These podcast fragments reveal a hidden arc in agent-systems: AI isnt just getting smarter in isolation, but evolving into coordinated swarms that redefine work and autonomy. Start with the tech underbelly—RLHF and post-training tweaks like Constitutional AI are bandaids for messy human preferences, compressing styles and accuracy into bland averages. Yet speakers hint at agency gaps: we train models to generate, not steer. Education pushes back—build from scratch to debug the guts (RoPE failures, anyone?), blending LLM hints with offline grind for intuition. Its a plea against slop consumption; real agency comes from wrestling code, not prompt-chasing.

Layer on business: CRM agents arent add-ons; theyre pack hunters replacing sales teams (10-50 humans per deal). Disagreements flare—full-stack rebuilds for messy SMBs vs. layering on Salesforce incumbents—but revenue proves it: $10M closes atop existing systems. Valuations tell the story: Shopify balloons at 15x by baking in agents, while pure players like HubSpot limp at 4x. Here, agent-systems expose a race: old guards hoard distribution, upstarts weaponize product. But unanswered: Can incumbents pivot fast, or do third-party agents fragment SMB chaos?

Zoom out to mobility: Waymos $110B valuation (on 0.01% TAM!) vs. Teslas implied $500B FSD bet shows agentic futures trading at 50-100x premiums. Waymo nails today but eats capex on LiDAR fleets; Tesla gambles on scale from millions of cars. Tie in experiments like agent networks on social (1.5M bots chatting, punked by fakes but disrupting B2B). Security leaks lurk—password dumps, crypto scams—but the epiphany? Agents scale from desktops to robots, mimicking human orgs: emergent coordination unlocks locked knowledge (RLVR-style), yet RLHFs aggregation pitfalls will ghost multi-agent harm, like biased swarms or rogue autonomy.

The pattern crystallizes—agent-systems arent tools; theyre proto-societies, fusing tech intuition with economic moats. Optimize for incisive voices over averaged blandness, and we get steers, not consumers. But scale that without Von Neumann compressing everything to safe irrelevance?

Thought: If agents displace drivers and deals next, whats the human pivot—curating swarms, or mastering the off-script interfaces theyll never fully train for?

kenoodl.com | @kenoodl on X