Venture with Grace

There are a handful of genuinely non-obvious observations - token maxing as a Goodhart's Law trap, the near-zero attrition at Anthropic as a structural signal, and intelligence-per-watt as the coming optimization frontier - but the episode is heavily padded with generic VC wisdom ('people matter more than anything') and the guest frequently meanders without landing sharp points.

“I think the concept of token maxing is a transient phenomenon. I think we'll look back on it in a few years and think it was like a really silly, uh, somewhat nonsensical concept.”

“The intelligence per watt today I think we're going to look back on in a few years and it's going to be crazy like how resource inefficient these models are being used.”

A few moderately contrarian takes emerge - applying Goodhart's Law specifically to token-maxing, framing neo-lab valuations as structurally mispriced risk-reward bets, and predicting TSMC/ASML monopolies will break - but the bulk of the episode recycles standard VC narratives (focus wins, people first, AI talent is scarce) that circulate constantly in the space.

“it's like so obviously so ripe for good hearts law issues where like if you, if you're optimizing for that, you can obviously just max that out in a way that's not helpful or productive”

“if you're doing that at 50 million or 20 million post, it's a lot different than if you're doing that at 2 billion posts”

Rob Toews is a legitimate AI-specialist VC with a real deal track record - co-incubating Cohere in 2020 pre-GPT-3, World Labs, and others at a credible $3B AUM firm - and his non-VC background (Zoox, White House AV policy) adds genuine texture; the knock is that he is a fund manager commenting on founders rather than a practitioner who built something at scale himself.

“for the past five years or so I've been at Radical where I lead their San Francisco office and have led a number of our investments”

“we incubated Cohere. Cohere was actually the very first company we incubated back in 2020”

The episode has a reasonable density of named companies, specific fund sizes, and named individuals (Karpathy, John Jumper, CTO of Workday joining Anthropic as IC), but lacks hard data, cited metrics, or analytical rigour - several numerical claims (e.g. 'trillion dollars in five years') are asserted without sourcing and the portfolio name-drops often substitute for genuine analytical depth.

“The early fund is a $650 million fund. The growth fund is an $800 million fund.”

“John Jumper joining from D Mind is crazy like that. You know, he won the Nobel Prize with Demis. You know, he did alphafold with Demis”

The host's questions are consistently multi-part, syntactically incoherent, and full of filler ('like', 'um', 'I guess like'), making it hard for the guest to deliver focused answers; there is zero pushback on any claim, no genuine follow-up drilling into specifics, and the session ends with a rapid-fire segment that wastes the final minutes on trivia.

“I want to start with like you know um, the investing strategy of like Radical and obviously you guys have like an early stage fund and a leader stage slash like grow stage fund and um, leading the Bay Area or like west coast, um, team like how does it kind of like came together”

“I wonder maybe we could start with like. I really agree with you all. Like uh, there's like a few different profiles.”