Iggy's Mission Control

🧠 How It Works

• Each session starts fresh — files are continuity
• Daily notes capture raw events, decisions, findings
• MEMORY.md = distilled long-term knowledge
• Periodic heartbeat reviews consolidate daily → long-term
• All 4 Telegram bots share the same workspace + memory files

🌙 Planned: Overnight Memory Agent

Run a local model (Llama 3.2 / Qwen 7B) on the Mac mini overnight to review daily notes, consolidate MEMORY.md, and clean up stale entries — zero API cost.

🤖 Charlie's Daily Monitoring Sweep

Every weekday at 09:00, Charlie runs a monitoring sweep: lists all cron jobs, checks for errors, compares against the dashboard calendar, and reports any discrepancies. If everything is healthy, he stays silent. If something's broken, he alerts in #charlie.

💡 Why MoE Changes Everything

MoE (Mixture of Experts) models have huge total parameter counts but only activate a fraction per token. This means:

• Speed: Only 22-37B params compute per token → fast inference
• Quality: 1T total params = the model has vastly more knowledge stored
• Catch: ALL parameters must be loaded into memory, even though only a few activate. So you need the RAM to hold the full model.
• RTX 5090 (32GB): Can only hold active params → runs MoE models at "small model" speed but loses quality of full model (offloading to system RAM kills speed)
• Mac Studio 512GB: Loads the ENTIRE model → full quality at moderate speed

📊 What Makes Sense?

🍎 The 512GB Sweet Spot

• Kimi K2 (1T/32B active) — Best agentic model, fits fully loaded (~550GB with Q4). Would need tight quantization.
• DeepSeek-R1 (671B/37B active) — Best open reasoning model. Fits comfortably at Q4 (~370GB).
• Llama 3.1 405B — Dense, all params active. ~230GB Q4. Fits easily. GPT-4 class.
• Qwen 3 235B (full load) — ~130GB Q4. Could run alongside Whisper + other tools simultaneously.
• Multiple models at once: Load Qwen 3 + Whisper + embedding model all in memory

⚡ Speed Reality Check

819 GB/s bandwidth on M3 Ultra. For a 405B Q4 model (~230GB):

• ~3.5 tokens/sec — readable but not instant
• Good enough for: overnight batch processing, long analysis, memory consolidation
• Not great for: interactive chat, real-time responses
• Waiting for M4 Ultra? Likely ~1,000+ GB/s → ~4.5 tok/sec, meaningful improvement

💰 Cost Comparison (once, not recurring)

• Mac Studio M3 Ultra 512GB: ~€10,000-12,000
• 2× RTX 5090 PC build: ~€6,000-7,000 (but only 64GB VRAM, can't run 405B)
• RTX 6000 Ada (48GB): ~€7,000 for GPU alone (still only 48GB)
• API cost equivalent: At ~$15/M tokens (Opus), €10k buys ~600M tokens. If you burn 1M tokens/day = 600 days of API. Mac Studio pays for itself in ~2 years of heavy use.