RC RANDOM CHAOS

Articles

Long-form writing on tech, culture, and the edges of the internet.

Zhipu built its own inference stack for GLM
LLM inferenceAI infrastructure

Zhipu built its own inference stack for GLM

GLM built its own inference infrastructure to serve LLMs cheaply on constrained, mixed hardware. Here's what breaks in generic stacks and what to copy.

10 min read
A 4B model outplans Postgres
postgres optimizationllm engineering

A 4B model outplans Postgres

A 4B model proposes faster Postgres query plans, but the validation gate and fallback - not the model - are what make the 81% speedup safe to run.

10 min read
A patch waits eleven days at the gate
Google PlayAndroid security

A patch waits eleven days at the gate

Google Play reviews now take a week or more. The real risk isn't malware slipping the gate - it's droppers that mutate after approval and slowed security patches.

6 min read
Gemini 3.8 Live broke two security assumptions
AI safetyprompt injection

Gemini 3.8 Live broke two security assumptions

Gemini 3.8 Live and Extended Thinking make ambient audio and video an untrusted AI input, reshaping prompt injection, logging, and privacy risk.

7 min read
Rubble where the records used to be
cloud resilienceboard risk

Rubble where the records used to be

A cloud provider states some Middle East data cannot be restored after a strike - what permanent loss means for board-level accountability and recoverability.

8 min read
The merged Claude isn't about a smarter model
AI workflowsLLM orchestration

The merged Claude isn't about a smarter model

Merging Claude Cowork and chat removed the seam that forced discipline. Now structured workflows, not prompts, decide whether your AI work holds up.

8 min read
The page never leaves your laptop
local LLM inferenceAI automation pipelines

The page never leaves your laptop

Run Mistral's multilingual models locally via Mozilla llamafile to tier web automation: cheap private steps stay local, hard reasoning escalates to the cloud.

9 min read
Baseten lost production GitHub admin in 25 minutes
github securityidentity boundary

Baseten lost production GitHub admin in 25 minutes

Admin access to Baseten's production GitHub was reached in 25 minutes. A briefing on the identity boundary that failed and what the number means.

8 min read
Every new feature makes iOS 27 less safe
iOS 27 securityprivacy

Every new feature makes iOS 27 less safe

iOS 27, iPadOS 27, and macOS 27 concentrate risk at one point: whether a capability re-checks identity at execution, not at the moment consent was captured.

9 min read
One attacker, three rivals, zero coincidence.
threat attributionAI company security

One attacker, three rivals, zero coincidence.

One actor is tied to OpenAI, Anthropic, and Meta incidents. The confirmed failure is collapsed separation, not any single company's controls.

7 min read
The bill in your admin channel isn't law
MastodonFediverse

The bill in your admin channel isn't law

The UK Ofcom licensing bill won't make Mastodon servers illegal - it's an unpassed Lords bill. Here's the law that actually applies to admins today.

7 min read
The model already broke your embargo
responsible disclosurevulnerability embargo

The model already broke your embargo

An OpenAI model reproduced the RubyGems caching vulnerability before public disclosure, exposing why embargoes fail against automated ingestion.

8 min read