Auset Brain

transcript-insights-march-28

3 min read

YouTube Transcript Insights — March 28, 2026

8 videos analyzed from .cursor/transcripts/3-28-2026/

MOST ACTIONABLE

1. Claude Certified Architect Exam (vizgFWixquE)

  • 5 domains: agent architecture (27%), tool/MCP integration (18%), Claude Code config (20%), prompt engineering, context management
  • 4-5 tools max per agent — more causes misrouting. Audit our agents.
  • Hooks for compliance (100% deterministic), prompts for style (90%)
  • Sub-agents work best with broad goals, not micromanagement
  • First 40% and last portion of context get most attention — middle gets fuzzy (“lost in the middle”)
  • Mo should take this certification

2. Google Gemini 3.1 Flash Live — Voice (Qt3zMBH-FNg)

  • Speech-to-speech — no text intermediary, lower latency
  • 70+ languages with real-time translation
  • Works in noisy environments, understands alphanumeric strings
  • Function calling support (calendar, email, tickets)
  • Free tier in Google AI Studio
  • Could replace Groq + MiniMax chain for Clara Voice

3. AutoDream — Memory Consolidation (6pjETAf2XhU)

  • 4-phase cleanup: Orientation → Gather Signal → Consolidation → Prune & Index
  • Fixes memory decay after 20+ sessions
  • Already enabled in settings (autoDreamEnabled: true)

4. Hidden Claude Code Settings (-73ERRlbDn0)

  • cleanupPeriodDays: 365 (applied), agent teams experimental (applied)
  • feedbackSurveyRate: 0 (applied), autoDream (applied)
  • bash_max_output_length and file_read_max_output_tokens NOT in schema yet

MARKET VALIDATION (We’re on the right track)

5. Xiaomi Mimo V2 (yJlP9iqStSE)

  • 1 trillion parameters, 1M token context, built for agent task execution
  • Shift from conversational AI to task execution — validates our agent-first approach
  • Parallel sub-agents, broad goals over step-by-step

6. Minimax M2.7 — Self-Evolving Agents (eRS7w-L-ypQ)

  • Agent harness: skills, memory, tools — AI improves through analyze-plan-fix-test-repeat loops
  • Handles full software project lifecycle (write → test → fix → ship)
  • Multi-agent coordination — different agents handle code, research, writing
  • This IS our pipeline — Opus → Sonnet → Haiku → agents → review → merge

7. Moonshot Kimmy K25 — Attention Residuals (bjWPni3LbEk)

  • Fixes “lost in the middle” — layers selectively surface relevant earlier context
  • 256K token context, runs 100 parallel agents in swarm mode
  • 1T params with only 32B active = efficient routing
  • Open source on HuggingFace

8. Baidu Chenfan OCR (Oz895snZVrg)

  • Single model replaces entire document processing pipeline
  • 192 languages, layout understanding, structured extraction
  • Low priority for us but validates trend: unified models > chained tools

KEY THEMES ACROSS ALL TRANSCRIPTS

  1. Agent-first > chatbot — every model vendor pivoting to task execution
  2. Parallel multi-agent swarms — 100+ agents coordinating (we do 8-10)
  3. CLI/MCP/API-first — tools ranked by agent usability, not human UX
  4. Speech-to-speech — text intermediary is a bottleneck being eliminated
  5. Self-evolving agents — the agent improves itself through loops
  6. Broad goals > micromanagement — give agents outcomes, not steps

ACTION ITEMS

  1. Take Claude Architect exam — Mo is already doing this work, get certified
  2. Audit agent tool counts — 4-5 tools per agent max, reduce where needed
  3. Evaluate Gemini 3.1 Flash Live for Clara Voice (speech-to-speech, free tier)
  4. Agent teams enabled — test in next session with coordinated swarm
  5. AutoDream enabled — will auto-consolidate memory going forward

Graph View