Just amazed how sophisticated attackers can be...
- inject a malicious prompt into Claude via a GitHub issue title
- evect the GitHub Actions cache by overloading it with garbage
- inject code into Cline to install OpenClaw on victims machines
adnanthekhan.com/posts/clinej...
The Claude C Compiler was fascinating but reading this does show that it's fairly elementry in its approach.
Not sure if it's a prompting issue or that there's not enough weight towards compiler optimizations in the LLMs?
harshanu.space/en/tech/ccc-...
That said, it still requires a lot of solid, good knowledge in software engineering to create this harness.
Far from vibe coding.
Not only are the results fascinating, but the steps the engineer had to take plays as a good reference.
It's also encouraging that the steps an Anthropic researcher is taking are nothing special and it's something humans would have had to do anyways.
www.anthropic.com/engineering/...
Interesting result on how having an index in AGENTS.md performed better than having Skills.
AGENTS.md feels more natural as it's about adding context whereas Skills is more an action.
vercel.com/blog/agents-...
> But invest primarily in your people’s ability to think clearly about complexity.
We've gone through many iterations of the "developers are unnecessary" yet never happened.
We're problem solvers tackling complexity and that never goes away.
www.caimito.net/en/blog/2025...
Anthropic has their original CLAUDE.md so seems like it's OpenAI's move to make AGENTS.md the defacto-standard.
Not sure if that means much though given it's really just natural language that gets thrown into the prompt🤔
Been reloading OpenAI's news all day to see GPT 5.2 release and found this news:
OpenAI co-founds the Agentic AI Foundation under the Linux Foundation
openai.com/index/agenti...
They're also going to contribute AGENTS.md to the foundation.
This "veto test" is so simple yet so spot on.
"the “veto test” used by Silicon Valley Product Group: If a person has veto power over your project or can stop it from launching, they’re a stakeholder you need to engage."
(taken from "Frictionless")
This is really unfortunate that the owner lacked respect to the contributer.
Even if they did use AI, they at least should have been open about it and explain the reason behind doing so.
joshua.hu/ai-slop-okta...
This a good lesson on
- how previous conceptions may not be adequate today
- it's not always about using the absolute best model (ROI matters)
www.levs.fyi/blog/2-years...
This bottoms up way is typically how technology spreads.
www.bloomberg.com/opinion/arti...
Ordered Codex/Claude Code/Amp for code.
Ordered Notion AI to write my docs.
Asked ChatGPT/Claude AI for research.
Hmm...Am I allowed to say this is MY work? 😂
High performers tend to be ones who constantly re-evaulate their process (i.e. "how" they work) and keep on improving it.
Everyone should listen to what Addy has to say about AI coding.
Very thoughtful and level-headed but yet very optimistic about AI.
newsletter.pragmaticengineer.com/p/beyond-vib...
Just donated to PSF.
Sad to see they had to withdraw their proposal to the US government but I think it's the right decision.
My donations small but just wanted show some people are behind them.
pyfound.blogspot.com/2025/10/NSF-...
Zed v0.210.4 release note screenshot
The latest Zed release note is huge!
I really think it's gaining momentum!
The small things make a big difference.
Been posting the same message to X, Bluesky, Threads, Mastodon. Some apps retry properly even in the background.
You can always resend it yourself but it's so much more pleasant to use.
- LLMs can be helpful for software development.
- LLMs for software development are being oversold by some people who stand to gain a lot of money from them.
- LLMs for software development are being undersold by some people who have decided for various reasons that they do not like the technology.
These 3 are exactly how I think about AI/LLMs in software development
steveklabnik.com/writing/i-am...
Really nice to see great support given to the Zig community.
I've recently been re-writing some of my scripts in Zig for faster execution too.
www.synadia.com/blog/synadia...
Been using Zig for some personal projects lately. Really impressive stuff. These improvements make the future look bright for Zig.
mitchellh.com/writing/zig-...
Fascinating deep-dive into Claude Code's architecture.
Simple grep commands prefered over RAG for code search and uses smaller models for 50% of operations.
Sometimes the elegant solution beats the sophisticated one.
minusx.ai/blog/decodin...
CloudKitchen's evaluation summary of tools.
Real talk on AI in engineering: CloudKitchens' practical eval of GenAI for dev—results are refreshingly grounded, no hype.
Really like their evaluation summary too.
techblog.cloudkitchens.com/p/study-and-...
Feels like it's becoming increasingly important to have "good taste" of what's better for your codebase.
It used to be that you had to spend time on code edits before you could tell but with LLMs like Claude Code, it's way easier to try it out.
Apple's new on-device LLM is post-trained on a special data structure that is and to be generated by the developer based on a Swift Generable annotation.
This is interesting.
First time I've seen an LLM integrated with a programming language in this way.
machinelearning.apple.com/research/app...
Google's email about their new coding agent, Jules, just landed in my GMail spam folder😂
Gmail's AI is desperate for survival by blocking people's access to other AIs😆
I'm starting to think static typing (or rather stricter rules enforced by the compiler) is the way to go when it comes to AI tools (Cursor, Cline, Claude Code, etc).
The compiler can pretty much direct the AI agent to apply the changes needed.
Wow...Claude Code has just upgraded my GitHub helper command written in Haskell from 9.10.1 to 9.12.1 and actually used it to create the pull request merely by reading the code (and maybe the help message it executed during the session).
Been bumping into scaling issues recently.
Might sound a bit inappropriate, but makes smile every time.
Makes everyone busy fighting but proves we are growing!
