David Zax hosts IBM’s Mixture of Experts with Kush Varshney, Sridhar Rajagopalan and Alexis Zamkow in this 28-minute discussion.
Mixed chips create a software optimization problem — 0:58
- Anthropic’s public debut frames a discussion about buying compute across NVIDIA, Amazon and Google.
- Rajagopalan argues that hardware supply limits push large buyers toward several chip families, each suited to different workloads.
- Small improvements at the hardware–software boundary can accumulate across a large fleet.
- Zax’s “hardware ice cream truck” metaphor captures the variety; exploiting those different architectures takes additional software work.
- Zamkow raises the competing concern: investors supplying compute can concentrate financial and infrastructure power in a small ecosystem.
Agent payments need bounded authority — 10:50
- The panel discusses ChatGPT Wallet: a person reviews a purchase, approves it, and the agent pays using a substitute card credential.
- OpenAI’s documentation confirms biometric approval and an initial rollout through dots for eligible US Pro users.
- Rajagopalan connects short-lived, narrowly scoped credentials to practices used by teams operating large online services.
- Zax describes a different arrangement: Claude used his logged-in browser to buy a cable after conversational approval.
- His later $850 projector purchase illustrates another risk: he checked the return policy but had yet to test the product.
- Varshney invokes the “market for lemons,” where buyers struggle to judge quality and sellers lack information about buyers.
- Caveat: the panel’s five-minute credential example is not a verified Wallet guarantee; OpenAI says approved recurring plans may permit later merchant charges.
Drug discovery depends on selecting viable molecules — 19:45
- Isomorphic Labs announced a $5 billion Series C on November 17, co-led by Greenoaks and Sequoia.
- Rajagopalan describes computational screening that reduces a huge candidate space before expensive laboratory and human testing.
- Better candidate selection could improve the odds of later success; the discussion supplies no measured clinical success-rate improvement.
- Zamkow extends the idea to biomarkers and selecting trial participants, explicitly acknowledging that the hoped-for acceleration remains unproven.
- Varshney’s Chef Watson analogy separates generating many possibilities from choosing useful ones.
- Toxicity and solubility matter alongside efficacy: a molecule must be tolerable and practical to administer.
- Learned predictions from molecular data can guide selection without requiring a complete simulation of the human body.
“The selection part is the hard part.” — Kush Varshney, 25:46