Descriptions:
Recorded at the AI Engineer (AIE) conference, this Latent Space crossover episode brings together Matthew Berman and swyx for a wide-ranging conversation that moves from conference origin story through chip geopolitics, government AI regulation, and ultimately to existential risk timelines. Swyx explains that AIE was modeled on how front-end and cloud engineering professionalized — each developing dedicated conferences, influencer ecosystems, and distinct tech stacks — and that he saw AI engineering following the same trajectory. The conference’s value proposition is framing itself as a neutral venue where all major labs compete on equal footing, which swyx argues both engineers and the labs themselves find valuable.
The chip discussion covers Etched, which came out of stealth during the conference week. Swyx frames it as a next-generation Cerebras — a specialized inference accelerator optimized for post-transformer workloads, built on the bet that GPT-3.5-style architecture will persist long enough to justify dedicated silicon. The group discusses why this bet is more defensible than it might appear: once a model architecture works at scale, organizations rarely migrate even as new architectures emerge.
The most provocative segment addresses AI governance and existential risk. The hosts analyze White House oversight of model release timing as an “art of the deal” opening gambit — extreme initial positions that get rolled back — and compare it favorably to fragmented state-level attempts like California’s SB 47. Swyx shares a p(doom) near 90% over a 50,000-year timescale, framing superintelligent AI emergence as analogous to how human civilization outpaced slower life forms — a natural transition that may not favor its creators.
📺 Source: Latent Space · Published July 10, 2026
🏷️ Format: Podcast
