ago

Novels are just lots of words

theangelofhistory.bsky.social•2 days ago

These agents better be bloody brilliant to somehow get us to cancel our contracts with Teams, Salesforce and Workday

skepticpunk.bsky.social•2 days ago

Isn't that supposed to be what Deep Research is?

hbarsquared.bsky.social•3 days ago

They're totally the best agents and definitely real, but they go to school in Canada so you can't meet then yet.

ilk.org•3 days ago

It’s exactly like all those guys who eat to sell you their secret foolproof method for getting rich.

tokilabs.co•2 days ago

The trajectory as of late is to become the “AWS of AI”. A lot of people want an easy way to play with LLMs, and the killer apps are entirely unclear. But they can at least host the infrastructure and provide tooling on top to bundle/lock-in services later down the line.

herring.land•2 days ago

They can’t make an agent reliably do the thing, because it turns out getting the random number to be repeatable is hard. So they’re going to outsource the risk of agent mistakes out to third parties, this is Tesla “self-driving” energy.

peterswimm.com•2 days ago

What I don't get is if there is a nontrivial amount of code on my end I need to author to make the agent work then why am I send the output to a cloud model anyways.

chrisgart.bsky.social•2 days ago

falken.bsky.social•3 days ago

As a developer, this is actually more in line with what I want from them. I don't want some magic oracle, I just want to be easily able to plug bits of their technology into my own systems.

Of course, they don't make money off people using tech, so it makes no sense for them to promote that.

edzitron.com•3 days ago

can you tell me how much this would actually cost to run an agent, though?

zeroghost.net•2 days ago

Remains to be seen at the moment depending on what hyper scaler you're using. The larger costs will come from things like document processing in industries where knowledge work is king such as legal contract management. Also depends on the actual orchestration. For complex work at scale - not cheap!

zeroghost.net•2 days ago

This is from a mid-market to enterprise perspective. The real money comes from compute spend on large tasks. Costs can go up exponentially, especially when using micro services like Azure Doc Intel etc ..

edzitron.com•2 days ago

There is no money in compute spend though. It's all horribly unprofitable

zeroghost.net•2 days ago

They've recently inked some deals around compute to bolster their ability to offer it as part of their efforts to make money. Outside of API calls, which can be costly at scale, it's hard to see how they will make money in any meaningful way when compared to their eval.

falken.bsky.social•2 days ago

The value of this SDK they are releasing is that it would lock you into OpenAI.

Right now, I am free to pick and choose the LLM component of my agent based on price, performance, etc. Also, given the flexible nature of LLMs, it might be easier to swap these out than traditional SW components.

falken.bsky.social•3 days ago

I would just use langchain to do the orchestration and then it is going to be down to estimating how often I think it is going to actually call out to which ever AI service choose to integrate.

It's really going to depend on what I have the agent doing and how many different services it uses.

jeffrush.bsky.social•2 days ago

Remember: You can't spell "Why the hell would I be paying this much for OpenAI" without "OpenAI"

invalidham.bsky.social•3 days ago

Selling shovels in a gold rush

seven3.design•2 days ago

The shovels:

cuw.bsky.social•3 days ago

So they are just making it easier for other devs to use their gigantic expensive unprofitable dataset for distillation for their own AI and then not have to use OpenAI products ever again. Ed I think this might be poorly planned out!

bnej80.bsky.social•2 days ago

The whole concept of an agent built with LLMs is whacky to me. We already have computer to computer automation, it's incredibly fast and cheap because it doesn't use a user interface.

Why the heck would a computer need to use an interface (language, UI) designed for humans?

ukplissken.bsky.social•3 days ago

Shovels. They’re selling shovels.

Microsoft have already moved to this for CoPilot.

Comments

Posting Rules

Reply