To be honest it's hard to bet against @cognition after watching this pod (and also as a Devin user myself). So many insightful insights: 1. "Building a company feels like a tree search" Scott describes his brain as running a constant strategy game, calculating moves, exploring decision trees, trying to find paths to victory. 2. The "salty" framework, losing hurts more than winning feels good, but not enough to stop trying. Both Scott and the great founders Senra studies (Larry Ellison, Michael Dell) share this trait: the pain of losing massively outweighs the pleasure of winning. But critically, Scott adds: "not by enough that it makes me want to stop trying." That asymmetry is what keeps people in the game. 3. Why agents working for MONTHS changes everything The METR report showed AI could do ~10-20 seconds of unassisted work a couple years ago. Now it's hours. Scott's first-principles question: why can't it be days, weeks, months? When an agent can work on a mission for a year, you stop giving it tasks and start giving it missions, solve this societal problem, build this video game from scratch, explore this novel materials science approach. The abstraction level of human work shifts from execution to vision. 4. Software economics are about to break, "one-time-use software" becomes viable Right now, building software only makes economic sense if it'll be used thousands/millions of times (the YouTube test). But most white-collar work is essentially one-time-use software that nobody would ever hire a team to build. When agents can write throwaway code for single-use tasks, that entire category of work gets automated. 5. "Survival mode → Creative mode" (Minecraft metaphor) Scott's co-founder's line: humanity has been in Minecraft survival mode, farming, staying safe and grinding. AI pushes us into creative mode with unlimited resources. The only question is what do you want to build? Scott genuinely believes this transition happens within 5 years. 6. How Cognition actually found PMF: the Nubank story They deployed THE ENTIRE COMPANY to Brazil for their first customer. They sat with Nubank engineers, debugged specific problems, manually taught Devin their codebase patterns. The early product failed on ~87% of tasks (13% on SWE-Bench). PMF came from finding the right task shape first: repetitive but not scriptable (migrations, version upgrades), but scoped enough for tight feedback loops. 7. "Measure ROI, not token spend" Scott pushes back hard on the "rank engineers by tokens spent" trend. His framing: measure actual output shipped, not API consumption. Cognition is incentive-aligned with customers to make token spend efficient because they're model-neutral. They don't profit from customers burning more compute. 8. Why focus lets startups beat giants, the Daniel Ek / Spotify principle "We're just going to care way more about [software engineering] than they are." Microsoft has GitHub, Copilot, VS Code, and an OpenAI partnership. Cognition has ~100 people and one obsession. Scott's argument: if you do everything, you lose to the trillion-dollar company that also does everything. If you make a concentrated bet on one thing and execute ruthlessly, the big company's diversification becomes a liability. Jimmy Iovine's Apple Music experience confirms this. 9. The independence stance, "we would sell if it were the most ambitious thing to do" Scott has turned down acquisition offers (he won't say how many). His reasoning isn't financial, he doesn't own a car, lives in an apartment. But he couldn't live with the outcome of not finding out what they could have built. "If we tried and we weren't good enough, I'd be salty but I could live with it. If we held back, I couldn't." 10. The first time Devin worked, Scott couldn't sleep The task was mundane: setting up MongoDB. But seeing an AI reproduce the entire debugging loop (hit error → Google it → try fix → hit different error → iterate) autonomously for the first time gave him a visceral sense of the exponential ahead. "Why shouldn't all software be built this way now?"

The only 5 ways to make yourself AI proof: 1. Develop Taste 2. Learn to sell 4. Build a high value network 3. Learn + create AI systems 5. Build an audience on social media

SpaceX will become one of the frontier AI labs after acquiring @xai and @cursor_ai They've been jointly training a model for months. It's shipping in both Cursor and Grok Build (xAI's coding CLI that already runs parallel agents). Think about what that means: Colossus (xAI's supercomputer) trains the model → Cursor's millions of developers generate the feedback → that feedback improves the next model → which ships back to Cursor. Closed loop. No API dependency on Anthropic or OpenAI. Now they own both the model AND the distribution

Everyone's scrambling to find a moat in the new AI world. I read Chamath's 138 page report to find what that means at every layer of the stack:

Multi Agents are FINALLY working. Everyone thinks multi-agent = throw 10 AIs at a problem and let them cook in parallel. That doesn't work in pratice. When multiple agents write code simultaneously, they make conflicting decisions — different styles, different edge case handling. It's chaos. Cognition and Anthropic independently discovered the same architecture. One agent writes. The others contribute intelligence. Anthropic calls it orchestrator-worker. Cognition calls it "single-threaded writes with multi-agent intelligence." Here are the 3 patterns that actually work: 1. The Self-Review Loop Devin reviews its own code with a separate clean-context agent. Why? Context rot — models make worse decisions at longer context lengths and The fresh reviewer catches what the coder can't see. Result: 2 bugs found per PR, 58% of which are severe. 2. Smart Friend Two frontier models (say Opus 4.7 and GPT 5.5) working together. Different models excel at different subtasks — some debug better, some reason visually better. The unlock is routing to each model's strengths. I personally use Codex to review code made by Claude Code and it catches critical issues constantly. 3. The Research Swarm Anthropic's version: one lead researcher (Opus) spawns parallel search agents (Sonnet) that fan out, gather info, and report back. these workers are read-only 90% improvement over single-agent. 15x more tokens but handles parallel research beautifully. The pattern everyone's converging on isn't "more agents." It's the right agents doing the right jobs, with one clear decision-maker. If you're interested in the full research, follow + comment AGENT and i'll send you the links :)

almost punched my cofounder for trading Kanye prediction markets during his concert… but this means we’ll escape the permanent underclass right?