Palash Bansal

809 posts

Palash Bansal banner
Palash Bansal

Palash Bansal

@repalash

building stuff https://t.co/jYKSNAYwG7 https://t.co/uFsnfqIINx https://t.co/CAtPuwbXAH

Katılım Eylül 2012
533 Takip Edilen3.3K Takipçiler
Sabitlenmiş Tweet
Palash Bansal
Palash Bansal@repalash·
More Math to Shaders!! Visualize(and edit) both the code and the visuals instantly. Now with data types and control flow, we can create function nodes with multi-line math to code. Here's SDFs for Box, Sphere, Torus and Plane in maths straight from iq's blog All in a browser!
English
11
36
402
27.8K
Palash Bansal retweetledi
Minjune Song
Minjune Song@minjunesh·
Your agent can create a backend by POSTing to: blitz.dev/api/v1/new-pro…<slug> This creates a backend at https://<slug>.app.blitz.dev with a D1 database, R2 storage, and a project token. No account needed! Point your agent to blitz.dev to learn more
Minjune Song tweet media
English
1
2
5
613
Minjune Song
Minjune Song@minjunesh·
We built Blitz, the cloud that agents use without login. If your agent can make network requests, it can use Blitz to deploy full-stack apps. Just say "deploy to blitz.dev" to your agent.
English
1
1
8
300
Palash Bansal
Palash Bansal@repalash·
@yusukebe How about trying AI for triage. It would be pretty good at it
English
0
0
0
285
Yusuke Wada
Yusuke Wada@yusukebe·
これ!Honoでも同じことが起こっている!非公開のセキュリティレポートが溜まっている!
yousukezan@yousukezan

Linuxカーネル開発を率いるリーナス・トーバルズ氏が、AIで発見された脆弱性報告の急増に苦言を呈した。同じ問題を別々の研究者が大量投稿し、セキュリティメーリングリストが「ほぼ管理不能」になっているという。 トーバルズ氏はLinux 7.1 RC4公開に合わせた週間報告で、AI利用者による重複報告がメンテナーの負担を急増させていると説明した。「既に修正済み」「数週間前に議論済み」と案内する作業ばかりになり、本来の開発や検証時間が奪われていると批判している。 特に問題視したのは、AIツールが同じコードパターンから同一バグを大量検出する点だ。非公開のセキュリティリストで処理すると、研究者同士が互いの報告を確認できず、重複がさらに悪化すると指摘した。AIで見つかるバグは「秘密情報ではない場合が多い」とし、単なる通報だけでは価値が低いとの考えを示している。 一方でAI活用自体を否定したわけではなく、「実際に役立つ形で使うべきだ」と強調した。AIが見つけた問題に対し、コードを理解し、修正パッチまで用意して初めて貢献になると述べ、「理解のないまま報告だけ送るな」と研究者らへ呼びかけている。 theregister.com/security/2026/…

日本語
2
27
220
48.9K
Matt Auerbach
Matt Auerbach@mauerbac·
I'm officially becoming a @Cloudflare customer and trying it out! Starting with R2
Matt Auerbach tweet media
English
20
0
113
12.2K
Palash Bansal
Palash Bansal@repalash·
@BenjDicken Normalise creativity rather than copy pasting the same Dino game everywhere for the sake of putting a mini game
English
0
0
1
46
Palash Bansal
Palash Bansal@repalash·
@yusukebe Yes, it can be in the form of checklists that they have to follow thoroughly and FAQs (frequent agent questions)
English
0
0
1
32
Yusuke Wada
Yusuke Wada@yusukebe·
ちょっとできてないけど、コントリビュータ側のAIエージェントが見るためのドキュメントをOSSライブラリサイドは用意したほうがいいな 直近では、セキュリティレポートが大量にきてるんだけど、脆弱性と見えても方針でよしとしてるものもあるので、そういうのを書いておく
日本語
2
1
18
4K
Palash Bansal
Palash Bansal@repalash·
@willeastcott It still can't see, there could be like a feedback loop to check the text wrt the current frame
Palash Bansal tweet media
English
1
0
0
93
Palash Bansal
Palash Bansal@repalash·
@thekitze Thats still a scam bro, even if you try to be a robin hood or whatever
English
0
0
0
92
kitze · supermac.io 🐦‍🔥
the amount of times i see a $20/mo charge notification and i can't be bothered to do something about it is insane. i might let it run for a few months before i get annoyed or it catches me at the right time. i'm just a lazy idiot when it comes to money. now imagine ACTUAL rich ppl's mindset when it comes to paying for saas etc lol there's AN ENTIRE WORLD OUT THERE> CHARGE THEM
English
14
0
49
7.3K
Palash Bansal retweetledi
Yusuke Wada
Yusuke Wada@yusukebe·
!!!!!!!
Yusuke Wada tweet media
ZXX
4
15
235
22.8K
Madhuri ✳︎
Madhuri ✳︎@Iruhdam24·
Can I be a forward deployed product designer?
English
3
0
26
2.8K
John Lindquist
John Lindquist@johnlindquist·
Honestly, Claude Code's /goal shipped too soon
John Lindquist tweet media
English
20
4
108
14.2K
Palash Bansal
Palash Bansal@repalash·
@PThorpe92 so if i make an unsafe.rs file with just 2 unsafe calls and use that module 100k times, is my code safer?
English
0
0
0
279
Dan Greenheck
Dan Greenheck@dangreenheck·
"AAA graphics aren't possible in the browser" Hold my beer 🍺
English
108
123
2K
116.1K
Ashley Peacock
Ashley Peacock@_ashleypeacock·
My first Cloudflare ship is in wrangler@4.90.1 🎉 It fixes remote bindings hanging indefinitely when closing a wrangler dev session A little side ship while I was working my way through onboarding, simple fix but tricky to nail down - fun way to learn Wrangler though
English
10
2
150
5.7K
Palash Bansal
Palash Bansal@repalash·
That also gives you a false sense of security. We tried a few months back with Opus 4.5. Simple test with a normalise file name method using snakecase with some constraints. It wrote nice looking code, and so many tests. After manual review of the code before deployment, it was using a regex. It was okay when it first wrote and I checked, but during some /simplify it moved the regex string out in a variable because someone might want to use the 6 char regex again in the snake case file. So new regex(/\W... Became r = "\W... It worked fine. All its written tests kept working. But the side effect of this. Any string that starts with capital W is stripped of W. And White becomes hite. Asking the developer they were like it's a basic snake case function, didn't expect opus to mess that up even, and we asked it to write a bunch of text which looked pretty good.
English
0
0
0
137
Glauber Costa
Glauber Costa@glcst·
There is a class of software that can operate particularly well under a logic of throw-it-see-if-it-sticks. For example tests: you can automatically generate tests as much as you want. If you generate bad tests, nothing bad really happens. But if you generate good tests, then you found a bug and the reward is there. This kind of software is particularly well suited for AI: brute force a lot of space, and if a small percentage of the generated code finds issues, you're good. Unfortunately for us, the other class of software that works well under this logic is exploits.
English
8
3
44
3K
Palash Bansal
Palash Bansal@repalash·
@CherryJimbo Doesn't work on mobile also. Cloudflare is embracing AI so much that it's going in slop territory. For those saying humans also make mistakes - yes, but most don't approve broken things for prod by lying about it and calling it done and stable.
English
0
2
16
1.8K
James Ross
James Ross@CherryJimbo·
Cloudflare launched their new marketing site today, and while it's pretty... - it's inaccessible (contrast WCAG issues, tab-indexes wonky) - inaccurate pricing info - spikes CPU usage - clearly rushed/vibed - so many ux issues Humans have clearly been removed from the process.
James Ross tweet media
English
43
9
474
77.7K
Palash Bansal
Palash Bansal@repalash·
@alightinastorm Cosplaying a retard coming back to bite? 😂 Btw that person made a pretty strong prediction, it's easy to prove them wrong by next month. Just don't burnout or they'll win
English
1
0
0
242
robot
robot@alightinastorm·
damn i just found out i was mentioned on reddit i don't get it, what do some people expect? i am building in hyper public, almost on every commit i'm recording a progress video and a devlog i set an insane goal of achieving AAA quality, which usually takes millions of $$ and huge teams 3 months in, i've achieved quite a lot and whenever i encounter a new problem space, i take a break from the game's development and develop a tool oss to make it happen, for everyone! Recap: First Demo: Figuring out as many problems as possible during "raw vibe coding" (cli/chat only) -> World modeling: Tested different AI harnesses for constrained modeling, from generative floorplans + extrusion to more sophisticated level editors with AI integrations -> Building character state machines is not a very good fit for code-only, as small changes require a lot of context for AIs, but may be solved with rewiring a node visually as a human. So i built the character + animation studio apps -> Iterating quickly across scenes and reusing assets: Built a vite-framework which connects the level editor and character studio to iterate faster -> Cheap but high quality vfx: Nothing exists, everywhere i searched (google, yt, forums) everyone seems to just happily pay embergen or houdini money to do it -> Higher fidelity of 3D human models: Tripo and other ai model generators are really great for props, but they won't get you to really high quality (and decent topology) and the biggest bottleneck currently for higher quality human characters is, YOU HAVE NO FREE HIGH QUALITY OPTIONS. Metahuman is pseudo-free. In between these breaks i always come back to the game and use these tools There's still SO much missing, from clothing for the models, to having a good default animation library (also here, mixamo is free but it is missing SO many animations!!) And more, so so much more. It's funny how i am actually very transparent and share the process daily and have to speedrun horizontally through all fields necessary, yet being compared to the influencers that show oneshot slop! My progress hasn't really slowed down, i'm hitting bigger problems which require more time and energy, even if vibecoded anyways new updates on the human creator soon, god forbid a robot takes a day off 😂
robot tweet media
English
37
3
113
8.9K