Yet another Hacker News

Typesafe-computer-use drives a Mac toward a goal for 1/50th of a cent per step

72 points rahimnathwani 2 days ago 37 comments

john_minsk 3 hours ago | parent

Super cool. Hope waitlist will move soon. I have a use case for it too.

are you the author? If so - what are your notes on using Jev in this scenario?

jasonjmcghee 3 hours ago | parent

Did you do the follow up questions? I was invited within a few hours of joining today.

It's also now on openrouter and cloudflare

baxtr 3 hours ago | parent

Same here. 8h hours later I was in.

sroussey 3 hours ago | parent

Curious about people’s experience here. I am working on a small model, verify by jev, and escalate to big model. Some cases, the small model is not a model but some regex.

cheap-confirm-escalate

Using jev as the confirm step.

Zaraif13 3 hours ago | parent

How does it do on OSWorld-verified? Recently read that even Fable 5 is just at 85% .

jgilias 2 hours ago | parent

This probably throws a spanner in the wheels there:

> Every piece of reasoning the frontier model does for free has to be rebuilt here as deterministic state.

EDIT: Not to shit on this though. I totally believe that some smart mixture of LLM-reasoning + Jev-style + determinism is going to be pretty amazing.

mmastrac 3 hours ago | parent

I'd be interested to see if using DiffusionGemma-as-Jev helps as you can feed the image directly into the model and it'll make decisions based on the image embeddings.

nowittyusername 2 hours ago | parent

I had a long talk with chat gpt about this today as well. I think its duable and prolly not too hard either, also you could do lotsa funky stuff with stitched frames of a video in one 4x4 grid for example and send that as one image for analysis. that way temporal understanding can be had for fractions of a second by jev... also because vlm works in pixel space you can get around the whole state machine issue as well, so many possibilities...

jdkoeck 2 hours ago | parent

I trailed off a few lines into the README. No human ever edited any of this. « LLM detected, project rejected ».

alex_suzuki 2 hours ago | parent

I stopped reading at “The honest caveat”…

antoniojtorres 2 hours ago | parent

Every third post in HN has this same complaint comment. Everybody knows.

soulofmischief 2 hours ago | parent

And every comment on HN calling out vibe-coded slop has this same complaint comment in turn. OP has an actual complaint, your complaint is just "stop complaining".

antoniojtorres 1 hour ago | parent

It’s not stop complaining as much as it is please contribute something of substance instead of venting hot air into a comment thread where everyone is already aware.

jdlshore 2 hours ago | parent

I appreciate it when people say something is slop. Saves me from wasting my time looking at it.

How many slop ideas have come to the front page, never to be heard from again because the execution isn’t actually any good? I’m guessing most of them.

theturtletalks 1 hour ago | parent

Just because the ReadMe is slop, doesn’t mean the code is slop. People are starting to make apps for themselves now and open-sourcing them so they’re not putting much thought into the ReadMe or distribution.

This just means ReadMe’s are less important now. I just have my terminal agent dig into the code and tell me what features are there. If the app is actually useful.

Even before AI, there were so many projects with subpar ReadMes, no screenshots, etc. But once you use the software, you realize how good it is.

Source: I maintain a massive collection of open-source alternatives and quality of open-source alternatives have increased a lot

tacomagick 1 hour ago | parent

If you do not respect your project to write your readme yourself chances are i will not care though.

theturtletalks 1 hour ago | parent

For many new coders, LLMs are so good at writing the code, asking it to write the ReadMe sounds like a good idea. Clearly it’s the first impression your project makes so handwriting it is important.

Being turned off by the project because of the ReadMe is your prerogative. I’m just suggesting you dig into the code sometimes, the ReadMe is not the be all, end all.

sampullman 54 minutes ago | parent

But it takes so little effort to wade through an LLM readme and cut out or edit it down to something palatable.

Just the bare minimum effort to do that would be nice.

pjmlp 40 minutes ago | parent

I apply the same rule to the code, unless it is on the job and I have to follow along with the herd, or else....

jdlshore 1 hour ago | parent

You’re right that a poor quality readme doesn’t mean a poor-quality product, but it seems more likely than not to me.

Slop is an instant tab close for me. If something’s good, it’ll come around again. I’ll catch it when there's some evidence that it’s worth my time.

Kwpolska 1 hour ago | parent

This project is 100% slop, as evidenced by all commits having a Claude attribution. https://github.com/awlevin/typesafe-computer-use/graphs/cont...

ithkuil 18 minutes ago | parent

Perhaps we should have README.md and README.ai the latter containing a bunch of stuff that humans don't want to read but agents can use to answer questions humans ask

no-name-here 2 minutes ago | parent

Your comment might be sarcastic, but AGENTS.md already exists for the explicit purpose of agents reading and using it.

ithkuil 20 minutes ago | parent

Hmm, I wonder if we can automate that

petterroea 1 hour ago | parent

It's better that everyone is loud about it then everyone giving up and being silently irritated. At least if people complain it's possible to read the room

IshKebab 48 minutes ago | parent

Clearly the people writing these annoying articles don't know.

I wish they'd just use Astra instead. It doesn't write this awful prose.

slacktivism123 8 minutes ago | parent

>Everybody knows.

Awesome!

I think these repo owners should be required to film themselves reading out their Claudemade readmes with a straight face.

Every honest caveat. Every seam. Every "Frames lie, so the walk prunes hard".

ako 1 hour ago | parent

I thought it was a fine informative readme, starts with the problem, outlines the core of the solutions, and some limitations. Everything I want to know in the first few paragraphs. No need to spend human time to improve it.

jimmySixDOF 1 hour ago | parent

not sure how this is innovative they show the System-1 model can play Doom right in the announcement [1] :

>Doom >We love how this doomo doomonstrates real-time intelligence and what can be doone with code + AI. The engineer behind it was worried about making 10 queries a second (which ends up costing ~$7/hour), but the rest of us agreed that was lower than expected! This is so fun we intend to not only release an in-depth walkthrough, but also host some events to hack on this.

[1] https://typesafe.ai/blog/introducing-system-one-models-and-j...

orbital-decay 1 hour ago | parent

Well nothing about the Doom demo or this entire model is new new either, is it? I don't even think Typesafe themselves are claiming anything novel, they say that they're focusing on practicality instead of chasing big numbers and AGI. Classifiers are older than generative models and are used everywhere. Fast classifiers are used in sampling machinery of every big model and for automation in agentic game plugins for years, except they're usually small and finetuned for the task, not general-use.

I think many people wondered why non-generative models are so underused on a big scale, well here's a long overdue attempt to market that which evidently goes well with people being interested in this again. The field has been captured by the vibe coding and valuation-goes-up hype and a bit. AI has a ton of low hanging fruits that are much more practical than using one tool that gets most attention for everything.

altmanaltman 36 minutes ago | parent

So it's kind of like AI hype growing up and rediscovering its roots because the future isn't futuring soon enough.

Surac 1 hour ago | parent

I did not understand what this is all about. Anyone with more brain than me can explain please?

orbital-decay 1 hour ago | parent

It's a quick proof of concept of computer use powered by Jev, a new general-purpose classifier model. Here it takes the description of the interface (it can't do images yet) and outputs commands. It's faster and many times cheaper than using frontier generative models like GPT to do the same.

smartbit 54 minutes ago | parent

Jev docs gives a good explanation https://docs.typesafe.ai/concepts/system-one

Also see recent discussion https://news.ycombinator.com/item?id=49717558#49718350

aruss 1 hour ago | parent

It seems like the more honest comparison would be to OCR the screen and send that as input to the LLM?

curtisblaine 49 minutes ago | parent

"The honest caveat"