frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Show HN: CUA-S1 – A System One Model for Computer Use

https://github.com/trycua/cua
14•frabonacci•3h ago
Hello HN! We're Dillon and Francesco from Cua.

We were wondering how many computer use tasks actually need a full general purpose LLM (e.g. gpt-6-astra, claude-opus-5 etc.) to think through all their decisions and steps. Some tasks require thinking about a plan, exploring different paths, recovering from failure. Other tasks are a question of making local decisions, like this value should go in this box, or should I check this box, or this element should be ignored.

We wondered how far we could go with a small model trained to only make these kinds of decisions.

Our inspiration was Typesafe's Jev and its System One Model framing. This is a nod to the dichotomy between thinking quickly, automatically, and intuitively (system 1) vs. thinking slowly, analytically (system 2), as described by Daniel Kahneman.

The interesting question for us was: what happens if you give a model an interface of current context, and a set of possible choices, and you ask it to return a probability for each choice? This kind of model does not generate output token by token like most LLMs do, but rather scores the options you give it, which you can check, trust, and use to drive your app's behavior.

CUA-S1 is our answer for narrow, specialized decision models for computer use. Our first release is CUA-S1-FORMS. We built this from ideas and code in jevlike, and then trained a second model just to handle form interactions. It has 706k parameters, and the original checkpoint is 2.8 MB.

The first training iteration took less than 30 minutes on synthetic data. Given a set of structured elements and values extracted from a document, it predicts whether to use the given value, CHECK, CLICK, or SKIP for each element. It does not predict new values for text fields, and does not consider screenshots. Element decisions are scored together, and your code can order the actions, and Cua Driver will execute them one at a time.

A first evaluation of this specialist vs. hosted Jev on our form task:

- For the whole decision set: 99.7% correct vs 83.6%.

- For the subset of steps that require an action: 100% correct vs 96%.

- For the subset of steps that are just leaving already-filled fields alone: 100% correct vs 74%.

The specialist was trained specifically for this task and convention (just press skip for already filled boxes), while hosted Jev has not been fine-tuned for it, so this is an experiment in scoped specialization.

We measured 7-9 ms to score a form locally vs. 260-280 ms per call to hosted Jev including network latency, though those samples measure different things and are not end-to-end form completion times.

Our interest here is in the space between a brittle script and a general agent loop. The content and layout of form fields vary enough that scripts get unwieldy, but the set of available decisions can remain narrow and well scoped. We want to explore the possibility of a general agent encountering something novel, and passing well understood decisions over to specialists like this.

That is a direction we are looking into. The current release is for forms only. We're open sourced the synthetic data generation, training, evaluation, and Driver integration under libs/cua-s1 with an MIT license.

Comments welcome! Especially if you are building computer-use agents and have run into a recurring decision that is too variable to script but is too narrow to call another LLM for.

Comments

badatnames•7m ago
Please give us a hyper general cookie consent popup dismissal add-on :)

Show HN: CUA-S1 – A System One Model for Computer Use

https://github.com/trycua/cua
14•frabonacci•3h ago•1 comments

Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash

https://cactuscompute.com/needle
220•HenryNdubuaku•1d ago•90 comments

Show HN: BreachScanner – Check if your email has been in data breaches

https://github.com/Aaronmike481/BreachScanner
2•thewalnut124•2h ago•3 comments

Show HN: An e-ink frame that hears birds and draws them as 1800s illustrations

https://github.com/arnegiacomo/fugleramme
2343•arnemunthekaas•4d ago•259 comments

Show HN: I wrote a custom assembler for CHIP-8 in C++

https://github.com/Tackx/c8-ass
24•tack1234•2d ago•4 comments

Show HN: CrabBoss, radio automation software

https://crabboss.vercel.app
2•sonyarianto•4h ago•0 comments

Show HN: 3D map of which newsletters link to the same things

https://reletter.com/graph
2•drpancake•1h ago•0 comments

Show HN: Scry, programmable internet search w/ congestion pricing

https://scry.io/
59•Xyra•1d ago•23 comments

Show HN: Seal – Letters and passwords that open for your family after you die

https://github.com/jasonepage/Seal
10•jpage2•15h ago•7 comments

Show HN: Ax-check.com – Can agents use your product?

https://www.ax-check.com/
33•0x63_Problems•2d ago•39 comments

Show HN: Jeff – A read-only CLI for semantic code review using Jev

https://github.com/Alurith/jeff
26•imalessandro•1d ago•4 comments

Show HN: Share your AI Setup, Learn from others

https://mysetup.ai/
237•steveybrown•2d ago•136 comments

Show HN: Snapdrop: Instantly share files between devices. No setup, no signup

https://snapdrop.me
107•Capira•1d ago•48 comments

Show HN: LiveWorld – Every 24/7 YouTube live camera on one globe

https://liveworld.info/
37•harisingh1612•19h ago•36 comments

Show HN: Hacking a $20 4G wireless hotspot into a texting device

https://bkovac.github.io/modem-thing/
208•bobili1234•4d ago•37 comments

Show HN: Craigslist for agent skills, curated by a human

https://skillbay.sh/
26•skeptrune•2d ago•18 comments

Show HN: AttaLambda: a language where types and data are made of untyped lambdas

https://attalambda.com
47•kserrec•5d ago•7 comments

Show HN: Capsule – Single-file web apps that save their data into SQLite

https://withcapsule.app/
377•bashtian•4d ago•167 comments

Show HN: Procedurally Generated Mineral Gems

https://gems.0x0000007a.com/
3•Dramatize•1d ago•0 comments

Show HN: I built a new version of my fun spatial 3D online meeting app

https://flat.social
97•pawelwentpawel•2d ago•52 comments

Show HN: How Stale Is Your AI? Release age and training cutoff for 20 models

https://stale.jock.pl/
80•joozio•3d ago•49 comments

Show HN: Free game to destroy any web page

https://page-rage.com/
18•alexreardon•1d ago•7 comments

Show HN: An OSS Python dependency scanner for exploited, unmaintained packages

https://github.com/binuka200/package-doctor
3•binukajayaweera•19h ago•3 comments

Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS

https://www.aclif.ai/
34•chris_marino•2d ago•17 comments

Show HN: OJ – A drop-in replacement for Vite in Rust

https://github.com/lovablelabs/oj
17•h1fra•1d ago•1 comments

Show HN: Pizza Bot – An inbox for AI agents that work in the background

https://github.com/pizza-bot-app/pizza-bot
60•jd_•4d ago•36 comments

Show HN: TypeSeer – on-device autocomplete for every text field on macOS

https://typeseer.com/
4•animateus•22h ago•1 comments

Show HN: Microsoft Office running with Wine on Linux with no virtualization

https://github.com/Tombert/office365_flake
88•tombert•1d ago•86 comments

Show HN: Macros with a Behringer FCB1010 MIDI Pedalboard in macOS

https://github.com/JamesRyanATX/fcbnerd
85•fretlessjazz•4d ago•21 comments

Show HN: Go Implementation of Systemd Time

https://gitlab.com/allddd/go-systemd-time
4•allddd•1d ago•0 comments