Dev Digest — April 2, 2026
🔥 HOT RELEASES
Google Gemma 4 Google's most capable open models yet. 4 sizes (31B Dense, 26B MoE, E4B, E2B), Apache 2.0 license, 256K context, native function calling for agents, multimodal (vision + audio), runs on everything from phones to workstations. The 31B model hits Arena Elo 1464 — competing with models 20x its size. 🔗 X: https://x.com/GoogleAI/status/2039735543068504476 📖 Blog: https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/ 💻 HuggingFace: https://huggingface.co/collections/google/gemma-4 💻 Ollama: https://ollama.com/library/gemma4
Qwen 3.6-Plus Alibaba drops a massive agentic coding upgrade. Claims to beat Claude Opus 4.5 on select benchmarks, SWE-bench Verified at 78.8. 1M context window, 65K output tokens, enhanced multimodal vision, 3x faster than Claude Opus 4.6 in early tests. Already #8 on Code Arena. Free on OpenRouter. Open-source smaller variants coming soon. 🔗 X: https://x.com/Alibaba_Qwen/status/2039705104723611829 📖 Blog: https://qwen.ai/blog?id=qwen3.6
Anthropic's April Fools Reveal The "Claude Code leak" from March 31 was a deliberate April Fools prank. The source code, Mythos benchmarks, 3,000 internal docs — all fabricated. They seeded fake assets in a staging environment and watched the internet go wild. Internally codenamed "Capybara" because capybaras sit calmly while everything escalates. 🔗 X: https://x.com/om_patel5/status/2039518876636168278
🧪 INTERESTING REPOS
eye-declare — Declarative Rust TUI Framework React-like components for terminal UIs, built on top of ratatui. Viewport + scrollback support, inline rendering. By the Atuin team. 🔗 X: https://x.com/orhundev/status/2039682606162157948 💻 GitHub: https://github.com/atuinsh/eye-declare
similar — Rust Diffing Library (Major Update) @mitsuhiko's diffing library gets histogram/hunt diffs, semantic fixes for inline diffs, no_std support, and perf improvements. 🔗 X: https://x.com/mitsuhiko/status/2039444228338520434 💻 GitHub: https://github.com/mitsuhiko/similar
QuantAgent — Multi-Agent LLM Trading Analysis Open-source system from Stony Brook, CMU, Yale, UBC, and Fudan. Four specialized AI agents (Indicator, Pattern, Trend, Decision) analyze markets simultaneously and synthesize into actionable trade decisions. 🔗 X: https://x.com/heyrimsha/status/2039613450834620725 💻 GitHub: https://github.com/Y-Research-SBU/QuantAgent
Feynman — Open Source AI Research Agent 🔗 X: https://x.com/tom_doerr/status/2039115357839929610 💻 GitHub: https://github.com/getcompanion-ai/feynman
Liquid AI Mobile Cookbook Collection of on-device AI apps running 100% locally with LFM models and Liquid's LEAP SDK. No cloud, no API keys. 🔗 X: https://x.com/paulabartabajo_/status/2039689619357261900 💻 GitHub: https://github.com/Liquid4All/cookbook
ML System Design Case Studies (300+) Curated list of real ML system case studies from top companies. 🔗 X: https://x.com/mdancho84/status/2039666749147140385 💻 GitHub: https://github.com/Engineer1999/A-Curated-List-of-ML-System-Design-Case-Studies
PointFree DebugSnapshots (Beta) New Swift library bringing exhaustive testing tools to reference types — making difficult-to-test types testable. 🔗 X: https://x.com/pointfreeco/status/2039707607955341401 📖 Blog: https://www.pointfree.co/blog/posts/205-beta-preview-debugsnapshots
Claude Code Elixir/Phoenix Plugin v2.6.1 56 files changed across 40 skills, 20 agents, and the hook system. Self-improving skills, auto-analysis of 513K TypeScript lines. 🔗 X: https://x.com/quatermain32/status/2039613764257874130 💻 GitHub: https://github.com/oliver-kriska/claude-elixir-phoenix
🎥 WORTH WATCHING
Slim pickings today for dev video content. The Gemma 4 launch has various demo videos embedded in the announcement tweets — worth checking the Google AI blog directly for walkthroughs of running it locally.
@victormustar — How to Run Gemma 4 Locally Step-by-step guide to running Gemma 4 on your own hardware with llama.cpp, including Chat UI + OpenAI-compatible endpoint. 🔗 X: https://x.com/victormustar/status/2039739591276581118
💡 TECHNIQUES & IDEAS
Self-Validating AI Workflows (Builder-Validator Pattern) Embed Puppeteer in Claude Code skills so the AI renders HTML, screenshots its own output, measures if text overflows, and iterates 3x before you see it. Running the same prompt twice improves quality (Google paper). Build the check into the system — stop being the QA layer. 🔗 X: https://x.com/aakashgupta/status/2039442140405895546
Claude Code for Video Editing One user had Claude Code scan 122 videos (100GB), extract thumbnails, identify scenes with specific people via image recognition, read slide text, and cut 4K clips — all from a single prompt. ~10 minutes vs half a day for a human editor. 🔗 X: https://x.com/masahirochaen/status/2039590592179789939
LACP v0.7.0 — Heterogeneous Agent Teams Mix Claude + Codex + Hermes agents in one session with difficulty-aware routing. Simple tasks → cheap models, hard tasks → expensive ones. RRF retrieval, active recall probes, dreaming consolidation for memory. 🔗 X: https://x.com/nyk_builderz/status/2039663977308029076
🔮 EMERGING TRENDS
Open models are closing the gap fast. Gemma 4 at 31B params competing with 600B+ models on Arena. Qwen 3.6-Plus matching Opus on coding benchmarks. The "you need a huge closed model" argument is weakening weekly.
Agentic coding is the new battleground. Every major release (Gemma 4, Qwen 3.6) now leads with "agentic capabilities" — function calling, tool use, multi-step reasoning. SWE-bench and Terminal-Bench are becoming the benchmarks that matter.
Self-improving agent systems. From LACP's difficulty-aware routing to the builder-validator pattern to Claude Code plugins that improve their own skills — the meta-pattern is agents that get better without human intervention.
On-device AI going mainstream. Gemma 4 E2B/E4B for phones, Liquid AI's mobile cookbook — the push to run everything locally is accelerating. Privacy + latency benefits are compelling.
Full digest: https://digest.99.cooking