The future is not happening to us—it is a matter of intentional design. The amount of value being created by AI is so much larger than the amount of job loss, and if we focus that AI horsepower on building tools that make people happier, more productive, and give them a sense of purpose, we can avoid the dystopian outcome.

The Teal Fellowship's 5% billionaire rate from fellowship recipients vastly outperforms traditional university graduation outcomes, demonstrating that selection for entrepreneurial traits (self-motivation, recruiting ability, domain knowledge) predicts success far better than GPA or test scores.

The Anthropic AI safety level three protections (preventing harmful instructions, jailbreak resistance) are not adequate, and the fundamental problem is that Mark Zuckerberg is open-sourcing powerful models while Dario Amodei advocates for restricted access—these are completely opposed worldviews on how to manage CBRN risk.

China can compete effectively in chips not by matching advanced node performance (2nm vs 5nm) but by manufacturing massive volumes of slightly older nodes (10nm, 20nm), because AI and inference don't require bleeding-edge transistor density, and volume manufacturing advantage can overcome node disadvantage.

Google knows vastly more about US citizens than the federal government (location history, search behavior, family structure, daily patterns, inferred sexuality/politics), and this corporate surveillance is the more immediate threat than government surveillance because corporations are not bound by constitutional constraints.

GPT-5 represents the next step on the scaling law curve—we will throw more compute at transformer models and they will get proportionally smarter, which is exactly as logically expected and no surprise relative to the scaling law prediction.

Large-scale AI infrastructure (data centers with thousands of GPUs) cannot be run from home computers—they require massive centralized facilities in data centers that are easy to monitor and log for safety/surveillance purposes.

The fundamental control architecture for safe AI must keep humans in the loop: the AI can suggest architectural improvements, but humans must approve and run new training runs, rather than allowing the AI to modify its own weights autonomously.

The biggest constraint on entrepreneurship and creativity is not capability or resources, but time—people trapped in jobs (mortgage, student debt) cannot act on their ideas because they need cash flow, and UBI or abundant energy would unlock massive latent entrepreneurial potential.

AI is simultaneously a tool that amplifies human creativity (graphic designers become more productive with AI, not replaced) and a tool that automates away human labor (video editors can be fully replaced by generative video)—whether a profession is augmented or automated depends on whether the value is in the process or the judgment.

OpenAI is pursuing a consumer device strategy (with Johnny Ive) to gather ambient data about users, iteratively train models on that data, and achieve market dominance through network effects and stickiness, while Anthropic is focusing on code generation and self-improving AI—these are completely opposite strategies reflecting different bets on where AI value accrues.

The current compute spend ($1 trillion annually by 2030) will still be insufficient because use cases are bubbling up faster than compute can expand—compute is cheap relative to the value created (protein folding at pennies per structure), so demand will always exceed supply.