No PriorsNo Priors Ep. 138 | The Best of 2025 (So Far) with Sarah Guo and Elad Gil
CHAPTERS
- 0:00 – 0:30
2025 AI highlights montage: what to expect in this “Best of (So Far)” episode
Sarah Guo frames the episode as a curated set of standout moments from No Priors in 2025 to date. She previews the range—from startup builders to legendary researchers—and sets up the theme of recognizing and exploiting new capabilities at the right time.
- •Recap format: favorite clips from the year so far
- •Guests span companies like Harvey, OpenAI, Glean, Abridge and more
- •Includes science and industry legends (Fei-Fei Li, Noubar Afeyan)
- •Theme: leaning into newly unlocked AI capabilities
- 0:30 – 1:00
Winston Weinberg (Harvey): spotting GPT-3’s hidden product moment in legal workflows
Harvey’s CEO describes how seeing GPT-3—before widespread hype—created a sharp conviction that something important was being missed. By mapping model capability to real legal work, they discovered a practical wedge into a conservative domain.
- •Early exposure to GPT-3 sparked surprise that few were using it
- •Connected model output to concrete legal workflows
- •Looked for an “opportunity hidden in plain sight” in professional services
- •Set up experimentation mindset rather than waiting for consensus
- 1:00 – 2:10
Harvey’s early validation: prompt chaining, blind evaluation by attorneys, and an OpenAI cold email
Weinberg details a scrappy evaluation: using legaladvice questions, crafting early chain-of-thought-style prompts, and having lawyers judge outputs without knowing AI was involved. The strong pass rate became proof enough to reach out directly to OpenAI leadership.
- •Used ~100 landlord-tenant questions as a test set
- •Created prompt chains before “chain of thought” was mainstream
- •Blind test with 3 attorneys; 86/100 answers deemed sendable without edits
- •Cold-emailed OpenAI’s general counsel; results surprised OpenAI itself
- 2:10 – 3:11
Dr. Fei-Fei Li: why spatial intelligence is one of evolution’s hardest problems
Fei-Fei Li explains spatial intelligence as the challenge of reconstructing a 3D world from limited sensory input, enabling navigation and interaction. She argues it’s foundational to intelligence and not fully solved even by humans and animals.
- •Spatial intelligence = building 3D understanding from visual signals
- •Enables navigation, interaction, and (for humans) sophisticated manipulation
- •Evolution solved it partially, but it remains deeply challenging
- •Frames spatial reasoning as core to next AI capability frontier
- 3:11 – 4:25
From mental models to editable 3D worlds: what better spatial AI could unlock for people
Fei-Fei contrasts everyday human limitations in imagining detailed 3D environments with the promise of tools that make 3D representation fluid and editable. She points to how training and expertise (architects/designers) shape these abilities and how AI could broaden access.
- •Most people struggle to generate complex 3D models mentally
- •Spatial skill improves with training and talent
- •AI could make 3D interactivity far more accessible and malleable
- •Implication: new workflows and creative possibilities built on spatial tooling
- 4:25 – 5:22
Brendan Foody: rapid job displacement, political backlash, and the core economic challenge
Foody predicts fast-moving workforce disruption across many roles, with painful social consequences and likely populist political reactions. He centers the hardest question as how society reallocates work and wealth as AI capabilities compound toward superintelligence.
- •Displacement across roles will happen quickly and painfully
- •Expect major political effects (populist movement)
- •Key problem: what displaced workers do next (e.g., support, recruiting)
- •Wealth allocation becomes central, especially under power-law gains
- 5:22 – 6:20
Where people go next: physical-world work, niche skills, and why atoms lag bits
Pressed on outcomes, Foody suggests more work shifts into the physical world and human-centric services, alongside niche skills. He argues physical automation progresses slower than digital due to fewer self-reinforcing improvement loops and real-world constraints.
- •More jobs in the physical world and roles needing human interaction (e.g., therapy, service)
- •New categories like robotics data creation
- •Physical-world automation lags digital transformation
- •Digital systems improve faster via self-reinforcing feedback loops
- 6:20 – 7:20
Dan Hendrycks: superintelligence geopolitics through the lens of nuclear deterrence
Hendrycks draws an analogy to nuclear strategy: shared vulnerability can deter first strikes. As AI becomes pivotal to national power, states may attempt deterrence to prevent rivals from converting AI advantage into a decisive strategic weapon.
- •Nuclear deterrence analogy: retaliation risk prevents first strikes
- •AI viewed as pivotal to national futures increases strategic tension
- •Fear: AI advantage used as a ‘super weapon’ to crush rivals
- •Deterrence dynamics may emerge around AI development and deployment
- 7:20 – 8:21
Escalation risks: preemptive cyberattacks, espionage, and AI’s destabilizing competition
Hendrycks outlines how rivalry could intensify into preemptive actions like cyberattacks on data centers and pervasive surveillance of AI projects. He sketches a multipolar dynamic where the U.S., China, and Russia respond to perceived breakthroughs and shifting power balances.
- •Potential for preemptive cyberattacks on AI infrastructure (e.g., data centers)
- •Escalation as AI R&D becomes highly automated and fast-moving
- •Russia/other actors reassess as AI visibly impacts software and economies
- •Espionage and ‘keeping tabs’ on projects becomes easier and more common
- 8:21 – 10:46
Noubar Afeyan: making entrepreneurship less random and more like a disciplined science
Afeyan recounts building companies as an immigrant entrepreneur and questioning why venture building is treated as improvisational and ‘gamey.’ He argues that in domains like healthcare and climate, entrepreneurship should be practiced as a rigorous profession—and AI can help systematize it.
- •Personal origin: founding companies when that path was atypical for young immigrants
- •Critique: entrepreneurship framed as random, idiosyncratic, score-keeping ‘game’
- •Call for entrepreneurship as a profession with repeatable methods
- •High-stakes sectors (healthcare, climate, food) demand rigor; AI can assist
- 10:46 – 12:58
OpenAI on reasoning models: uncertainty awareness + tool use drives better test-time scaling
Brandon McKinzie and Eric Mitchell explain how reasoning models can recognize uncertainty (e.g., weak visual confidence) and become more effective when they can act through tools. Tool use allows the model to spend tokens productively—like manipulating an image or delegating computation—steepening performance gains at inference time.
- •Models can transparently indicate uncertainty (“I can’t see it well”)
- •Tool use enables actions like cropping/manipulating images for visual reasoning
- •Improves token efficiency and increases test-time scaling slope
- •Delegating to tools beats forcing the LLM to do verifiable computation internally
- 12:58 – 14:05
Isa Fulford: training DeepResearch—the ‘taste and grind’ behind breakthrough browsing behavior
Fulford describes the moment a new dataset and training algorithm first ‘worked’ for browsing tasks—surprising even with strong prior conviction. She also notes the counterintuitive failures: models can do impressively smart steps and then make baffling mistakes, highlighting remaining headroom.
- •Conviction that training on browsing tasks would work—then amazement when it did
- •Breakthroughs often come from data generation, iteration, and taste
- •Hands-on ‘visceral’ validation is a key part of development
- •Models still fail in surprising ways; smart behavior can coexist with silly errors
- 14:05 – 16:30
Arvind Jain (Glean): turning “graveyard” enterprise search into a viable category via SaaS and scale
Jain explains why enterprise search historically failed—especially pre-SaaS, when data access and deployment were too hard. SaaS APIs, interoperability, and cloud-scale infrastructure created the conditions for a turnkey unified search product, born from an internal pain point at Rubrik.
- •Enterprise search was a ‘graveyard’ due to data access and deployment hurdles
- •SaaS changed the game: APIs, interoperability, uniform versions across customers
- •Origin story: internal inability to find info across hundreds of SaaS tools at Rubrik
- •Modern enterprises face internet-scale internal content (billions of documents) requiring scalable systems
- 16:30 – 18:21
Abridge and healthcare’s human impact: purpose, retention, and giving time back to clinicians
A story from Abridge illustrates how AI can materially improve clinicians’ lives by reducing documentation burden and extending careers. Elad Gil shares how frontline feedback—both critical and emotional—keeps teams aligned on mission and reminds them why the work matters.
- •Doctors provide detailed workflow feedback; teams treat it as essential input
- •AI tools can reduce burnout and delay retirement for clinicians
- •Emotional impact: clinicians spending more time with family (dinner story)
- •Mission and purpose as durable motivation beyond hypergrowth metrics
- 18:21 – 18:58
Closing reflection and calls to action: AI as a hinge moment, more conversations ahead
Sarah closes by framing these clips as evidence of a historic inflection point and previews more builder-and-thinker interviews to come. She invites listeners to engage via reviews, comments, guest suggestions, and subscribing across platforms.
- •AI described as a ‘hinge moment’ in history
- •Promise of more episodes with leading builders and thinkers
- •Engagement asks: reviews, YouTube comments, guest recommendations
- •Where to follow: social channels, subscriptions, and transcripts/newsletter
