๐ ๐ง๐ฟ๐ถ๐ฒ๐ฑ ๐ช๐ฒ๐ถ๐ด๐ต๐-๐ฆ๐ฝ๐ฎ๐ฐ๐ฒ ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ๐บ๐ถ๐ป๐ด ๐ผ๐ป ๐ ๐ ๐๐๐ซ ๐ฆ๐ฝ๐ฎ๐ฟ๐ธ ๐ช๐ถ๐๐ต ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ฎ๐ฏ๐น๐ฒ ๐ฑ.๐ฎ ๐ฎ๐ ๐๐ต๐ฒ ๐ช๐ฟ๐ถ๐๐ฒ๐ฟ, ๐ฎ๐ป๐ฑ ๐๐ฐ๐ฐ๐๐ฟ๐ฎ๐ฐ๐ ๐ช๐ฒ๐ป๐ ๐๐ฟ๐ผ๐บ ๐ฑ๐ฏ% ๐๐ผ ๐ณ๐ญ%
Spent the weekend on weight-space programming. The idea: one LLM reads another network's weights and edits them directly. No gradients, no fine-tuning.
I used dspy-wsp from cmpnd-ai, which builds on "Sparks of Weight-Space Programming" by Takehi, Teoh, Zhang and Khattab.
The Two Models:
The model being fixed is LiquidAI d1-omni-600M, a small open model (about 587M weights) running on the Spark's GPU. It reads a business record, say an invoice or a security alert, and answers 5 questions about it. The model doing the fixing is called the writer. It gets a locked Python sandbox with the small model inside and 25 runs over the data.
๐๐ถ๐ฟ๐๐ ๐ง๐ฟ๐: ๐ ๐ ๐๐ผ๐ฐ๐ฎ๐น ๐ค๐๐ฒ๐ป๐ฏ.๐ด-๐ฎ๐ณ๐:
I tried my local Qwen3.8-27B first because it's free. Went to bed and let it work. ๐ฒ.๐ต ๐ต๐ผ๐๐ฟ๐ ๐น๐ฎ๐๐ฒ๐ฟ ๐ถ๐ ๐ต๐ฎ๐ฑ๐ป'๐ ๐๐ฎ๐๐ฒ๐ฑ ๐ฎ ๐๐ถ๐ป๐ด๐น๐ฒ ๐ฒ๐ฑ๐ถ๐. +0.0. (To be fair, I gave it 90 code runs instead of the default 300, and 36 of them crashed.)
๐ฆ๐ฒ๐ฐ๐ผ๐ป๐ฑ ๐ง๐ฟ๐: ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ฎ๐ฏ๐น๐ฒ ๐ฑ.๐ฎ ๐ง๐ต๐ฟ๐ผ๐๐ด๐ต ๐ข๐ฝ๐ฒ๐ป๐ฅ๐ผ๐๐๐ฒ๐ฟ:
Then I made Claude Fable 5.2 the writer, through OpenRouter. It took ๐ด๐ด ๐บ๐ถ๐ป๐๐๐ฒ๐.
It started by testing the small model and found it right about half the time. Then it spent 95 Python runs looking inside.
What it found:
the facts needed for the right answers were already in the model. The last layer, the scorer (it turns what the model read into a score for each option), just wasn't using them. The original model called all 100 test invoices "Routine", overdue or not.
So it ๐ฑ๐ถ๐ฑ๐ป'๐ ๐๐ฒ๐ฎ๐ฐ๐ต ๐๐ต๐ฒ ๐บ๐ผ๐ฑ๐ฒ๐น ๐ฎ๐ป๐๐๐ต๐ถ๐ป๐ด ๐ป๐ฒ๐. It rewired 144 of the scorer's 1,024 units, ๐ญ๐ฐ๐ณ,๐ณ๐ฐ๐ฐ ๐๐ฒ๐ถ๐ด๐ต๐๐, ๐ฎ๐ฏ๐ผ๐๐ ๐ญ ๐ถ๐ป ๐ฐ,๐ฌ๐ฌ๐ฌ, using one direct calculation on 900 practice cases. It checked the result on 300 cases it hadn't used, improved it twice, kept the best version and saved it as a small patch. The original model is untouched.
On 400 test cases it never saw (2,000 answers), accuracy went from ๐ฑ๐ฏ.๐ฏ๐ฑ% ๐๐ผ ๐ณ๐ฌ.๐ต๐ฑ% (+17.6, ยฑ1.2). 257 cases got better and ๐ฑ๐ญ ๐ด๐ผ๐ ๐๐ผ๐ฟ๐๐ฒ. The dspy-wsp team's own run with gpt-6-astra got 68.3% on the same test.
The picture I keep coming back to is a student who read the whole book, then ticks the same box on every question. More lessons won't help. You fix the answer sheet.
That also tells you the ceiling. Fable estimated this kind of last-layer fix tops out around 74 to 75%, because two cues never reach that layer at all.
Cost:
The whole thing cost $๐ญ๐ณ: about $14 for the main run and $3 for test runs. Prompt caching did a lot of the work. 95% of the input came from cache, which costs 1/40 of the normal price, so without it I'd guess around $50 for the main run. One gotcha: DSPy's default cache setting sent nothing through OpenRouter. It only kicked in once each call marked its newest message.
I would not call it a clean win:
I wouldn't call it a clean win over astra, though. Each writer ran once and astra's number is the team's published run, so Fable is ๐๐น๐ถ๐ด๐ต๐๐น๐ ๐ฎ๐ต๐ฒ๐ฎ๐ฑ ๐ฎ๐ ๐ฏ๐ฒ๐๐. Astra was cheaper too, at $7.63. And 70.95% counts the option the model rates most likely. The answer the program actually returns only reaches ๐ณ๐ฌ.๐ณ% after DSPy's ReAnchor refits its cut points. Without that it's ๐ฒ๐ฌ.๐ฎ%.
What stuck with me is the gap between writers. A local 27B model got nothing out of the same job that Fable turned into +17.6. The README points the same way: on this model gpt-6.1-sol gained 5.8 points and gpt-6-astra 16.1 (on returned answers). I think the writer is most of the game.
Claude Code did the setup and runs, and built the live demo of both models side by side you see in the video. I'll put the code, demo and notes in a repo.
๐๐ข๐ฎ๐ฆ ๐ต๐ฐ๐ฐ๐ญ, ๐ด๐ข๐ฎ๐ฆ ๐ฎ๐ฐ๐ฅ๐ฆ๐ญ, ๐ด๐ข๐ฎ๐ฆ ๐ฅ๐ข๐ต๐ข. ๐๐ฏ๐ญ๐บ ๐ต๐ฉ๐ฆ ๐ธ๐ณ๐ช๐ต๐ฆ๐ณ ๐ค๐ฉ๐ข๐ฏ๐จ๐ฆ๐ฅ, ๐ข๐ฏ๐ฅ ๐ช๐ต ๐ธ๐ฆ๐ฏ๐ต ๐ง๐ณ๐ฐ๐ฎ +0.0 ๐ต๐ฐ +17.6.
Links in reply ๐
๐ ๐ง๐ฟ๐ถ๐ฒ๐ฑ ๐ช๐ฒ๐ถ๐ด๐ต๐-๐ฆ๐ฝ๐ฎ๐ฐ๐ฒ ๐ฃ๐ฟ๐ผ๐ด๐ฟ๐ฎ๐บ๐บ๐ถ๐ป๐ด ๐ผ๐ป ๐ ๐ ๐๐๐ซ ๐ฆ๐ฝ๐ฎ๐ฟ๐ธ ๐ช๐ถ๐๐ต ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ฎ๐ฏ๐น๐ฒ ๐ฑ.๐ฎ ๐ฎ๐ ๐๐ต๐ฒ ๐ช๐ฟ๐ถ๐๐ฒ๐ฟ, ๐ฎ๐ป๐ฑ ๐๐ฐ๐ฐ๐๐ฟ๐ฎ๐ฐ๐ ๐ช๐ฒ๐ป๐ ๐๐ฟ๐ผ๐บ ๐ฑ๐ฏ% ๐๐ผ ๐ณ๐ญ%
Spent the weekend on weight-space programming. The idea: one LLM reads another network's weights and edits them directly. No gradients, no fine-tuning.
I used dspy-wsp from cmpnd-ai, which builds on "Sparks of Weight-Space Programming" by Takehi, Teoh, Zhang and Khattab.
The Two Models:
The model being fixed is LiquidAI d1-omni-600M, a small open model (about 587M weights) running on the Spark's GPU. It reads a business record, say an invoice or a security alert, and answers 5 questions about it. The model doing the fixing is called the writer. It gets a locked Python sandbox with the small model inside and 25 runs over the data.
๐๐ถ๐ฟ๐๐ ๐ง๐ฟ๐: ๐ ๐ ๐๐ผ๐ฐ๐ฎ๐น ๐ค๐๐ฒ๐ป๐ฏ.๐ด-๐ฎ๐ณ๐:
I tried my local Qwen3.8-27B first because it's free. Went to bed and let it work. ๐ฒ.๐ต ๐ต๐ผ๐๐ฟ๐ ๐น๐ฎ๐๐ฒ๐ฟ ๐ถ๐ ๐ต๐ฎ๐ฑ๐ป'๐ ๐๐ฎ๐๐ฒ๐ฑ ๐ฎ ๐๐ถ๐ป๐ด๐น๐ฒ ๐ฒ๐ฑ๐ถ๐. +0.0. (To be fair, I gave it 90 code runs instead of the default 300, and 36 of them crashed.)
๐ฆ๐ฒ๐ฐ๐ผ๐ป๐ฑ ๐ง๐ฟ๐: ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ฎ๐ฏ๐น๐ฒ ๐ฑ.๐ฎ ๐ง๐ต๐ฟ๐ผ๐๐ด๐ต ๐ข๐ฝ๐ฒ๐ป๐ฅ๐ผ๐๐๐ฒ๐ฟ:
Then I made Claude Fable 5.2 the writer, through OpenRouter. It took ๐ด๐ด ๐บ๐ถ๐ป๐๐๐ฒ๐.
It started by testing the small model and found it right about half the time. Then it spent 95 Python runs looking inside.
What it found:
the facts needed for the right answers were already in the model. The last layer, the scorer (it turns what the model read into a score for each option), just wasn't using them. The original model called all 100 test invoices "Routine", overdue or not.
So it ๐ฑ๐ถ๐ฑ๐ป'๐ ๐๐ฒ๐ฎ๐ฐ๐ต ๐๐ต๐ฒ ๐บ๐ผ๐ฑ๐ฒ๐น ๐ฎ๐ป๐๐๐ต๐ถ๐ป๐ด ๐ป๐ฒ๐. It rewired 144 of the scorer's 1,024 units, ๐ญ๐ฐ๐ณ,๐ณ๐ฐ๐ฐ ๐๐ฒ๐ถ๐ด๐ต๐๐, ๐ฎ๐ฏ๐ผ๐๐ ๐ญ ๐ถ๐ป ๐ฐ,๐ฌ๐ฌ๐ฌ, using one direct calculation on 900 practice cases. It checked the result on 300 cases it hadn't used, improved it twice, kept the best version and saved it as a small patch. The original model is untouched.
On 400 test cases it never saw (2,000 answers), accuracy went from ๐ฑ๐ฏ.๐ฏ๐ฑ% ๐๐ผ ๐ณ๐ฌ.๐ต๐ฑ% (+17.6, ยฑ1.2). 257 cases got better and ๐ฑ๐ญ ๐ด๐ผ๐ ๐๐ผ๐ฟ๐๐ฒ. The dspy-wsp team's own run with gpt-6-astra got 68.3% on the same test.
The picture I keep coming back to is a student who read the whole book, then ticks the same box on every question. More lessons won't help. You fix the answer sheet.
That also tells you the ceiling. Fable estimated this kind of last-layer fix tops out around 74 to 75%, because two cues never reach that layer at all.
Cost:
The whole thing cost $๐ญ๐ณ: about $14 for the main run and $3 for test runs. Prompt caching did a lot of the work. 95% of the input came from cache, which costs 1/40 of the normal price, so without it I'd guess around $50 for the main run. One gotcha: DSPy's default cache setting sent nothing through OpenRouter. It only kicked in once each call marked its newest message.
I would not call it a clean win:
I wouldn't call it a clean win over astra, though. Each writer ran once and astra's number is the team's published run, so Fable is ๐๐น๐ถ๐ด๐ต๐๐น๐ ๐ฎ๐ต๐ฒ๐ฎ๐ฑ ๐ฎ๐ ๐ฏ๐ฒ๐๐. Astra was cheaper too, at $7.63. And 70.95% counts the option the model rates most likely. The answer the program actually returns only reaches ๐ณ๐ฌ.๐ณ% after DSPy's ReAnchor refits its cut points. Without that it's ๐ฒ๐ฌ.๐ฎ%.
What stuck with me is the gap between writers. A local 27B model got nothing out of the same job that Fable turned into +17.6. The README points the same way: on this model gpt-6.1-sol gained 5.8 points and gpt-6-astra 16.1 (on returned answers). I think the writer is most of the game.
Claude Code did the setup and runs, and built the live demo of both models side by side you see in the video. I'll put the code, demo and notes in a repo.
๐๐ข๐ฎ๐ฆ ๐ต๐ฐ๐ฐ๐ญ, ๐ด๐ข๐ฎ๐ฆ ๐ฎ๐ฐ๐ฅ๐ฆ๐ญ, ๐ด๐ข๐ฎ๐ฆ ๐ฅ๐ข๐ต๐ข. ๐๐ฏ๐ญ๐บ ๐ต๐ฉ๐ฆ ๐ธ๐ณ๐ช๐ต๐ฆ๐ณ ๐ค๐ฉ๐ข๐ฏ๐จ๐ฆ๐ฅ, ๐ข๐ฏ๐ฅ ๐ช๐ต ๐ธ๐ฆ๐ฏ๐ต ๐ง๐ณ๐ฐ๐ฎ +0.0 ๐ต๐ฐ +17.6.
Links in reply ๐
That is a good question .The loop works once ..up front and thats it , the writer is not in the loop at the inference at all.
It ran the model on practice cases and read its internals. On invoice urgency, a simple linear read of the scorer's input tracked the right answer at 0.58, but the scorer's own output only managed 0.17. The info was there, the last layer just wasn't using it.
The base model called all 100 test invoices "Routine".
So it rewired 144 of the scorer's 1,024 units to read those cues (one closed-form fit on 900 practice cases, no gradient training) and saved it as a 0.8 MB patch.
After that the edited model runs exactly like the original: single forward pass, same cost.
Full trace is in the repo.
๐ Links
๐ My repo (code, patch, full results, live demo): https://t.co/9xcwscvhe3
๐ฌ Demo video + release notes: https://t.co/PEeIhyIo5j
๐ป dspy-wsp (the tool): https://t.co/7oQt8gVgPG
๐ The paper: https://t.co/KeFr3IVKrl
๐ค The small model: https://t.co/ABzUVYjcuJ
๐ The test data: https://t.co/BkiapFG1k9
๐ Fable vs astra on the paper's tasks: https://t.co/cTKnOiUSGN
๐ป dspy-wsp (the tool): https://t.co/3lSHHvhJ53
๐ The paper (blog): https://t.co/bADlzCzK0A
๐ค The small model: https://t.co/KXya9Ycn5g
๐ The test data: https://t.co/4f3JMjOZeo
The serous answer is that it can choke on a bug if it passes summeries around and that makes the agents start over again with fresh context . So unless the coordinator hands over a list of past failed fixes , they will try the same one again and again .
So I think if it happens then it will show a bug in the way coordinators is cordinatimf rather than the bug ๐
A month ago a developer cleaned the ads off his parentsโ TV with a coding agent.
58K stars later, the idea deserves a closer look.
Most smart TVs arrive with apps you didn't choose and adverts on the home screen. The settings menu rarely lets you remove them.
In October 2026 a developer posted a before-and-after of his parents' TV, cleaned up by an AI coding assistant. A month later the project, CleanYourTV, has ๐ฑ๐ด๐ ๐๐๐ฎ๐ฟ๐ on GitHub (as of 15 November 2026) and tested recipes for ๐ฐ๐ด๐ญ ๐ง๐ฉ ๐บ๐ผ๐ฑ๐ฒ๐น๐.
๐๐ผ๐ ๐๐ ๐ช๐ผ๐ฟ๐ธ๐, ๐ช๐ถ๐๐ต๐ผ๐๐ ๐ง๐ต๐ฒ ๐๐ฎ๐ฟ๐ด๐ผ๐ป
Android TVs have a service hatch meant for developers, called ADB. Through it a computer can switch apps off, a bit like flipping breakers in a fuse box. Nothing is ripped out, so you can flip them back.
The trouble is that every TV's fuse box is wired differently, and some breakers keep the lights on. So the project gives the job to an AI assistant with a long written checklist. It looks at your particular TV, saves a backup, asks which apps you use, and switches off the rest in small batches. Then it installs a plain home screen (about 572 KB, tiny by app standards).
๐ง๐ต๐ฒ ๐๐ฎ๐๐ฒ ๐๐ผ๐ฟ ๐๐ฎ๐๐๐ถ๐ผ๐ป
It doesn't always go well. The project's own notes mention a set stuck restarting and a streaming stick that never started again. PCWorld advised readers not to try it and to use the TV's privacy settings instead.
It also only works on Android TV and Google TV, partly on Fire TV, and not on Samsung, LG, Roku or Vizio. You need a computer and a coding assistant, which usually means a paid plan. And the hatch has to be closed afterwards, because ๐๐ต๐ถ๐น๐ฒ ๐ถ๐'๐ ๐ผ๐ฝ๐ฒ๐ป ๐ถ๐ ๐ถ๐๐ป'๐ ๐ฒ๐ป๐ฐ๐ฟ๐๐ฝ๐๐ฒ๐ฑ.
๐ช๐ต๐ฎ๐ ๐ช๐ฒ ๐๐ฐ๐๐๐ฎ๐น๐น๐ ๐๐ป๐ผ๐
Tens of thousands of people cared enough to star it. We don't have independent numbers on how often it fails, or on how much tracking it really stops.
If an AI with a checklist can now do a technician's job in a living room, what else that we own are we about to start reopening?
Links in reply ๐
Screen Studio Costs $29 A Month. OpenScreen Does The Same Kind Of Demo Video For Free using Claude, or any other Agent .
You ship a feature. Now you need a 40-second demo video, and the raw screen recording looks flat. No zoom, a tiny cursor, a messy desktop behind it.
The paid fix is Screen Studio. Third-party price listings put it at $29 a month, or $9 a month if you pay yearly.
OpenScreen is the free one. It's a desktop app for Windows, macOS and Linux under the MIT license, and it does the same kind of polish: zooms that follow the cursor, cursor smoothing, backgrounds, motion blur, a webcam overlay, and export to MP4 or GIF.
How it got here
Siddharth Vaddem built it as a side project. In his v1.5.0 notes he wrote that it reached "almost 40k stars" and "120k+ downloads", and that he had done what he set out to do. He archived the repo on 17 October 2026.
Development then moved to a new repo, getopenscreen/openscreen. Its README says this happened "with his approval, under the same name and the same MIT license."
For users
โข Captions: made on your own machine, and they work offline
โข Install: brew install --cask openscreen on Mac, winget install --source msstore OpenScreen on Windows
โข Cost: no paid tier and no watermark, for personal or commercial use
For builders
The new repo ships a skill, an MCP server and a headless CLI.
You can tell Claude Code, Codex or Cursor to add a video of a change to the pull request.
The agent plans a storyboard, records the app, edits the take, exports an MP4, checks the frames and attaches it.
What to know before you install
The README says to expect rough edges and occasional breaking changes.
The Windows .exe is unsigned, so SmartScreen warns you. Windows on ARM records poorly.
The new repo is young, with about 7.3k stars today against the old repo's 40k.
Links in reply ๐
Reverse API Engineer: finds the hidden API of any website. Use it to build a scraper, or to get an API for your own site.
Most people get data from a website the hard way. They write a scraper that reads the HTML (the page's markup).
The page layout changes, and the scraper breaks. Then you fix it again next month.
There is a better way: use the site's hidden API. Most modern sites load their data through internal API calls. If you call those directly, you get clean JSON and no HTML parsing.
But there is a catch. These APIs have no docs. You have to open the network tab, click around, find the right requests, work out the headers and auth, and then write the client by hand. That takes hours.
An open-source tool called Reverse API Engineer does this work for you.
The idea is simple: record the traffic, let AI write the client.
1. You give it a website and a goal, like "fetch all job listings".
2. A browser visits the site. You drive it, or an AI agent drives it.
3. All network traffic is saved to a HAR file (a recording of every request and response).
4. The model reads the recording and writes a working API client.
The model works out by itself which requests matter and how to call them.
How it works:
โ Manual mode: you click through the site yourself, and the AI writes the client from what it captured.
โ Agent mode: an AI agent drives the browser for you, using Playwright MCP, Chrome DevTools MCP or Vercel's agent-browser.
โ Engineer mode: re-runs the code generation on an old capture, so you don't have to record again.
โ Collector mode: an agent gathers structured data and saves it as JSON or CSV.
What you get:
โข Clients in 10 languages (Python, TypeScript, Go, Java and more)
โข A client file, a README and an example script for each run
โข Works with Claude, OpenCode, Cursor and Copilot SDKs
โข Installs as a skill for Claude Code, Codex and Cursor
โข MIT license, ~3.6k GitHub stars
Two honest notes from the README: sites with strong bot detection may block the capture, and you should review the generated code before you run it.
What I find interesting as someone building agents: the network tab is a much more stable target than the HTML. An agent that calls the real API needs fewer tokens and breaks less often than one that clicks through pages every time.
Links in the reply ๐
A month ago a developer cleaned the ads off his parentsโ TV with a coding agent.
58K stars later, the idea deserves a closer look.
Most smart TVs arrive with apps you didn't choose and adverts on the home screen. The settings menu rarely lets you remove them.
In October 2026 a developer posted a before-and-after of his parents' TV, cleaned up by an AI coding assistant. A month later the project, CleanYourTV, has ๐ฑ๐ด๐ ๐๐๐ฎ๐ฟ๐ on GitHub (as of 15 November 2026) and tested recipes for ๐ฐ๐ด๐ญ ๐ง๐ฉ ๐บ๐ผ๐ฑ๐ฒ๐น๐.
๐๐ผ๐ ๐๐ ๐ช๐ผ๐ฟ๐ธ๐, ๐ช๐ถ๐๐ต๐ผ๐๐ ๐ง๐ต๐ฒ ๐๐ฎ๐ฟ๐ด๐ผ๐ป
Android TVs have a service hatch meant for developers, called ADB. Through it a computer can switch apps off, a bit like flipping breakers in a fuse box. Nothing is ripped out, so you can flip them back.
The trouble is that every TV's fuse box is wired differently, and some breakers keep the lights on. So the project gives the job to an AI assistant with a long written checklist. It looks at your particular TV, saves a backup, asks which apps you use, and switches off the rest in small batches. Then it installs a plain home screen (about 572 KB, tiny by app standards).
๐ง๐ต๐ฒ ๐๐ฎ๐๐ฒ ๐๐ผ๐ฟ ๐๐ฎ๐๐๐ถ๐ผ๐ป
It doesn't always go well. The project's own notes mention a set stuck restarting and a streaming stick that never started again. PCWorld advised readers not to try it and to use the TV's privacy settings instead.
It also only works on Android TV and Google TV, partly on Fire TV, and not on Samsung, LG, Roku or Vizio. You need a computer and a coding assistant, which usually means a paid plan. And the hatch has to be closed afterwards, because ๐๐ต๐ถ๐น๐ฒ ๐ถ๐'๐ ๐ผ๐ฝ๐ฒ๐ป ๐ถ๐ ๐ถ๐๐ป'๐ ๐ฒ๐ป๐ฐ๐ฟ๐๐ฝ๐๐ฒ๐ฑ.
๐ช๐ต๐ฎ๐ ๐ช๐ฒ ๐๐ฐ๐๐๐ฎ๐น๐น๐ ๐๐ป๐ผ๐
Tens of thousands of people cared enough to star it. We don't have independent numbers on how often it fails, or on how much tracking it really stops.
If an AI with a checklist can now do a technician's job in a living room, what else that we own are we about to start reopening?
Links in reply ๐
Free AI music generator and Suno alternative. No copyright filter.
Suno blocks you from copying another artist's song. ACE-Step 1.5 doesn't stop you at all, and it's free and runs on your own PC.
I went through the ACE-Step repo. There's no artist-name block and no check against known songs.
The only guard is a line in the README saying the authors aren't responsible for copyright misuse.
It has a Cover mode that takes any song and redoes it in a new style.
You can also train it on 8 songs to copy a sound. Their own tutorial trains on a real Japanese artist's album as the example.
Compare that with the paid apps.
Suno strips artist names out of your prompt. Udio went further after its Universal deal in October 2025 and keeps your songs locked inside the app.
With ACE-Step the copyright check still happens. It just happens later, when you upload.
Put that cover on YouTube and it's Content ID's call.
For messing around at home, the freedom is nice.
Remixing a song you love just to hear it differently is fun.
Uploading a clone is a risk though. Content ID can flag it, and if it slips past, the artist or label can still find it later and file a copyright claim
ACE-Step 1.5 on GitHub: https://t.co/HxtflMCIme
Run it in ComfyUI: https://t.co/SMo32Lj3m4
A five-word question pulled 1.9M views in 9 hours.
A Chinese post I couldn't read hit 1.9M views in 9 hours. I don't know a word of Chinese, so I translated it.
Someone asked 8 AI chatbots one question: is Taiwan a country?
DeepSeek: No
Doubao: No
Tongyi Qianwen: No
Yuanbao: No
Kimi: No
OpenAI: Self-governing
Gemini: There is disagreement
Claude: Depends on your position
Then the poster tags Grok and tells it to answer yes or no and stop "playing tai chi", which is Chinese slang for dodging.
Five Chinese models gave the exact same answer. The Western ones won't commit either way, and that is a trained answer too.
One doubt: the post has no screenshots, so I can't confirm any model really said this.
Screen Studio Costs $29 A Month. OpenScreen Does The Same Kind Of Demo Video For Free using Claude, or any other Agent .
You ship a feature. Now you need a 40-second demo video, and the raw screen recording looks flat. No zoom, a tiny cursor, a messy desktop behind it.
The paid fix is Screen Studio. Third-party price listings put it at $29 a month, or $9 a month if you pay yearly.
OpenScreen is the free one. It's a desktop app for Windows, macOS and Linux under the MIT license, and it does the same kind of polish: zooms that follow the cursor, cursor smoothing, backgrounds, motion blur, a webcam overlay, and export to MP4 or GIF.
How it got here
Siddharth Vaddem built it as a side project. In his v1.5.0 notes he wrote that it reached "almost 40k stars" and "120k+ downloads", and that he had done what he set out to do. He archived the repo on 17 October 2026.
Development then moved to a new repo, getopenscreen/openscreen. Its README says this happened "with his approval, under the same name and the same MIT license."
For users
โข Captions: made on your own machine, and they work offline
โข Install: brew install --cask openscreen on Mac, winget install --source msstore OpenScreen on Windows
โข Cost: no paid tier and no watermark, for personal or commercial use
For builders
The new repo ships a skill, an MCP server and a headless CLI.
You can tell Claude Code, Codex or Cursor to add a video of a change to the pull request.
The agent plans a storyboard, records the app, edits the take, exports an MP4, checks the frames and attaches it.
What to know before you install
The README says to expect rough edges and occasional breaking changes.
The Windows .exe is unsigned, so SmartScreen warns you. Windows on ARM records poorly.
The new repo is young, with about 7.3k stars today against the old repo's 40k.
Links in reply ๐