E135: Wagner rebels, SCOTUS ends AA, AI M&A, startups gone bad, spacetime warps & more

Sacks argued that Prigozhin’s mutiny was driven partly by Wagner’s planned absorption into Russia’s Defense Ministry, which threatened his income and status. He said roughly 8,000 fighters entered Rostov, 3,000 headed toward Moscow, and the lack of elite backing left Prigozhin to accept a Lukashenko-brokered deal. (Reuters)

The affirmative-action debate exposed a separate Harvard preference system: more than 43% of white admits were athletes, legacies, donor-linked applicants, or children of faculty and staff. NBER estimated roughly three-quarters of those white ALDC admits would have been rejected as ordinary applicants, making legacy and athletic preferences a major parallel advantage. (National Bureau of Economic Research)

Calacanis said AOL once refused to promote him from SVP to EVP explicitly because he was white, while offering equivalent bonus compensation. He presented it as a first-person example of informal demographic balancing affecting a corporate promotion decision, rather than as a general claim about all DEI programs.

Sacks disclosed that his firm was preparing to invest in MosaicML at a $400 million post-money valuation before Databricks’ acquisition interrupted the round. He said another investor then offered a $700 million valuation, while MosaicML’s founder privately told him the acquisition was underway; the announced headline price was about $1.3 billion.

Calacanis said four of seven companies in his latest accelerator batch would not have been economically viable before recent language-model APIs. He also said startups he meets typically test six to eight models and do not consistently choose OpenAI, implying the early market was already multi-model rather than winner-take-all.

Inflection AI’s $1.3 billion funding round looked less like ordinary startup financing than a massive compute buildout. Chamath estimated that 22,000 H100 GPUs alone represented about $900 million of capex, arguing that much of the capital was effectively flowing through to Nvidia and cloud providers rather than funding a conventional software organization.

IRL’s collapse gave the panel a concrete example of boom-era venture diligence failure: a board investigation found 95% of its claimed 20 million users fake after a $170 million round at a $1 billion-plus valuation. Sacks argued giant funds magnify errors because they must deploy $200 million where smaller funds might risk $10 million.

NANOGrav’s 15-year pulsar-timing dataset was the episode’s strongest science item, providing evidence for a very-low-frequency gravitational-wave background. Friedberg said the method could eventually map supermassive black-hole systems and add a new “fingerprint” of the universe alongside cosmic microwave background measurements. (nanograv.org)