
@eeexevi
iAccount based inIndonesia
About this account
- Account based in
- Indonesia
- Connected via
- Web
Account-level information from X, not a live location or the device used for a specific post.
/tmp/universe/milky_way/solar_system/earth.sys
Joined August 2020
- Tweets1.6K
- Following334
- Followers214
- Likes2.8K
Thanks to onnxruntime-react-native, react-native-skia, and @PaddlePaddle 's paddle ocr, now [email protected] added support to run the OCR natively on mobile! Local-first & offline!
The PP-OCRv6 is also supported now. Check out the demo repo: github.com/PT-Perkasa-Pilar-…
VEXEEE retweeted
Introducing Amos 1, four open-weight decision models from 0.5B to 27B.
> 94.9% on Decision Bench, ahead of Perplexity, Gemini and Jev
> beats Jev by 10 points on JevBench hard
> reads photos and answers with boxes
> decisions in 27 to 122 ms, free API, Apache 2.0
VEXEEE retweeted
Introducing Indra, a 1M parameter OCR model:
- 49 pages/s on CPU, 11x faster than PaddleOCR and 17x Tesseract
- Beats them on OmniParseBench text (41.3 vs 35.8 and 19.3)
- Full layout and line bounding boxes
VEXEEE retweeted
Zodのリポジトリに、マルウェアっぽいコミットが見つかったというHNのスレッドから。
自分もTypeScriptのバリデーションはほぼZod頼みだから、他人事じゃない。依存上げるときに差分まで見てるかって言われると、正直怪しいかも。
・対象は「Fix bugs and improve performance」というコミット。バージョンを4.6.5から4.6.6に上げつつ、meta.test.mjsを追加していた
・中身はEthereumのスマートコントラクトからC2のホスト名を取り、loaderを落として実行するステージ1ローダ
・package.jsonにpreinstallで「node meta.test.mjs」を仕込む形
・ただし公開されたnpmの[email protected]には、そのファイルもpreinstallも入っていない。SLSAのprovenance付きで、tarballもクリーンと確認できる
・履歴上は後からCleanup bad commitsで巻き戻され、別経路で4.6.6が出し直された、という見方がHNでも広がっている
・一方で「なぜこんな露骨な差分が通ったのか」「メンテナのアカウントは無事か」という不安は残る
サプライチェーン攻撃の未遂か、途中で止められたのかはまだ確定していない。でもコミット履歴を見ただけで止まる世界になってきた感覚はある。
github.com/colinhacks/zod/co…
VEXEEE retweeted
"The purpose of software engineering is to control complexity, not to create it." – Pamela Zave
More true now than ever.
VEXEEE retweeted
What actually changes:
- Deno runtime: 1 more year of monthly releases, stays open source
- Deno Deploy: 6 more months, migration help to Workers
- JSR: stays, moves to Cloudflare
- rusty_v8 is heading into workerd
We are excited to announce that Deno is joining Cloudflare!
deno.com/blog/cloudflare
VEXEEE retweeted
If this is not clickbait, I don't what this is. 🙄
Such a waste of energy for humanity, what is wrong with some people 😒
VEXEEE retweeted
very cool e2e use case:
you can set up a visual regression check with a fast model to catch inconsistencies/flaws on all your marketing pages (on all viewports).
It costs just $0.07 for all TesterArmy pages.
VEXEEE retweeted
Generates cinematic product videos using Claude Code and Remotion. video-shotcraft includes 157 shot recipe cards, motion previews, and automates storyboard, animation, and sound design workflows.
github.com/Vincentwei1021/vi…
VEXEEE retweeted
Cancel your Adobe subscription ASAP.
Everything Photoshop can do, running free inside a single browser tab.
ImageFlow will be launching in 24 hours.
VEXEEE retweeted
If you're a @herdrdev power user across multiple machines:
-Mobile: use Collie by @altanBLN
-Desktop: use herdr-gpui by @fabienpenso
Thank me later 🫡
VEXEEE retweeted
Opus/Sonnet 5.5 are surprisingly good at generating logo grids
Prompt I use in the comments if anyone wants to try it 👇
VEXEEE retweeted
Added two new skills to my skills repo recently.
/mobile-native - makes your web app feel native by doing stuff like disabling hover on touch devices.
/break-ui - tries to break your UI by adding long names, unusual emails, long lists, and fixes it.
github.com/emilkowalski/skil…
VEXEEE retweeted
speak once, hear yourself in another language 🗣️🌍
Index-Echo-S2ST is a new open model that dubs your speech and keeps your voice: talk in English or Chinese, get it back in English, Chinese, Japanese or Spanish
▶️ on Spaces hf.co/spaces/hugging-apps/in…
VEXEEE retweeted
The security-audit skill from @Cloudflare in action, pretty powerful
VEXEEE retweeted
VEXEEE retweeted
We release a technical blog on training PocketTTS, our 100M parameter on-device TTS, through drifting, a recent one-step generative objective from Deng et al. Less than 1% WER, high quality and voice cloning. To the best of our knowledge, it's the first speech model, and the first autoregressive model, trained this way.
































