Technical blog on backend, DevOps, AI, and software engineering.
OpenAI unveiled Jalapeño, its first custom inference chip with Broadcom — up to 1.9x more performance per watt. OpenAI is building its own hardware.
Shipyard, the team behind IPFS's core implementations and public gateways, winds down September 30. IPFS isn't dying — but who ships the fixes?
Ramp's July data shows Fable 5 taking just 8% of Anthropic model spend. Opus 5, half the price, overtook it in business spending within a month.
rust-analyzer can eat gigabytes of RAM. Rust Glancer flips the architecture — index once, store to disk, idle under 100MB.
Cerebras claims the CS-4 generates tokens 30x faster than GPUs by putting 44 GB of SRAM directly on the wafer — hitting 1,000 tokens/sec on 10-trillion-parameter models.
arrayref's maintainer account was hijacked to ship a proc-macro1 typosquat that runs malware at compile time — no need to even run the app.
Go 1.27 delivers generic methods — the feature developers have waited years for — plus a faster JSON stack, post-quantum signatures, and native UUIDs in the standard library.
Cursor ships Origin, its first in-house code host, pitched for the AI-agent era. The debate it sparked is about trust, not features.
DuckDB 2.0's preview turns the embedded analytics engine into a real server, adding the VARIANT type, triggers, and a 40x faster recursive query.
Stripe has agreed to buy OpenRouter for over $7B — the 'Stripe of AI' now belongs to the real Stripe. The AI routing layer just got revalued.
MCP has no fixed hostname — an AI agent's traffic looks like any other HTTPS API call. Cloudflare now detects and blocks it at the network layer.
Google's open-source HEIR compiler converts plaintext AI models into ones that run on encrypted data — bringing private inference to healthcare and finance.
DeepSeek open-sourced its Harness agent — a Claude Code rival — while raising API prices. Cache hits jump 6x starting August 16.
A rare data race in SQLite's checkpoint logic hid in plain sight for 16 years. It took Tailscale six months and 19 corruption incidents to track it down.
After 3 years of development, Mojo 1.0 is here — a language from the creator of LLVM & Swift that combines Python syntax with C/Rust-level performance.