โ† All digests

Dev Digest โ€” March 4, 2026

๐Ÿ”ฅ HOT RELEASES

SkyDiscover โ€” Open-Source AlphaEvolve Modular framework for LLM-driven optimization that matches or exceeds Google's closed-source AlphaEvolve across 200+ tasks. Two new adaptive algorithms, +34% median score on Frontier-CS problems, discovers system optimizations beyond human-designed SOTA. The AI can even optimize its own optimization process on the fly. ๐Ÿ”— X: https://x.com/shulynnliu/status/2028892335875276919 ๐Ÿ’ป GitHub: https://github.com/skydiscover-ai/skydiscover

AReaL v1.0 โ€” One-Click Agentic RL From Huawei's InclusionAI team. Drop-in reinforcement learning for any existing AI agent โ€” adds a transparent proxy that captures all agent-LLM interactions for RL training with zero code changes. Open-source SOTA on tau2-bench, includes a new PyTorch-native 5D-Parallel Engine called Archon. ๐Ÿ”— X: https://x.com/jxwuyi/status/2029226785880301838 ๐Ÿ’ป GitHub: https://github.com/inclusionAI/AReaL

GPT-5.3 Released (and the drama) OpenAI shipped GPT-5.3 with 77 tokens/sec throughput at ~1B user scale. Meanwhile: massive community backlash โ€” users report it's worse than 5.2, the model identifies itself as GPT-4o internally, and the US State Department is running GPT-4.1 instead. GPT-5.4 teased for "soon." ๐Ÿ”— X: https://x.com/AILeaksAndNews/status/2028913080910909899 ๐Ÿ”— X (criticism): https://x.com/kexicheng/status/2029091002309591438

Qwen 3.5 Small โ€” Local AI Keeps Winning Alibaba dropped Qwen 3.5 Small, reportedly the most powerful free local model right now. Runs via Ollama, compatible with Claude Code. The 0.8B variant handles OCR surprisingly well โ€” one dev tested Kindle pages and handwriting recognition successfully. ๐Ÿ”— X: https://x.com/JulianGoldieSEO/status/2029164963944673758 ๐Ÿ”— X (OCR demo): https://x.com/stevibe/status/2029126123465130326

Agento v0.6.0 โ€” Open-Source AI Agent Platform Built on Claude, now with OpenTelemetry observability (full traces, metrics, structured logs), chat favorites, inline title editing, and working directory filters for Claude Code sessions. ๐Ÿ”— X: https://x.com/shaharia/status/2029259941790576940 ๐Ÿ’ป GitHub: https://github.com/shaharia-lab/agento


๐Ÿงช INTERESTING REPOS

SuperSplat โ€” 3D Gaussian Splat Editor Free, open-source browser-based tool for inspecting, editing, optimizing and publishing 3D Gaussian Splats. No install needed โ€” built on PlayCanvas web tech. ๐Ÿ”— X: https://x.com/rsasaki0109/status/2029029061637640211 ๐Ÿ’ป GitHub: https://github.com/playcanvas/supersplat

berezinopen/design-system โ€” AI Frontend Agent Config An .agent folder you drop into your codebase to make LLMs follow strict frontend rules โ€” semantic z-scales, fluid typography, Lenis scroll integration, 15 contextual UI patterns. Works with Cursor, Windsurf, OpenCode, Codebuff. ๐Ÿ”— X: https://x.com/Berezin12345/status/2029270538066055171 ๐Ÿ’ป GitHub: https://github.com/berezinopen/design-system

FerrisKey โ€” IAM Built in Rust Open-source Identity & Access Management platform for modern infrastructure. 500+ stars, 30+ contributors, growing in the cloud-native ecosystem. ๐Ÿ”— X: https://x.com/bonnal_nathael/status/2029195555172180024 ๐Ÿ’ป GitHub: https://github.com/ferriskey/ferriskey

Obsidian + Claude Code + MCP Template AI vault template that connects Obsidian to Claude Code via MCP โ€” 39 skills, searchable in 2 seconds. Open-source template for building your own AI second brain. ๐Ÿ”— X: https://x.com/Aiwithdhruv/status/2029168206259339740 ๐Ÿ’ป GitHub: https://github.com/aiagentwithdhruv/obsidian-ai-vault

GitHub Command Center Keyboard-first (Ctrl+K) GitHub interface built with Go 1.21 backend + React 18/TS frontend. Proxies GitHub API to beat CORS/rate limits. MIT licensed. ๐Ÿ”— X: https://x.com/PKMTSolar/status/2028911354921287978 ๐Ÿ’ป GitHub: https://github.com/paulmmoore3416/github-command-center

SO-ARM100 Parallel Gripper (3D Printed Robotics) Open-source parallel gripper for SO-ARM100/101 robot arms โ€” 1.5 kg payload, ~$76 in parts, 30 min assembly. Full STLs and source code. ๐Ÿ”— X: https://x.com/robo_nine/status/2029262721103827178 ๐Ÿ’ป GitHub: https://github.com/roboninecom/SO-ARM100-101-Parallel-Gripper


๐ŸŽฅ WORTH WATCHING

Claude Builds a Full VC Fund P&L Model One prompt, no follow-ups โ€” 7 linked Excel tabs, 25 portfolio companies, American waterfall carry, J-curve visualization. "McKinsey charges $6,625/day for this. This cost $0.20 in API tokens." ๐Ÿ”— X: https://x.com/chiefofautism/status/2029244084720738781

Generalized Non-Exponential Gaussian Splatting Research breakthrough: sub-linear/super-linear 3DGS variants achieve similar quality to original but with up to 4x speedup on complex real-world captures by reducing overdraws. ๐Ÿ”— X: https://x.com/janusch_patas/status/2029104953839476917

K-Search: Rethinking GPU Kernel Generation From Ion Stoica (Berkeley). Using LLMs as co-evolving Intrinsic World Models for strategic planning in multi-step code transformations โ€” much more sample-efficient than pure heuristic search. ๐Ÿ”— X: https://x.com/istoica05/status/2028978605255192603


๐Ÿ’ก TECHNIQUES & IDEAS

20 Security Prompts for Vibe Coders Viral thread (2.3K likes) โ€” you can't just tell AI to "make everything secure." Specific, actionable security prompts for AI-assisted development covering real attack vectors. ๐Ÿ”— X: https://x.com/DeFiMinty/status/2029049629287895528

"Code is a Disposable Artifact. The Prompt is the Asset." Hot take gaining traction: if it takes 6 passes to fix a bug with an agent, you're doing it wrong. Write a strict behavioral test, delete the file, let the model regenerate it. ๐Ÿ”— X: https://x.com/Prompt_Driven/status/2029263763250266501

Agent Architecture: Cyclical, Not Reactive CREAO's agent-to-agent interaction pattern: separate roles, constrained objectives, structured responses, memory across turns. Shift from Userโ†’Agentโ†’Output to Agent Aโ†’Protocolโ†’Agent Bโ†’Updated Stateโ†’Agent A. ๐Ÿ”— X: https://x.com/CreaoAI/status/2028965464618008628


๐Ÿ”ฎ EMERGING TRENDS

Claude Hits #1 on App Store โ€” People moving from ChatGPT to Claude. Key insight: the app and API are "completely different things" โ€” app had 2 outages this week, API stayed up. Enterprise users should use the API. ๐Ÿ”— X: https://x.com/sharbel/status/2029243288734048750

Anthropic's Free Claude Course Stack โ€” Multiple viral threads about 13 free official courses covering Claude Code, API, Agents, MCP, and AI Fluency. Full learning roadmap at zero cost. ๐Ÿ”— X: https://x.com/charliejhills/status/2029149899988816327

Vibe Coding Security Debate โ€” The 2026 vibe coding wave is spawning a security counter-movement. Developers pushing back on "just tell AI to make it secure" with structured security prompting frameworks and audit tools.

Local AI Models Going Mainstream โ€” Qwen 3.5 Small running locally via Ollama with Claude Code integration. The 0.8B variant doing credible OCR. Local inference quality keeps climbing while cost drops to zero.


Full digest: https://digest.99.cooking