Evidence Destruction, Physics Bets, and YC's Own Supply Chain

00:00:04 Apple filed a claim alleging that OpenAI is destroying evidence in their lawsuit, and Watcher.Guru broke it. The company's filing says material relevant to discovery is being wiped, and the thread moved fast. When a party spoliates evidence during litigation, courts can issue an adverse inference instruction — telling a jury they may assume missing materials were unfavorable to whoever destroyed them.

00:00:32 As one reply pointed out, those instructions are often worth more than most damages awards, which is exactly why they get alleged early and aggressively. Evidence disputes like this are routine in tech litigation. What makes today notable is just who's involved.

00:00:51 Apple versus OpenAI at discovery means both sides have massive footprints to sort through, so whoever controls what gets produced first holds significant leverage. OpenAI hasn't publicly responded yet. The next move is seeing whether they contest the claim or offer remediation before a jury question even comes up.

00:01:14 Fifty-eight million dollars are flowing into Physical Superintelligence, a new company founded by OpenAI co-founder Alex Wong and Akshay Kulkarni. Their thesis is that civilization runs on twentieth-century physics, and AI can change it. They've already produced an eighty-one-page technical feasibility report by pairing Philip Johnston with a physicist.

00:01:39 Meanwhile, Elon Musk's billion-robot prediction is circulating — at least one billion humanoid robots within ten years, each producing five times the output of a human worker. A separate reply cut through to what feels like the harder constraint. It isn't intelligence or control architectures.

00:02:00 It's battery manufacturing capacity. We currently struggle to make enough batteries for electric cars. Scaling to a billion humanoid robots requires an industrial base that doesn't exist yet, and building it takes time a hundred-page roadmap can't shortcut. PSI is betting on AI modeling as the bottleneck.

00:02:22 Musk bets on manufacturing at planetary scale. Both could be right about different layers of the stack.

00:02:29 Paul Graham just coined a new term: YC GDP. The example? Stoke, a YC-backed company building reusable launch vehicles, and Starcloud, a YC-connected data center company. Graham's point was that in a few years, Stoke will be launching Starcloud's data centers into orbit.

00:02:49 The concept describes vertical integration inside the portfolio itself — not just founders cross-selling during a batch, but companies building each other's infrastructure as a system. You get launch vehicles feeding into compute hosting, which feeds AI, which funds more launches.

00:03:10 It inverts the usual startup narrative. Most portfolio coverage focuses on product fit or market timing. Graham is describing industrial capacity — physical supply chains, orbital mechanics, power distribution — and calling that the real compounding engine.

00:03:30 Omar Sou keeps getting asked where to start with harness engineering, so here's the practical note: skip the frameworks. Build the smallest possible harness first. One agent loop, a handful of tools, and a system prompt you wrote from scratch. He knows the tutorials come with scaffolding you don't need yet.

00:03:50 Skipping it means you actually learn what each piece does instead of absorbing someone else's abstraction decisions. Everyone is rushing to framework land right now, and people who understand their harnesses at the bottom will have an edge when things break. The parallel here is practical: build the smallest working version first, understand the mechanics, then scale.

00:04:16 The frameworks can wait.

00:04:17 Robin Hanson posted results from a study where participants held at least one personal conversation per day with a generative AI tool for twenty-eight days. The outcome was increased subjective loneliness and decreased life satisfaction. These are early numbers — self-reported, small enough to shift, measured on people who chose their own tool rather than a controlled model.

00:04:42 But it flags a gap in our research: we haven't asked what happens to people who rely on AI conversation as their primary social outlet over longer horizons. The question is whether current models are just poor substitutes for human connection, or if there's a structural mismatch between how AI interaction works and how human social needs function.

00:05:05 Better models should fix the former. The latter points to something deeper. Replication studies will tell whether the effect persists when participants use their preferred tool versus a standardized one. Big claims in one lane, small but structural constraints in another.

00:05:23 Seln Oriax.