Superhuman's growth plateaued because it was hyperfocused on a single ICP (founders); preventing a plateau requires proactively expanding to additional ICPs.

Valar Atomics refused to call itself a nuclear startup until it had split its first atom, making hardware execution the identity bar.

As models improve, they will be better at designing their own agent architecture than humans, making human-fixed architecture a bottleneck.

Almost no food on Earth has not been genetically modified, making the natural-vs-GMO distinction false.

Leveraged ETF investors do not switch funds to save fees because leveraged ETFs are trading vehicles, not long-term holdings.

Paul Rand's logo design terms were non-negotiable: a flat $100,000 fee, exactly one logo presented, and the client must accept it.

Agent products that cannot be tried immediately and require heavy implementation are dead on arrival.

Chess.com's founder inverted the standard VC-backed startup playbook, instead of hiring a Stanford technical cofounder, raising money, getting an office, targeting a big market, and spending on paid acquisition, he hired a friend from San Jose State, worked remotely, raised no money, paid nothing for customer acquisition, and picked a tiny market.

The energy system will never return to January 2026 conditions; the 'grand bargain' is dead.

The next lab to run Evo's virus-design method has no obligation to adopt Arc's safety choices, so similar results could be produced without safeguards.