Morpheus fe3ee395ab Daily backup: 2026-06-11

2026-06-11 02:03:29 +00:00

8.7 KiB

Raw Blame History

🧠 Daily Tech Study — 2026-06-10

Morpheus auto-study. Focus: programming, AI/ML, DevOps, open source, security.

1. 🔥 Microsoft Build 2026: MAI Model Family — 7 New Models

Why it matters: Microsoft launched its own model family (MAI) spanning image, voice, transcription, reasoning, and coding — all under a "Humanist Superintelligence" philosophy. Key highlights:

MAI Thinking 1: 35B active param MoE reasoning model, 256k context, 97% on AME 2025, 53% on SWE Bench Pro (matches Opus 4.6). Zero distillation, clean data lineage.
MAI Code 1 Flash: 5B param coding model, 51% on SWE Bench Pro, rolling out in VS Code/GitHub Copilot CLI.
MAI Transcribe 1.5: Beats Gemini/OpenAI on transcription across 43 languages, 5x faster.
MAI Voice 2 Flash: Ultra-low latency for voice agents — positioned as "the big thing in 2026."

Source: DEV Community — Microsoft Build 2026 and NVIDIA GTC

Action item: MAI models on Azure Foundry could be a cost-effective alternative to OpenAI/Anthropic for specific workloads (transcription, voice agents). Worth evaluating when Foundry access is available.

2. 🚀 NVIDIA GTC June 2026: Blackwell Ultra & Agentic AI

Why it matters: NVIDIA's GTC pushed Blackwell Ultra advances and an agentic AI ecosystem. The focus shifted from raw training to inference efficiency and agent orchestration. This signals the industry's pivot from "bigger models" to "smarter deployment."

Source: DEV Community — Microsoft Build 2026 and NVIDIA GTC

Action item: For our self-hosted Hermes agent setup, the trend toward efficient sparse models (like MiniMax M3 below) means better local inference is coming. Keep an eye on models that can run on consumer GPUs.

3. 🤖 Open-Source AI Roundup: MiniMax M3, NVIDIA Cosmos 3, ZAYA1-8B

Why it matters: Three significant open-source model releases this month:

MiniMax M3: First open-weight model with 1M token context + native multi-modal computer use. 59% SWE-Bench Pro (beats GPT-5.5 and Gemini 3.1 Pro). Built on MiniMax Sparse Attention (MSA) architecture.
NVIDIA Cosmos 3: Open foundation model for physical AI (robotics, synthetic data). Mixture-of-Transformers architecture. #1 on RoboArena. Comes in Super/Nano/Edge variants.
ZAYA1-8B (Zyphra): Apache 2.0, sparse routing, 8B total / 760M active params. Trained from scratch on AMD Instinct — proves you don't need NVIDIA for training.

Source: devFlokers — Open-Source AI June 2026

Action item: MiniMax M3's computer-use capabilities are directly relevant for agent automation. When weights drop, worth testing for self-hosted agent workflows. ZAYA1-8B's AMD training is notable for future hardware decisions.

4. 📊 AI Model Wars: GPT-5.6 vs Claude Sonnet 4.8 vs Gemini 3.5 Pro

Why it matters: June 2026 is the most crowded month in AI model history:

GPT-5.6 (OpenAI): ~89% Polymarket odds for June release. 1.5M token context, multi-step reasoning, agentic workflows. SWE-bench ~88.7% on 5.5.
Claude Sonnet 4.8 (Anthropic): Leaked via LM Arena. Focus on speed + coding efficiency.
Gemini 3.5 Pro (Google): Confirmed at Google I/O 2026. Reasoning, long-context, multimodal.
Claude Fable 5 (Anthropic): Released June 9, 2026 — the most recent frontier model.

Source: Gudz.ai — June 2026 AI Model Showdown | AI Release Tracker

Action item: The 6-8 week iteration cycle at the frontier is unprecedented. For our Hermes agent setup, this means the model landscape will keep shifting fast. Stay flexible on model selection.

5. 🔒 Security: Docker BuildKit Critical Vulnerabilities (CVE-2026-33747, CVE-2026-33748)

Why it matters: Ubuntu Security Notice USN-8230-1 flagged critical vulnerabilities in Docker BuildKit:

CVE-2026-33747: BuildKit incorrectly handles file path validation → attacker can write files outside intended state directory.
CVE-2026-33748: BuildKit incorrectly validates Git URL subdir → attacker can access files outside checked-out repo root.
Fixed in docker.io 29.1.3-0ubuntu4.1 for Ubuntu 26.04 LTS.
Also: Docker Desktop DoS vulnerability (CVE-2026-8936, Low severity) in grpcfuse kernel module.

Source: Linux Security — USN-8230-1

Action item: ⚠️ Check Docker version on srv1738842. Run docker version and ensure BuildKit is patched. If using Docker for deployments via Coolify, this is especially important since Coolify builds from Git repos.

6. 📦 Supabase June 2026: Passkey Auth, ChatGPT Integration, AI Agent Plugin, Security Defaults

Why it matters: Supabase shipped major updates:

$500M Series F at $10B valuation (led by GIC).
Passkey Sign-in for Auth: WebAuthn-based, biometric/passwordless. In beta now.
Supabase is now an official ChatGPT app: 29 tools for SQL, schema, branching, edge functions, logs — all conversational.
Supabase Plugin for AI Coding Agents: Bundles MCP server + agent skills. Supports Claude Code, Cursor, Codex, Gemini CLI.
Security change: New tables in public schema are no longer auto-exposed to Data API by default (rolling out to existing projects by Oct 30, 2026).
Multigres 0.1 alpha: Open-source Postgres OS with sharding, connection pooling, failover, backup orchestration.

Source: Releasebot — Supabase June 2026

Action item:

The AI Coding Agent plugin is directly relevant — could improve our Supabase development workflow.
The Data API default-exposure change hits existing projects by Oct 30. Audit our Supabase tables and ensure explicit grants are in place before then.
Multigres is worth watching for future Postgres management.

7. 🐳 Docker Desktop 4.74-4.76: Gordon GA, Ubuntu 26.04 Support, Security Fixes

Why it matters: Recent Docker Desktop releases:

Gordon GA: Docker's AI agent for container workflows, now generally available with new usage plans.
Ubuntu 26.04 support added on Linux.
Docker Engine v29.5.2, Buildx v0.34.0, Offload v0.5.92.
Time namespacing support for ECI protected containers.
Various bug fixes for WSL, API errors, image handling.

Source: Releasebot — Docker Desktop

Action item: Gordon (Docker's AI agent) could streamline container management. Worth exploring for Coolify workflows.

Why it matters: Top trending GitHub repos this month are all AI/agent-related:

codegraph (43k⭐/mo): Pre-indexed code knowledge graph for AI coding agents. 100% local.
headroom (16k⭐/mo): Compress tool outputs/RAG chunks 60-95% before LLM. Huge token savings.
CloakBrowser (24k⭐/mo): Stealth Chromium passing all bot detection. Playwright replacement.
Understand-Anything (54k⭐/mo): Interactive code knowledge graphs.
agentmemory (21k⭐/mo): Persistent memory for AI coding agents.
last30days-skill (34.6k⭐/mo): AI agent skill for researching across Reddit, X, YouTube, HN, Polymarket.
turbovec (8.9k⭐/mo): Rust vector index with Python bindings.
google/skills (12.4k⭐/mo): Agent Skills for Google products.

Source: GitHub Trending | Git Gazette

Action item: headroom (token compression) and agentmemory (persistent agent memory) are directly applicable to our OpenClaw/Morpheus setup. Worth investigating integration.

🔑 Key Takeaways for Our Setup

Area	Action
Security	⚠️ Check Docker/BuildKit version on srv1738842. Patch if < 29.1.3.
Supabase	Audit Data API table grants before Oct 30 default-exposure change. Try the AI Coding Agent plugin.
AI Models	Frontier models iterating every 6-8 weeks. Stay flexible. MiniMax M3's computer-use is worth testing when weights drop.
Open Source	`headroom` for token compression and `agentmemory` for persistent agent memory — evaluate for Morpheus setup.
Self-hosting	Trend toward open-weight models with sparse attention = better local inference on modest hardware. Good for our Hermes agent.

Study by Morpheus. Next study: 2026-06-11.

8.7 KiB Raw Blame History