No PriorsNo Priors Ep. 71: The Best of 2024 (so far) with Sarah Guo and Elad Gil
CHAPTERS
- 0:00 – 0:45
Mid-year “Best of 2024” kickoff: AI across research, hyperscalers, and startups
Sarah sets up a mid-year compilation episode highlighting standout moments from prior No Priors conversations. She frames the breadth of topics—from frontier research to practical company-building—and signals that links to full episodes will be included.
- •Mid-2024 recap format: favorite clips + context
- •Coverage spans AI research, hyperscalers, and upstarts
- •Promise of episode list for listeners to revisit full interviews
- 0:45 – 2:14
Fintech’s AI whitespace: merchant identity, compliance, and risk decisions (Emily Glassberg Sands, Stripe)
Emily explains that robust merchant identity and business understanding are foundational problems where AI can create outsized impact in fintech. She highlights how accurately mapping merchants to complex regulatory constraints underpins credit, supportability, and network requirements.
- •AI-driven identity resolution: who the merchant is and what they sell
- •Regulatory mapping as a hard, high-value problem
- •Downstream implications for lending, supportability, and card network/bin sponsor rules
- •Opportunity area for both startups and incumbents
- 2:14 – 3:14
AI-native integrations: making payments setup seamless without armies of engineers
Emily describes an opportunity to dramatically reduce the engineering overhead required to integrate financial products. LLMs can already generate code, but the challenge is building reliable automated integrations as edge cases and decision complexity increase.
- •Goal: user-specific financial integrations with minimal developer work
- •LLMs show promise for code generation and automation
- •Hard part is robustness as workflows and decisions get complex
- •No-code as a baseline; AI could push this much further
- 3:14 – 4:39
Payments data as a growth engine: using AI to help businesses run better
Beyond improving payment flows, Emily argues payments data can help businesses understand performance and make better decisions. She notes incentive alignment for platforms like Stripe—helping customers grow expands the overall economy and platform growth.
- •Payments data enables deeper business insight, not just transactions
- •Question of who builds it (startup vs incumbent) and how it monetizes
- •Aligned incentives: customer growth drives platform growth
- •Potential macro impact: “grow the pie” and GDP
- 4:39 – 6:32
Design with AI copilots: why iteration beats one-shot prompting (Dylan Field, Figma)
Dylan explains that creative work can’t be fully specified in a single prompt because good design encodes context, culture, and user psychology. He emphasizes the iterative loop—continuous back-and-forth with an agent—as the key unlock for AI-assisted creation.
- •Great design includes cultural/temporal context and emotional state
- •Prompts can’t capture all constraints and tacit knowledge
- •AI is most useful as a first draft + iterative collaborator
- •Agent interaction over time is central to creative augmentation
- 6:32 – 9:03
Will AI replace designers? Likely augmentation first—and more software overall
Responding to concerns about job displacement, Dylan argues near-term outcomes are more about access and efficiency than replacement. He predicts a surge in software creation, shifting engineers toward higher-level abstraction and increasing the total amount of designed software.
- •Short-term: designers remain essential due to context and judgment
- •Near-term impact: augmentation, efficiency, and broader access
- •Engineering shifts: higher abstraction, less hand-written code line-by-line
- •Future: much more software exists, even if interfaces evolve
- 9:03 – 10:19
Building humanoid robots at speed: iterate, prototype, and continuously ship (Brett Adcock, Figure AI)
Brett outlines Figure’s bias toward building and testing over extended analysis. He describes a philosophy of continuous hardware/software updates and an iterative approach to learning quickly through prototypes.
- •Iterative design as core thesis: build/test to surface problems
- •Continuous improvement mindset: hardware and software never ‘done’
- •Fast prototyping and testing to validate assumptions
- •Velocity requires tight learning loops
- 10:19 – 12:59
Hardware program management: requirements, design gates, and longer timelines
Brett walks through a structured development process: defining customer-driven requirements, then moving through conceptual, preliminary, and critical design reviews with company-wide involvement. He contrasts hardware’s methodical planning with software’s faster iteration cycles due to longer lead times and higher costs of mistakes.
- •Start with customer needs → translate into measurable requirements
- •Design gates: conceptual, preliminary, critical design reviews
- •Company-wide participation in key reviews
- •Hardware demands more up-front rigor because timelines are much longer
- 12:59 – 14:59
Sora’s creative impact: enabling artists’ storytelling (OpenAI Sora team)
The team reflects on how artists use Sora to tell stories more easily, citing examples like Shy Kids’ short featuring the character “Airhead.” They emphasize that the most inspiring outcomes come from creators’ narratives, not just visually striking clips.
- •Sora’s value is story enablement, not only visual novelty
- •Example: Shy Kids’ narrative short as a creative unlock
- •Artists push beyond the team’s own ideas
- •Focus on creator intent and expressive possibilities
- 14:59 – 16:22
From playful prompts to new media: timelines for short films and beyond
Asked about professional adoption, the team avoids exact timelines but expects increasing film-like content over the next few years. They also predict entirely new content formats and interaction paradigms as creators adapt to generative video workflows.
- •Cautious on exact timelines; expects growth in film creation over years
- •Generative video may spawn formats beyond traditional film
- •New interaction modes emerge when models ‘respond’ to intent
- •Historical analogy: evolution from shorts to long-form (e.g., Pixar)
- 16:22 – 17:57
Is video a path toward AGI? World modeling, simulation, and robotics implications
The discussion shifts to “world simulation” and whether video models could become useful for simulation workloads. The team highlights robotics as a key application, arguing video captures physical dynamics (joints, contact, motion) that are critical for embodied intelligence.
- •Video models as implicit world models with physical priors
- •Potential future: simulation capabilities inside generative models
- •Robotics as a major beneficiary of learning from raw video
- •Physical accuracy (motion/contact) as an important capability signal
- 17:57 – 19:12
Working with AI engineers: English as interface, fundamentals still matter (Scott Wu, Cognition/Devin)
Scott argues that while natural language becomes an increasingly powerful interface, core engineering fundamentals remain valuable. He compares this to networking: many use the internet, but deep understanding still matters for those building and troubleshooting systems.
- •“Hottest programming language is English,” but not sufficient alone
- •Foundations: computer internals, logic, algorithms still useful
- •Analogy: TCP knowledge remains valuable despite abstraction layers
- •AI changes interfaces, not the need for technical rigor
- 19:12 – 21:06
The future software engineer: architect + product thinker amid rapid AI impact
Scott predicts the engineer role evolves toward solution decomposition, technical architecture, and product-oriented decision-making. He’s cautious about long-term singularity forecasts, but confident that near-term economic and workplace impacts will arrive quickly and that the industry is still early in the cycle.
- •Engineer as hybrid of architect and product manager
- •Core work: break down problems and specify solutions well
- •Hard to predict superintelligence timelines with confidence
- •Near-term: large real-world impacts; still early since ChatGPT’s launch
- 21:06 – 22:49
Building trust in AI: Scale’s data lifecycle, evals, and transparency (Alexandr Wang, Scale AI)
Alex explains trust comes from improving the full AI lifecycle: high-quality data, rigorous evaluation, and repeated measurement. He argues that evaluation is essential for governments, enterprises, and labs to assess safety, capability, and deployment risk.
- •Data abundance + data quality as the foundation
- •Lifecycle loop: data → train → evaluate → iterate
- •Measurement builds confidence and enables responsible adoption
- •Stakeholders (governments/enterprises/labs) need ongoing evaluation
- 22:49 – 25:55
Why evals are hard: benchmark overfitting, held-out tests, and expert assessment
Alex details the pitfalls of academic benchmarks, including contamination and overfitting, and describes Scale’s held-out GSM1K evaluation to reveal gaps between reported and real capabilities. He advocates expert human evaluation, public leaderboards, and continuous testing infrastructure as models exceed human-level performance in narrow areas.
- •Measuring ‘intelligence’ is philosophically and practically difficult
- •Benchmark contamination/overfitting undermines confidence
- •Held-out evals (e.g., GSM1K) reveal true vs reported performance
- •Need expert evaluators, transparency, leaderboards, and continuous eval platforms
