- Backend
- Languages
- Deep dive
- Web
GitHub rewrote the Copilot runtime — more than 800,000 lines of production Node/TypeScript — into Rust, across 128 incremental PRs, and the dogfooding detail is the point: the migration was largely executed by Copilot itself. The resulting native runtime supports in-process FFI for six SDK languages, cutting the Node/V8 and process overhead that limited per-machine density. It is the clearest public evidence yet that agentic coding can carry a rewrite of this size in production, not just greenfield code.
Python Workers on Cloudflare are now generally available: Python runs in Pyodide/WebAssembly inside the same V8-based workerd runtime as JavaScript Workers, which means one serverless platform serves both languages with the same deployment model. The path is stable, and the boundaries are now well documented — multiprocessing and threading are unavailable inside the WebAssembly VM, so packages that assume forked processes or real threads will not port cleanly. The practical upshot is that Python's enormous data and AI library ecosystem has a credible edge runtime.
The most interesting deep-dive this week is a look at DeepSeek-V4.1-Flash, a 552B-parameter multimodal MoE built for input-heavy, long-horizon agent workloads with up to 1M-token context. Its asymmetric design — activating 8B parameters per token during prefill versus 16B during decode — plus its CED/KV-cache approach makes long-context serving materially cheaper, which is exactly the cost structure agentic workloads need when most of the work is reading, not writing. The framing matters: after years of models treating context as an undifferentiated blob, the industry is rediscovering the classic memory hierarchy — cheaper paths for the bulk of input processing, pricier ones where decisions get made.
DuckDB-WASM can now keep a database file persistent in the browser using the Origin Private File System, and OPFS is supported across all major browsers — meaning a real analytical database can survive reloads entirely client-side, with no server and no storage backend. For dashboards, local-first data apps, and offline analysis, that removes the last structural excuse for shoving everything through localStorage or IndexedDB adapters. The DuckDB write-up walks through the mechanics of a persisted WASM database over OPFS: what flushes when, what the durability guarantees are, and where the browser file-system API still bites.