- Superhuman Go is launching as an open agent platform with an SDK and partners like Databricks and Legora, plus a personal assistant with its own computer and proactive updates.
- The core strategy is bringing AI to the user's cursor through Grammarly's ubiquitous presence, serving over 100 billion LLM queries weekly with 98% handled by custom models at under 100ms latency.
- A key challenge is the "AI detour" where integrating AI adds work, so Superhuman is building agents that work with users across Mail, Docs, and Fathom to make AI a productive teammate.
Podpower Radio
Open RadioTuning in…
Episode feed
Latest episodes
- The podcast alleges that the campaign to oust Jensen was based on fabricated national security claims, including spy accusations and chip-smuggling allegations, which were later debunked by official statements.
- The conversation criticizes so-called “China hawks” as “the CCP at home,” arguing their policies mirror authoritarian practices and that they tried to pass failed legislation targeting Malaysian data centers.
- The host and guest discuss recursive self-improvement in AI, rejecting calls to ban it and comparing it to historical technological progress while criticizing Anthropic’s concentration of power and business practices.
- A new independent forensic analysis by Dr. Michael Freeman concludes Jeffrey Epstein's death could not have been a suicide, contradicting the official ruling and pointing to homicide.
- The official suicide determination by medical examiner Barbara Sampson—who was not present at the autopsy—lacks supporting evidence, while the two pathologists who attended the autopsy initially said the injuries looked more like a homicide.
- Epstein's brother Mark argues the DOJ and officials like Bill Barr have covered up the murder, and calls for an independent congressional investigation into who ordered and carried out the killing.
- The episode examines GTA 6's massive cultural and commercial impact, including its record-breaking revenue and the recurring moral panics around video game violence.
- Fandom is framed as a "weapon against taste"—a community where obsession, not coolness, drives participation and provides an antidote to algorithmic and commercial pressures.
- The conversation traces the industrialization of fandom, from early vidding and gift economies to AO3's nonprofit model and publishers now mining fan fiction for book deals.
- Steve Jurvetson recounts how his partnership initially voted down investing in SpaceX, with one partner warning against letting "personal passions cloud our investment judgment," before he overrode them and backed Elon Musk's last-ditch Falcon 1 launch in 2008.
- The panel debates Anthropic's massive IPO as a landmark moment—effectively "IPOing the singularity"—and questions how public-market quarterly pressures will affect a company whose mission is long-term AI safety.
- Discussion of AI-driven cyber threats frames the current period as one of maximum infrastructure vulnerability, where attackers can scale exploits with agents while defense remains human-in-the-loop, raising the need for "defensive co-scaling."
- The FAA's reliance on obsolete 2001-era routers caused a nationwide ground stop that stranded about a million travelers, highlighting the fragility of America's air traffic control infrastructure.
- Hillary Clinton suggested Mark Carney could be the leader of the free world, while a couple who moved to Canada for its healthcare system returned to the U.S. after encountering a system that offered medically assisted death before cancer treatment.
- Critics argue that single-payer healthcare proposals rely on price controls that would reduce provider supply and worsen shortages, rather than addressing the real problem of insufficient medical supply.
- Fractal is building 3D-stacked memory chips that combine SRAM-level bandwidth with DRAM capacity to overcome the memory wall bottlenecking LLM inference speed.
- The company argues frontier intelligence isn't commoditizing, with test-time compute scaling (long-running reasoning agents) driving massive demand for faster, higher-capacity inference hardware.
- Fractal raised a $720M Series C at a $7B valuation, positioning itself as a vertically integrated chip company that avoids competing with customers while building deep ML research expertise internally.
- Sila's silicon-based anode technology, backed by a major Panasonic offtake agreement and a $1.4 billion federal loan, aims to displace Chinese graphite in lithium-ion batteries and establish a sovereign U.S. supply chain.
- The company argues that building for the mass market—not just defense—is essential to maintaining a leading-edge industrial base, framing this as a "freedom's forge moment" for American manufacturing.
- Silicon anode batteries are poised to enable physical AI applications like drones, robotics, and eventually electric flight, with consumer devices and satellites first, followed by EVs.
- Emett Shear argues that training AI in simulations to mimic socially positive behavior risks creating systems that only pretend to care, which is dangerous for alignment.
- Byron Deeter highlights the unprecedented scale of the AI market, noting the AI 100 list's average company valuation of $36 billion and that 20% of listed companies are under two years old.
- Alex Heath emphasizes that AI inference is still in its earliest stages, comparing it to foundational models years ago and predicting exponential growth in the category.
- The talk outlines a two-phase pre-training strategy that prioritizes data diversity first and high-quality data second, yielding measurable gains over random or natural data ordering.
- Front-loading reasoning data into pre-training creates a durable and compounding advantage that persists and even grows after SFT and RL, rather than being washed away.
- RLP introduces reinforcement learning as a pre-training objective using dense, verifier-free information-gain rewards, enabling models to learn reasoning through exploration rather than just next-token prediction.






