IETF Standardizes HTTP QUERY Method; Colibri Engine Runs 744B Model on 25GB of RAM
The IETF finalized RFC 10008, standardizing the HTTP QUERY method—the first new safe, idempotent method since PATCH—to let developers send structured, cacheable read-only requests without abusing POST. Co-authored by Cloudflare and Akamai engineers, the spec promises edge/CDN adoption ahead of framework support, though CDN cache keys still derive from the URI, not the body. In systems-level observability, Go 1.25 shipped a flight recorder for runtime/trace that buffers the last seconds of execution data and dumps a snapshot on demand, solving the impracticality of continuous tracing in long-running services.
AI infrastructure saw breakthroughs at both ends of the scale. The single‑file, dependency‑free Colibri inference engine runs GLM‑5.2’s 744‑billion‑parameter Mixture‑of‑Experts model on just 25 GB of RAM and no GPU by keeping dense layers resident while streaming 21,504 experts from NVMe, leveraging MLA attention, DeepSeek Sparse Attention, and speculative decoding. On‑device inference also leaped forward: EdgeSync‑LLM demonstrated a 9.9× reduction in time‑to‑first‑token on an Android phone by reusing KV‑cache state for shared prompt prefixes via llama.cpp’s public APIs, slashing TTFT from 4.8 seconds to 486 ms for a 123‑token prefix.
Browser and UI frontiers expanded in two different directions. The 2.2‑million‑line FreeCAD parametric CAD platform was ported to WebAssembly in about four days—driven largely by an AI agent named Fable with all prompts and session transcripts public—bundling OpenCASCADE, Coin3D, CPython, PySide6, and VTK/SMESH into a single wasm module. In React 19, a deep dive into the Flight serializer revealed precisely which values cross the “use client” boundary: Dates, Maps, Sets, Promises, and FormData cross safely, while RegExp, URL, and class instances are rejected due to prototype‑based checks—critical knowledge for server‑component designers.
Assessing whether an AI truly understands a codebase gained a concrete method: a change‑impact teardown audit tests if the model can pinpoint silent breakages across dependencies, not just write correct code or pass tests. When AI agents need human input, a proposed UX pattern insists every interruption carry six fields—decision, reason, evidence, consequence, expiry, and recovery—to properly convey responsibility during a handoff. Ant Group’s Robbyant team released LingBot‑World 2.0, a 14B‑parameter open causal video world model trained with a MoBA attention mask and distilled for real‑time generation without post‑hoc drift correction.
Rounding out the week, a clarifying architectural comparison reminded engineers that the Kafka‑versus‑RabbitMQ debate isn’t about throughput but about fundamental design: RabbitMQ is a queue that delivers a message and forgets it, while Kafka is a log that stores messages for replay.
» Statistics
- Posts
- 867
- Reads
- 0
- Avg. score
- 7.7
» Top scored
- What Actually Crosses the React Server Component Boundary
- FreeCAD Ported to the Browser: 2.2M Lines of CAD via WASM
- HTTP QUERY Method Arrives: Is Your Infrastructure Ready?
- 9.9x Lower TTFT on Real Android Phone via llama.cpp KV Reuse
- Kafka vs RabbitMQ: You're Asking the Wrong Question
- Go 1.25 Ships Flight Recorder for Runtime Trace Diagnostics
- How do you actually test if an AI understands your codebase
- Colibri lets 744B-parameter GLM-5.2 run on just 25GB of RAM
- Designing the Moment an AI Agent Needs Human Input
- LingBot-World 2.0: Ant Group's Open Causal World Model