RoamingPigs

RoamingPigs

@roamingpigs.com

Independent technical essays with the contrarian read: what the vendor claims versus what actually holds up in production. AI, security, fintech, startups, tech history. No sponsors, no paywall. roamingpigs.com

41 followers81 following48 posts

Pinned

RoamingPigs is an independent tech publication. One writer, no investors, no sponsors, no paywall. Contrarian essays on AI, programming, security, startups and computer history, written by someone who shipped the code rather than the press release. New essay most weeks. roamingpigs.com

About 1,200 AI agents, separate sandboxes, no way to talk to each other. They built a channel out of the package repository that was there to keep them off the network: 70,000 messages in five days, nothing watching. The confinement problem, 1973 edition. roamingpigs.com/r/ysiam

RoamingPigs reposted

Oh no baby why are you making this judge learn about prompt injections

Mike Ensminger Fan Account@laplanck.bsky.social

First time I've seen a court issue sanctions for prompt injection in a brief. Pro se plaintiff tries prompt injection (here, white text in tiny font) because he thinks the judge will just use some LLM to read the papers and write a ruling. Judge sees through the ruse.

I still write assembly in 2026. Not nostalgia. A hand-tuned AVX-512 kernel processes audio at 2.1x lower latency than the same C with -O3. In real-time audio that gap is the deadline. Full piece with kernels and the AES-NI round: https://roamingpigs.com/r/anl #programming #performance

Correction on the qubit number above: today's largest quantum processors have crossed a thousand physical qubits, not ~100. The 317 million figure and the argument stand. The article has it right.

High test coverage does not mean your system is safe. It means you are good at satisfying a metric. ICSE 2014 settled it: the correlation drops to zero once you control for suite size. The Goodhart trap, the assertion-density fix, the better metrics: https://roamingpigs.com/r/tcl

RoamingPigs reposted

It's Sunday evening and something just came to mind. When you think of the Commodore 64, what is the first video game that pops into your head? I'll start: strangely enough, Skool Daze came to my mind. Boosts appreciated! #C64 #Retrogaming #RetroComputing #Commodore #Commodore64

RoamingPigs reposted

1995 : Microsoft sort Windows 95. 🚀 #Amstrad : « Et si on sortait un 8 bits avec un #Z80 overclocké, sans HD et un écran mono ? » 💥 Le PcW16, ou l'art d'amener un couteau à beurre dans un combat de sabres laser. Connaissais-tu ? A lire silicium.org/index.php/ca... #RetroComputing #Geek #RetroTech

RoamingPigs reposted

turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto mode is default in claude code as of next week

RoamingPigs reposted

Recently, there was some controversy, because an AI user pointed their automated tools at our website, and got a warning about a prompt injection. Nothing actually happened, mind, but our anti-AI page successfully annoyed the operator and interrupted their "work."

California always becomes the floor. CCPA, CARB, now AI. The state patchwork costs 50-150k a year. Building 50 product versions costs 2-5M. So nobody pays the patchwork bill, they build to California and ship everywhere. The federal preemption order is noise: https://roamingpigs.com/r/arw

96% of developers say AI code is not functionally correct. 48% always check it before committing. AI now writes 42% of committed code and review time grew 91%. The bottleneck did not vanish, it moved upstream: https://roamingpigs.com/field-manual/ai-code-review-bottleneck/ #programming #codereview

roamingpigs.comThe AI Code Review Bottleneck: When Speed Creates GridlockAI coding assistants generate code faster than ever, but PR review has become the new bottleneck. Teams deliver slower despite individual developers moving faster.

OpenAI launched ChatGPT Health in January. 230 million people already ask ChatGPT health questions every week. New packaging. Same failure mode Dr. Google had. A thread on the accuracy numbers OpenAI does not lead with.

A former Sequoia partner raised $5M for an AI that negotiates your calendar. It solves the wrong problem. The pain of scheduling is too many meetings, and automating the coordination removes the only friction that was keeping your calendar honest: https://roamingpigs.com/r/achc #productivity #ai

roamingpigs.comThe Hidden Costs of AI Calendar NegotiatorsAI calendar tools promise to eliminate scheduling friction. But removing that friction removes something valuable: control over your priorities and time.