Grok 4.7 Intelligence, Performance and Price Analysis
7 by theanonymousone | 0 comments on Hacker News.
7 by theanonymousone | 0 comments on Hacker News.
Monday, September 21, 2026
Sunday, September 20, 2026
Saturday, September 19, 2026
New top story on Hacker News: Show HN: CUA-S1 – A System One Model for Computer Use
Show HN: CUA-S1 – A System One Model for Computer Use
12 by frabonacci | 0 comments on Hacker News.
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.
12 by frabonacci | 0 comments on Hacker News.
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.
Friday, September 18, 2026
New top story on Hacker News: Show HN: Ax-check.com – Can agents use your product?
Show HN: Ax-check.com – Can agents use your product?
13 by 0x63_Problems | 12 comments on Hacker News.
I'm the co-founder of Gauge, and I built ax-check.com to quickly test how well coding agents can onboard to your product. You'll get a scorecard, specific suggested fixes, and three full coding sessions that show how agents read your site and use your product. I built this because similar checks were too noisy. Most suggested obscure technical changes that don't actually make a difference in agent experience (or AX, hence ax-check.com). This check starts by using DeepSeek 4.1 Flash to try to find key information about your product, starting from the homepage. In actual agent traffic data, we've seen that the key pages are the homepage, llms.txt, pricing, and the docs site (by traffic volume, and by influence), so we focus on those and ignore the rest. We also find that content negotiation for Markdown is legitimately helpful for agents to complete tasks faster and find what they're looking for, so the scan tests that your key pages can serve Markdown. The other key piece is that we run actual coding agents in sandboxes, and have them try to onboard to your product. You can see the full trace and watch it happen live (we kick it off fresh when you enter a new site). We surface interesting findings like hallucinated URLs, inaccurate docs instructions, or product confusion. It also detects whether the agents could complete a fully working onboarding autonomously, without being blocked by a login wall. This is still controversial, but I think finding ways to let agents safely onboard autonomously is going to be table stakes within a year for developer tools in particular. The whole site is agent-friendly itself! You can generally just talk to your coding agent about ax-check.com and it can do the rest. Would really appreciate any feedback to make this useful.
13 by 0x63_Problems | 12 comments on Hacker News.
I'm the co-founder of Gauge, and I built ax-check.com to quickly test how well coding agents can onboard to your product. You'll get a scorecard, specific suggested fixes, and three full coding sessions that show how agents read your site and use your product. I built this because similar checks were too noisy. Most suggested obscure technical changes that don't actually make a difference in agent experience (or AX, hence ax-check.com). This check starts by using DeepSeek 4.1 Flash to try to find key information about your product, starting from the homepage. In actual agent traffic data, we've seen that the key pages are the homepage, llms.txt, pricing, and the docs site (by traffic volume, and by influence), so we focus on those and ignore the rest. We also find that content negotiation for Markdown is legitimately helpful for agents to complete tasks faster and find what they're looking for, so the scan tests that your key pages can serve Markdown. The other key piece is that we run actual coding agents in sandboxes, and have them try to onboard to your product. You can see the full trace and watch it happen live (we kick it off fresh when you enter a new site). We surface interesting findings like hallucinated URLs, inaccurate docs instructions, or product confusion. It also detects whether the agents could complete a fully working onboarding autonomously, without being blocked by a login wall. This is still controversial, but I think finding ways to let agents safely onboard autonomously is going to be table stakes within a year for developer tools in particular. The whole site is agent-friendly itself! You can generally just talk to your coding agent about ax-check.com and it can do the rest. Would really appreciate any feedback to make this useful.
Thursday, September 17, 2026
New top story on Hacker News: Show HN: Craigslist for agent skills, curated by a human
Show HN: Craigslist for agent skills, curated by a human
10 by skeptrune | 4 comments on Hacker News.
hey, i'm nick (@skeptrune on X). i have been using ai for a lot of things i'm not an expert in and thought it would be nice to try and buy paid skills for helping it accomplish those tasks more efficiently. for example, redlining contracts, creating ai generated videos, different website designs, and more. curious to see if this resonates with folks here. i figure a more engineering'y audience that's ai coding-agent forward would have similar problems to what inspired me to make this
10 by skeptrune | 4 comments on Hacker News.
hey, i'm nick (@skeptrune on X). i have been using ai for a lot of things i'm not an expert in and thought it would be nice to try and buy paid skills for helping it accomplish those tasks more efficiently. for example, redlining contracts, creating ai generated videos, different website designs, and more. curious to see if this resonates with folks here. i figure a more engineering'y audience that's ai coding-agent forward would have similar problems to what inspired me to make this
Wednesday, September 16, 2026
Tuesday, September 15, 2026
Monday, September 14, 2026
Sunday, September 13, 2026
Saturday, September 12, 2026
Friday, September 11, 2026
Thursday, September 10, 2026
Wednesday, September 9, 2026
Tuesday, September 8, 2026
Monday, September 7, 2026
New top story on Hacker News: Show HN: Wg-admin – web UI for an existing WireGuard host
Show HN: Wg-admin – web UI for an existing WireGuard host
5 by vsergione | 2 comments on Hacker News.
Reads /etc/wireguard, lets you add/edit/remove peers, applies with wg syncconf (no interface bounce). Does not install WireGuard or rewrite your PostUp/NAT. That's all
5 by vsergione | 2 comments on Hacker News.
Reads /etc/wireguard, lets you add/edit/remove peers, applies with wg syncconf (no interface bounce). Does not install WireGuard or rewrite your PostUp/NAT. That's all
New top story on Hacker News: 216M Spy TVs – The LG Smart TV Problem [video]
216M Spy TVs – The LG Smart TV Problem [video]
169 by treve | 549 comments on Hacker News.
https://ift.tt/EQWoGU7...
169 by treve | 549 comments on Hacker News.
https://ift.tt/EQWoGU7...
Sunday, September 6, 2026
New top story on Hacker News: Ask HN: UK Rescue Rocket Sheds/Houses Information
Ask HN: UK Rescue Rocket Sheds/Houses Information
11 by burnt-resistor | 6 comments on Hacker News.
I can't seem to find pictures or significant details (nothing on Wikipedia) about the coastal buildings purportedly used to contain shipwreck rescue rocket apparatus. Many of the buildings still exist and are Grade II listed. Anyone have any additional anecdotal/local/oral tradition information that isn't found in the little blurbs online on the few websites mentioning them?
11 by burnt-resistor | 6 comments on Hacker News.
I can't seem to find pictures or significant details (nothing on Wikipedia) about the coastal buildings purportedly used to contain shipwreck rescue rocket apparatus. Many of the buildings still exist and are Grade II listed. Anyone have any additional anecdotal/local/oral tradition information that isn't found in the little blurbs online on the few websites mentioning them?