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
1•frabonacci•1h 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.

Does AI Assistance Enhance or Erode Expertise? Evidence from 3-Month Experiment

https://www.nber.org/papers/w35720
1•m-hodges•1m ago•0 comments

Jev example use cases from community

https://jevable.com/
1•theanonymousone•2m ago•0 comments

Checkout offered by banks and credit unions. Alternative to OneLink

https://www.paze.com/
1•dzonga•6m ago•0 comments

Montessori gave us great start. Still switched to public school for first grade

https://www.businessinsider.com/switched-from-montessori-public-school-decision-2026-9
2•theanonymousone•6m ago•0 comments

The paradox at the heart of AI and science – Terence Tao [video]

https://www.youtube.com/watch?v=svl_1upFpQo
1•lr0•7m ago•0 comments

Lawsuit: Illegal Agreement of Anthropic, OpenAI, SpaceXAI, Google on AI Slowdown

https://www.independent.co.uk/news/lawsuit-anthropic-google-openai-lawyers-b3052942.html
3•mdp2021•7m ago•0 comments

Unblock AI Shield – Non-invasive perimeter hardening with zero-disk RAM retentio

https://unblock-shield.vercel.app/?lang=en
1•tendenciaiatufu•9m ago•0 comments

My company let candidates use AI tools in interviews. Here's what I saw

https://www.reddit.com/r/EngineeringManagers/comments/1urw740/my_company_lets_candidates_use_ai_t...
1•bucket2015•10m ago•0 comments

Maximizing Return-on-Feedback

https://avivbenyosef.com/maximizing-return-on-feedback/
1•levlaz•15m ago•0 comments

Post-Quantum Cryptography for Healthcare Professionals

https://www.udemy.com/share/10fqoh3@_zy7B-GA7JTZkdnE0X0rOZnG0NE6YPPmu77H_UQmBQ1fHdvfPwlC9xJeEyAvq...
1•remopulcini•19m ago•0 comments

In 2024 I fine-tuned an LLM. Jev could have removed the side quests

https://kasra.blog/blog/classification-and-jev/
1•jc4p•19m ago•1 comments

AI agent governance patterns as executable decision tables (MIT license)

https://github.com/scubagraham/outthebox-open-patterns
1•scubagraham•22m ago•0 comments

Almost Never Use AI to Write Anything Substantive

https://erichgrunewald.substack.com/p/why-you-should-almost-never-use-ai
5•erwald•23m ago•0 comments

Uber Ordered to Pay $40M to Parents of Passenger After Driver Left Her

https://www.panish.law/2026/09/uber-ordered-to-pay-40-million-to-parents-of-emily-normandin-parke...
2•sensanaty•27m ago•1 comments

CNN, MS Now and Politico Barred from White House After Trump Announces Ban

https://www.nytimes.com/2026/09/19/business/media/trump-white-house-press-ban.html
2•layer8•28m ago•0 comments

Ask HN: ICLR went from 143 submissions to 62,301 submission IDs – what happened?

4•luxshan_t•29m ago•1 comments

Britain created Iraq to control its oil, then installed its own king [video]

https://www.youtube.com/watch?v=Dh7nJNKrDdo
2•teleforce•29m ago•0 comments

Props Are Not a Design System

https://vitonsky.net/blog/2026/09/18/design-system/
3•RohanAdwankar•30m ago•0 comments

Russia reports 'powerful' cyberattacks on second day of parliamentary vote

https://www.aljazeera.com/news/2026/9/19/russia-reports-powerful-cyberattacks-on-second-day-of-pa...
2•mikhael•30m ago•0 comments

Infillion Buys Foursquare

https://app.dealroom.co/news/note/infillion-buys-foursquare-adding-location-data-to-its-ad-platform
4•edent•31m ago•0 comments

A Machine That Can Have Grandchildren

https://artemandreenko.com/blog/a-machine-that-can-have-grandchildren/
2•miovoid•35m ago•1 comments

Show HN: Built a local guardrail layer for Claude Code

https://github.com/ioni-dev/mati
1•fennect•36m ago•0 comments

Regolith Desktop status updates Sep 2026

https://regolith-desktop.com/posts/status_update_09_26/
1•thinkmassive•38m ago•0 comments

Trump announces agreement to secure 'permanent' control over Greenland security

https://www.nbcnews.com/politics/trump-administration/trump-announces-agreement-denmark-secure-pe...
3•ijidak•38m ago•0 comments

I still haven't bought into true RSI

https://www.interconnects.ai/p/where-i-stand-on-rsi
2•pretext•39m ago•0 comments

Periodic Table – Ptable

https://ptable.com/?lang=en
2•thunderbong•42m ago•0 comments

Leaving DigitalOcean, one site at a time

https://keith.is/post/leaving-digitalocean/
26•carnevalem•45m ago•4 comments

China's Electric Aviation Advantage Is a Transport System, Not a Prototype

https://cleantechnica.com/2026/09/19/china-electric-aviation-transport-system/
3•gumby•48m ago•0 comments

A Workaphile's Apology

https://moalquraishi.wordpress.com/2026/08/10/a-workaphiles-apology/
1•maxall4•50m ago•0 comments

Show HN: MNIST64 – 99.5% accuracy MNIST BNN on a stock Commodore 64

https://github.com/jarnoh/mnist64
1•jmagic•52m ago•0 comments