No PriorsNo Priors Ep. 128 | With Andrew Ng, Managing General Partner at AI Fund
CHAPTERS
- 0:00 – 1:29
Andrew Ng joins: capability growth beyond just scaling
Sarah introduces Andrew Ng and opens with the big question: where will future AI capability gains come from. Andrew argues that while scale still has some room, the industry’s focus on scaling is partly PR-driven, and other vectors are increasingly important.
- •Andrew’s background (Google Brain, Coursera, AI Fund, Amazon board)
- •Capability growth has multiple vectors, not just bigger models
- •Scaling is getting harder and returns are diminishing
- •Agentic workflows, multimodality, and application work are major drivers
- •Wildcards/new paradigms (e.g., diffusion-style approaches) could matter for text too
- 1:29 – 2:44
Why Andrew coined “agentic AI”: a spectrum, not a binary
Andrew explains why he introduced the term “agentic AI” to move the conversation past unproductive debates about what counts as an agent. He frames agency as a continuum and notes how quickly marketing hype latched onto the term.
- •Motivation: stop debating “is it an agent?” and start building
- •Agency exists on a spectrum from low to highly autonomous systems
- •Term adoption accelerated due to marketers applying it broadly
- •Marketing hype has outpaced real business progress (though progress is real)
- 2:44 – 4:05
What’s blocking real agents: evals, guardrails—and especially talent
Elad asks what’s still missing for true agents in production. Andrew lists technical gaps (reliable computer use, guardrails, evaluation) but emphasizes the biggest bottleneck is talent: teams that can run disciplined eval-driven error analysis outperform those that iterate randomly.
- •Computer-use agents are still unreliable for production
- •Evals and guardrails are foundational and still too hard/slow
- •Biggest barrier is talent and disciplined engineering practice
- •Systematic error analysis with evals is the key differentiator
- •Tooling and skills lag behind the amount of automatable work
- 4:05 – 6:08
Why building agentic workflows isn’t easily automated (yet)
Sarah pushes on whether AI can automate the engineering process of building agents. Andrew argues much of the needed knowledge is proprietary and trapped in people’s heads, so humans must still do heavy lifting—at least in the next year or two.
- •Agent-building often requires external, company-specific context
- •Key knowledge isn’t in public pretraining data or manuals
- •Idea of AI “avatars” interviewing employees is plausible but not near-term
- •Human PMs/engineers must encode business-process judgments
- •Proprietary data and tacit knowledge are major constraints
- 6:08 – 7:18
Bleeding edge examples: coding agents work; “computer use” mostly doesn’t
Andrew highlights coding agents as the most impressive, economically valuable agentic systems today—alongside question-answering. He contrasts this with shopping/browsing “computer use” demos that look good but aren’t production-ready.
- •Two big value buckets: Q&A assistants and coding agents
- •Coding agents can plan, create checklists, and execute multi-step work
- •Andrew’s favorite current tool: Cloud Code (as of recording)
- •Computer-use agents (shopping/browsing) remain demo-heavy
- •Reliability and production readiness separate real value from hype
- 7:18 – 8:09
Why coding wins: resources, clear ROI, and builders as users
Elad asks why coding agents feel magical compared to other agent use cases. Andrew attributes it to obvious economic value, massive resource allocation, and the fact that builders are often power users with strong product intuition—“capitalism solving research problems.”
- •Coding has massive, measurable economic value (strong incentives)
- •More talent and iteration cycles concentrated in coding tools
- •Creators are often the users, improving product instincts and feedback
- •Not necessarily a fundamental barrier—priority and investment matter
- •“Capitalism is great at solving fundamental research problems”
- 8:09 – 9:04
Bootstrapping models: AI writing code and generating training data
Elad asks when models will effectively build themselves. Andrew notes major labs already use AI to write substantial code and highlights an emerging pattern: using older models in agentic loops to generate data (e.g., puzzles) to train the next generation.
- •Foundation model teams already use AI for significant coding work
- •Agentic workflows can generate synthetic training data for successors
- •Example pattern: older model “thinks longer” to create hard tasks
- •Next model learns to solve tasks faster/with less compute
- •Progress will come from multiple complementary mechanisms
- 9:04 – 9:55
“Vibe coding” vs AI-assisted coding: rapid engineering is still hard work
Elad probes Andrew’s dislike of the term “vibe coding.” Andrew argues it trivializes the cognitive demands of software building; AI increases speed, but the work remains deeply intellectual and mentally exhausting—better framed as rapid engineering.
- •“Vibe coding” implies uncritical acceptance of suggestions
- •AI-assisted coding accelerates building but doesn’t remove rigor
- •Engineering remains cognitively demanding; productivity comes with fatigue
- •Better framing: rapid engineering, not casual improvisation
- 9:55 – 11:34
Startups change: coding speeds up, product management becomes the bottleneck
Andrew explains how AI-assisted coding compresses build cycles from months to days, shifting the limiting factor to product decisions and user feedback. When you can ship daily, waiting a week for feedback hurts, so teams lean more on strong internal mental models and customer empathy.
- •Build time collapses; prototype work becomes dramatically cheaper/faster
- •The core loop shifts bottleneck from engineering to product management
- •Faster shipping makes slow user feedback cycles more costly
- •Teams rely more on gut informed by accumulated customer understanding
- •Deep customer empathy enables rapid, correct decision-making
- 11:34 – 12:50
Can AI speed product management? Early attempts lag coding acceleration
Elad asks about tools that automate PM (user interviews, simulated user flocks). Andrew sees promise but says these tools aren’t yet boosting PM productivity nearly as much as coding tools boost engineers, reinforcing the PM bottleneck.
- •AI for design/PM is emerging but not yet transformative
- •AI-assisted user interviews and simulated user panels are early-stage
- •Calibration/realism of simulated users remains challenging
- •PM acceleration is currently behind engineering acceleration
- 12:50 – 19:22
Founder profiles in 2025: technical product leaders and relentless work ethic
Sarah asks how the successful founder profile has changed. Andrew argues that in fast-moving AI cycles, founders must understand what models can and can’t do; technical, AI-fluent leaders have a major advantage, and working hard still strongly correlates with success.
- •Many 2022-era startup workflows no longer work in 2025
- •AI fluency is now critical for strategy and product direction
- •Technical leadership becomes more important during disruption
- •Work ethic matters; nuance for life phases and personal constraints
- •Bold conviction and decisiveness are common among world-changers
- 19:22 – 22:46
Finding great product people: empathy, synthesis, and serving one ICP early
Sarah challenges the scarcity of great product leaders. Andrew argues the key trait is empathy—synthesizing many signals into a strong mental model of the user—and shares a lesson from trying (and failing) to convert engineers into PMs. He also notes startups can go far by serving one representative user profile early on.
- •Great PMs build accurate user mental models from messy signals
- •Empathy is a strong correlate of strong product instincts
- •Not everyone should be forced into PM roles (don’t demoralize engineers)
- •Early startups can optimize for a single ICP/user persona
- •Some products (e.g., developer tools) can rely on “we are the user” intuition
- 22:46 – 27:19
Leading and hiring in the AI age: everyone must be AI- and code-literate
The conversation turns to how leadership and hiring standards shift as tools evolve. Andrew argues that not using LLMs makes people less effective, and at AI Fund even non-engineers learn to code to communicate precisely with computers; Elad notes product/design now value rapid prototyping with AI tools.
- •Leaders can’t rely on 2022 playbooks; AI changes the baseline
- •LLM competence is becoming as essential as web search skills
- •Coding literacy helps non-engineers specify tasks and automate work
- •Hiring preference shifts toward candidates fluent in AI tooling
- •Example: junior with AI tool fluency can outperform senior without it
- 27:19 – 32:12
Keeping teams small (carefully): leverage, coordination costs, and when to scale
Sarah and Elad debate the “stay small” ethos. Andrew believes teams can be smaller than before, but cautions that market dynamics matter (winner-take-all requires speed). Elad warns that underhiring can be a trap when incumbents arrive, even though small teams and AI can be extremely efficient.
- •AI increases output per person; coordination costs make small teams powerful
- •But staying too lean can slow capture before incumbents respond
- •Winner-take-all markets push toward aggressive scaling and speed
- •Small teams were sometimes effective even pre-AI (e.g., Minecraft anecdote)
- •Key is leverage and distribution strategy, not lean-ness as a virtue
- 32:12 – 37:39
What industries transform next: concrete ideas, obsession with speed, and automation in investing
Sarah asks what’s next for AI applications. Andrew avoids vague “AI will transform X” theses, favoring concrete, testable ideas and rapid iteration; he also describes which VC/studio workflows could be automated first (research, reporting), while founder-judgment and trust remain human-heavy due to context gaps.
- •Top-down theses are less useful than specific, actionable wedges
- •AI Fund screens for concreteness: feasibility + customer demand can be tested fast
- •Economist job-disruption studies can inspire where to look
- •Investing ops ripe for automation: competitive research, market research, LP reporting
- •Hard to automate: founder quality assessment, relationship/trust-driven recruiting
- 37:39 – 42:11
Helping first-time technical founders + Andrew’s 5-year AI belief: individual empowerment
Andrew, Sarah, and Elad discuss how to support first-time technical founders through peer groups, complementary hires, and VC/studio pattern recognition (more “reps” in fundraising, speed, customer feedback loops). Andrew closes with a contrarian-leaning view: AI will make individuals far more capable than most expect, across work and personal tasks, for those who embrace it.
- •Founder support: peer groups, complementary hires, learning by doing
- •VC/studios add value via pattern recognition and repeated reps (fundraising, speed)
- •Prioritize building a product users love; fix other issues later when possible
- •Biggest underappreciated impact: AI dramatically amplifies individual capability
- •Those who adopt AI will be disproportionately more productive and empowered
