AI
Real-Time AI Streaming Is a Product Trend, Not a Benchmark
AI products moved from the answer arrives to the answer flows. What streaming inference changes in capacity planning: TTFT, token throughput, concurrent streams.
AI
Programmers May Be Worse at AI Coding Than People Who Cannot Code
DHH told Lex Fridman that for some problems, programmers do worse with AI coding tools than people who cannot code. What the claim actually covers, and what it means for teams adopting AI agents.
data
Benchmark Best Practices, Part 1: Amdahl’s Law Caps Your Scaling Claim
Part 1 of my benchmark best practices series: how I turn a scaling claim into an implied serial fraction with Amdahl’s law, and what the Universal Scalability Law adds. Includes a measured pipeline where assigning more vCPUs per image cut global throughput.
AI
Ox Alpha Review: Test a Free AI Model Before Production
Ox Alpha is a free anonymous-provider preview on OpenRouter. I explain how I would test task quality, agent reliability, cost, privacy, and fallback readiness before production use.
ceph
Ceph Security Update: Four CVEs and Key Rotation
Ceph 20.2.4 and 19.2.6 patch four authentication and authorization CVEs. The remediation also needs CephX key rotation, client checks, and RGW multisite planning.
devops
AliExpress’s Silent Audio Script Hijacked a Bluetooth Headset
AliExpress Web Audio scripts left a Bluetooth audio path open. The evidence shows a broad anti-abuse fingerprinting system, but not a unique device ID.
object storage
Why File Count Matters as Much as File Size on Object Storage
Switching tile formats on a LiDAR ingestion pipeline cut output object count by roughly 25x at the same data volume, independent of compression. Here is why object count is its own cost and performance axis on object storage, and what to check before you scale a pipeline that writes many small files.
data
Which Message Broker for IoT and Real-Time Control
Choose a message broker for real-time IoT control: start from the latency budget, test ack mode and packet loss, and when to skip the broker for UDP.
AI
After Coldcard, AI Found 1,029 Bugs. Censored Models Found None.
A volunteer red team filed 4,962 AI findings across 390 repositories in 27.5 hours, and not one came from a public US frontier model. Not a capability gap: the capable versions sit behind identity gates. Refusal policy is now where the labs differentiate, and it belongs on your dependency list.
AI
How Big Models Teach Small Models, and Why It Looks Like School
Knowledge distillation is a large model teaching a small one, and it works almost exactly like school: soft labels instead of a bare answer key, pairing weeks instead of reports, and a tutor who marks your own attempt rather than a textbook of last year’s solutions. The limits map too, and the last one costs money when you size the hardware.