AI development
-
Crush yolo mode now blocks curl, but rm -rf still runs
Crush yolo mode at v0.101.0 blocks curl, sudo and a configured kubectl, while rm -rf and bash -c 'curl' still run. Run at v0.100.0 and v0.101.0.
-
Is an MCP phone gateway safe? Where Pinky draws the line
Is an MCP phone gateway safe? Pinky lets Claude Code drive a real phone. I ran its approval tests and 40 button labels to see where it stops.
-
LongCat-2.5 weights add up to more than its model card says
The LongCat-2.5 weights sum to 1.78 trillion parameters, not 1.6, and the 4-bit repo is 1,336 GB. I added up the files and counted the asterisks.
-
Edge0 145B fits in 24GB VRAM if you add the RAM
Edge0 145B fits 24GB VRAM, but its own page puts 31 to 38 GiB on host RAM. A local LLM for coding on 24GB VRAM, with the sums checked.
-
vLLM vs SGLang when the MoE model doesn’t fit in VRAM
vLLM vs SGLang for an MoE model bigger than your GPU: paged experts run 1.94 times faster than llama.cpp at FP8, 87.7 vs 45.2 tok/s, per the PR's tables.
-
Who maintains the official MCP servers? One person, lately
Who maintains the official MCP servers? I counted five releases: one account wrote 575 of 591 commits. Every number is a curl away.
-
OpenSSH releases are coming faster because AI finds the bugs
OpenSSH releases now come faster because AI finds the bugs. I counted the gaps, diffed the 10.6p2 fix and wrote the checks a small team can run today.
-
Claude Code privacy is a plugin, so test your own agent
Claude Code privacy in practice: the account email sat in the model's context. I pointed opencode and Codex at a request logger and read what they send.
-
DeepSeek coding agent has one prompt rule worth stealing
The DeepSeek coding agent is MIT open source. I installed it in 8 seconds, ran 1,197 tests, and found one prompt rule worth stealing.
-
MCP Skills: I Tried Them, and Verified Is Not Trusted
MCP skills ship in the official TypeScript SDK 2.4. I ran them, and a hostile server's prompt injection still came back verified: true. Here is why.