Meta adopts QUIC for Instagram Live

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

📹 Meta adopts QUIC for Instagram Live

🤖 Muse Wonder debuts as a frontier model for AI agents

🔗 OpenCode 2.0 unifies terminal and apps

🪵 Cloudflare Logpush adds SQL log filtering

🐘 Postgres uuidv7 keys nearly match bigint

Plus: 🎁 8 more stories you might like, 🧰 6 tools, and 📚 5 papers.

📹 Meta adopts QUIC for Instagram Live LINK

• Meta has switched Instagram Live's video ingest on iOS and Android to Media over QUIC (MoQ), an emerging IETF standard it is building collaboratively with Google, Akamai, Cisco, and Cloudflare, replacing its decade-old in-house RUSH protocol.
• MoQ's core abstraction is a Track, an addressable media stream pushed through caching relays that forward objects without parsing payloads, letting two subscribers run at different latencies, roughly 100 ms for interactive use versus several seconds for buffered playback.
• Rather than rewriting the backend, Meta terminates MoQ at edge endpoints that translate frames to RUSH format, and the new path matched RUSH on quality and reliability, the parity target that now sets the baseline for extending MoQ to live playback.

🤖 Muse Wonder debuts as a frontier model for AI agents LINK

• Meta re-entered the frontier model race with Muse Wonder and Muse Wonder Full, announced yesterday, claiming on agentic software, legal and business-automation tasks they match or beat top models from OpenAI, Anthropic and Google.
• Independent benchmarker Artificial Analysis scored Wonder Full 56 on its Intelligence Index, tying Claude Sonnet 5.5 and topping GPT-6 Astra and Gemini 4 Argon at 53, though Claude Opus 5.5 still leads at 58; the standard Wonder scored 52.
• The standard model ships via API at $4/$20 per million input/output tokens with a ~1M-token context window, but Wonder Full is locked to paid Muse Code subscriptions with no production API, and neither ships open weights unlike Meta's Apache-licensed Muse Glimmer.

🔗 OpenCode 2.0 unifies terminal and apps LINK

• OpenCode 2.0 shipped today, formalizing the open-source coding agent's "major overhaul" that runs a single background service linking its terminal, web, and desktop interfaces alongside a rebuilt API and embeddable SDK.
• The server now starts automatically as one background process, so a session begun in the terminal can continue in the desktop or web app, while a new Code Mode lets agents write JavaScript calling multiple MCP tools in one execution to cut context bloat.
• Subagents and shell commands can now run in the background, but 1.0 plugins won't work with 2.0 and old server integrations need updating; existing users must upgrade manually via opencode upgrade --major or /update, as auto-updates were dropped.

🪵 Cloudflare Logpush adds SQL log filtering LINK

• Cloudflare Logpush now lets teams write SQL queries that filter, redact, and reshape log records before export, moving processing that previously required a separate service directly into the delivery pipeline.
• A Transformer is a SQL query tied to a specific dataset: records failing the WHERE clause are dropped, and the SELECT output becomes the exported NDJSON record, with execution powered by an embedded Apache DataFusion engine that processed a synthetic 5,000-row batch at roughly 1.6 million rows per second.
• Available on Free, Pro, Business, and Enterprise plans via Transformer Studio or the API, Transformers fail closed, retrying rather than shipping untransformed data, and cut downstream cost by discarding records like health checks before you pay to export and store them.

🐘 Postgres uuidv7 keys nearly match bigint LINK

• PostgreSQL 18's new time-ordered `uuidv7()` primary keys load 10 million rows in 32.3 seconds versus 29.7 for a bigint identity column, within 10%, while the old random `uuidv4()` key took 60.7 seconds, roughly double.
• Because PostgreSQL heaps append rows regardless of key, v4's damage lands on the B-tree index: random keys split pages to 70% fill versus 90% for ordered keys, and each post-checkpoint page touch logs a full 8 kB to WAL, costing three to four times the WAL per row.
• Fix an existing v4 table with a one-line `SET DEFAULT uuidv7()`, 5 ms, new rows only, which cut insert WAL to 250 bytes per row immediately; stick with v4 only where embedded creation timestamps shouldn't leak, since the ID reveals its millisecond.

Other news you might like

• Context Mode: keep bulky MCP tool output out of your agent's context window LINK

• AirLLM: run 70B LLMs on a single 4GB GPU LINK

• When you just need to build a tiny desktop app: TinyJS LINK

• A slightly faster slimmer curl LINK

• Parsing JSON at tens of gigabytes per second… in Python LINK

• flux9s GA: Flux cluster state, from the terminal LINK

• Chrome vs. Safari vs. Firefox: AI Debugging MCPs Tested LINK

• Big Arrow on the Screen: let your AI agents paint arrows, boxes and text on your Mac screen LINK

🧰 Trending tools

Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a chatroom when your allowance runs out. LINK

Kilo Code for JetBrains: open-source AI coding agent for JetBrains IDEs that lets you build and iterate with flexible model choices across projects. LINK

Mastra Factory: a TypeScript framework for building AI agents with staged workflows, memory, streaming, evals, tracing, and a Studio UI for testing. LINK

Kilo Code: open-source mobile AI coding agent that pushes notifications when sessions need input, so you can dictate prompts and attach files remotely. LINK

Monospace from Directus: wraps any SQL database with instant REST and GraphQL APIs plus a no-code app for collaborative data management. LINK

Meridian: open-source, locally-run activity tracker that logs work with context, revealing where your time goes and which blockers stall projects. LINK

📚 Trending papers & reports

Machine-learning quality checks get a plug-in framework that continuously watches live AI systems for data shifts and accuracy drops, catching real-world failures that standard software testing misses before they disrupt production. LINK

Optimization model repair automatically catches and fixes the errors AI makes when turning a business decision problem into a solvable model, correcting mistakes without breaking the parts that were already right. LINK

Verification flakiness gets measured across real code, showing that tiny harmless code tweaks can make automated proof checks stall or time out, and that optimizing how queries are structured would both speed them up and calm the unpredictability. LINK

AI code reviewers judging machine-written database specs barely beat a dumb baseline that just measures text length, yet they still correctly named the specific defect 84% to 92% of the time. LINK

Code cleanup tracking measures whether each round of tidying up messy software actually improves quality, checking before-and-after changes in size, complexity, code smells, bugs, and security flaws across open-source projects. LINK

See you tomorrow for a new dose of ☕️ Devshot!