The Twenty Minute VCEthan Mollick: Why OpenAl Abandons Products, The Biggest Opportunities They Have Not Taken | E1184
CHAPTERS
- 0:00 – 0:34
OpenAI’s product gap vs the AGI race mindset
Ethan argues OpenAI repeatedly abandons promising products because the organization is optimized for scaling toward AGI rather than refining end-user tools. He frames today’s revenue as almost incidental and highlights a broader contradiction: much of the ecosystem behaves as if steady scaling will solve everything.
- •OpenAI’s tendency to drop or deprioritize product lines
- •Chatbot + API as the main durable offerings so far
- •AGI ambition as the organizing principle (“machine god”)
- •Ecosystem assumption that scale/development will fix issues
- •Startups implicitly betting against near-term AGI
- 0:34 – 2:32
Ethan’s background: entrepreneurship, education, and translating AI
Ethan explains how his career spans entrepreneurship, teaching at Wharton, and early exposure to AI at MIT’s Media Lab. He describes becoming a public interpreter of AI’s implications and a conduit between labs and the broader world.
- •Co-founded a startup credited with inventing the paywall
- •Wharton entrepreneurship professor; MIT training
- •Worked with Marvin Minsky at the Media Lab
- •Long-running focus on teaching/entrepreneurship at scale
- •Social media + lab access created a reinforcing “go-to” role
- 2:32 – 4:35
LLaMA 3.1: open-weights parity and what it unlocks
The conversation turns to Meta’s LLaMA 3.1 and the significance of an open-weights model approaching GPT‑4-level capability. Ethan emphasizes diffusion: once models are downloadable and fine-tunable, delayed second-order effects begin to show up globally.
- •Open weights as a major distribution and innovation catalyst
- •Competitive parity is real, but closed labs likely have more to ship
- •Open availability changes national and enterprise access dynamics
- •Model doesn’t uniquely ‘feel’ better than top closed models—yet
- •Expect a wave of downstream experimentation and weird effects
- 4:35 – 5:53
The ‘transient winner’ problem and what actually matters
Harry notes the weekly social-media narrative of shifting model leadership. Ethan argues this matters less than whether and when capabilities top out, and cautions against over-indexing on leaderboard churn.
- •Social media amplifies short-lived dominance narratives
- •Mainstream users stick with familiar tools despite model jockeying
- •Key question: when capabilities plateau and what the ceiling is
- •No obvious ‘secret breakthrough’ disclosed in LLaMA techniques
- •Enthusiasts can track benchmarks; most should focus on outcomes
- 5:53 – 8:25
A four-outcomes framework for AI’s future (from fizzle to ‘machine god’)
Ethan lays out four broad trajectories: stagnation/fizzle, superintelligence, and two middle paths characterized by slower or steadier improvement. He argues the middle scenarios—especially linear, incremental gains—are under-discussed but most operationally important for society and firms.
- •Outcome 1: models stall and integration is slow/limited
- •Outcome 4: AGI/superintelligence and unpredictable societal shift
- •Middle scenarios: continued improvement without sudden takeoff
- •Even current systems aren’t deeply integrated into work yet
- •Linear gains (e.g., percentile improvements) may be most realistic
- 8:25 – 9:56
Escape velocity vs iPhone-like plateau: jagged intelligence and unknown ceilings
Using iPhone progress as an analogy, Harry asks why AI won’t plateau. Ethan explains that sustained exponentials can happen when underlying tech swaps out, but it’s unclear what AI’s true capability ceiling is and whether ‘jaggedness’ can be smoothed.
- •Tech can sustain exponentials via successive underlying breakthroughs
- •Core uncertainty: what ‘max intelligence’ looks like in practice
- •AI is ‘jagged’: excellent at some tasks, poor at others
- •Jaggedness limits full substitution for human work today
- •Open question: can the jagged edges be systematically overcome?
- 9:56 – 12:10
Compute, data, algorithms—and the overlooked bottleneck: human systems
Asked to pick the main bottleneck, Ethan reframes the debate: most users don’t care about architectural details; they care about usable capability. He introduces ‘reverse salient’ theory—progress is constrained by the lagging subsystem—and predicts bottlenecks will rotate as money and talent swarm them.
- •Most users don’t care whether it’s LLMs, MoE, Mamba, etc.
- •Human/organizational systems are a major limiting factor
- •Reverse salient: innovation focuses on the current lagging constraint
- •Likely sequence of shifting bottlenecks (data pipelines, etc.)
- •Capital and prestige concentrate where constraints block progress
- 12:10 – 13:53
Bad analogies in AI: why ‘steam power’ beats ‘picks and shovels’
Ethan critiques the common ‘picks and shovels’ framing and argues AI adoption is more like steam power in factories: the value accrues to the people who can connect a general-purpose capability to specific workflows. Skilled ‘artisans’ inside organizations are key to translating model power into real productivity.
- •‘Picks and shovels’ is vague and misleads GTM thinking
- •Steam engine value came from adaptation to concrete processes
- •Patents and openness accelerated diffusion historically
- •Modern analogue: convert LLM ‘back-and-forth power’ into workflows
- •Winning means enabling internal builders and process redesign
- 13:53 – 15:28
Why providers don’t ship ‘AI for Dummies’: obsession with scaling and rumor-based learning
Harry challenges the lack of practical onboarding and guides. Ethan argues labs prioritize scaling above all else because they believe bigger models will solve downstream issues; this leaves users with ‘documentation by rumor’ and a chatbot-first UX that intimidates many.
- •Labs optimize for scaling toward AGI; manuals feel like a distraction
- •Product work risks rapid obsolescence as models leapfrog
- •Organizations built bespoke tools that ChatGPT quickly obsoleted
- •No clear guidance on strengths/weaknesses; learning happens via Twitter
- •Result: widespread underuse and shallow use patterns
- 15:28 – 18:53
Open vs closed models: benefits, real risks, and the need for fast-follow governance
Ethan supports openness for entrepreneurship and global access in areas like health and education, while warning of immediate misuse once guardrails are removed. He advocates a ‘fast reaction’ model for regulation and monitoring—observe harms quickly and respond—rather than rigid pre-regulation or laissez-faire extremes.
- •Open models expand access and innovation; aid under-served regions
- •Risks: scalable spearphishing, catfishing, and security externalities
- •Debate is often corporate strategy rather than societal design
- •Fast-follow regulation (Joshua Gans): monitor, learn, react quickly
- •Need feedback loops/monitoring to understand real-world impacts
- 18:53 – 22:11
Regulation, Europe’s ecosystem, and why VC remains intensely local
The EU AI Act prompts discussion about whether heavy regulation slows development and adoption. Ethan argues ‘either/or’ regulation narratives miss the tradeoffs and points out Europe’s broader innovation constraints; he also explains research showing VC investing is geographically local due to networking and monitoring needs.
- •Pre-emptive, inflexible regulation can hinder experimentation
- •But ‘no regulation’ ignores real downside risks and societal goals
- •Europe’s challenges are multi-causal beyond regulation alone
- •Empirical results: relocation to Silicon Valley improves startup outcomes
- •VC-company distance and direct flights materially affect investment
- 22:11 – 24:40
What AI labs miss about business: abandoned tools, thin use-case thinking, and end-user discovery
Ethan argues major labs don’t understand large-company realities and release half-finished products without a deep use-case roadmap. He cites Code Interpreter as a transformative tool that was not fully productized, and claims many of the best use cases are discovered by managers and end users—not technologists.
- •Labs lack large-company experience and empathy for enterprise workflows
- •Products can be brilliant but left partially unsupported
- •Code Interpreter as a ‘world-changing’ capability for analysts
- •Documentation skews technical; end-user onboarding is weak
- •Use-case innovation is emerging bottom-up from non-technical users
- 24:40 – 28:23
Why companies under-adopt AI: ‘secret cyborgs,’ weak policies, and perverse incentives
Ethan reports that only a small minority have meaningful hands-on time with modern models, largely due to poor onboarding and unclear rules. Inside firms, many high performers use AI secretly because they fear punishment, reputational harm, or being rewarded with layoffs or extra work; policy design becomes the lever for unlocking productivity gains.
- •Low real adoption: many tried old models; few invested 10+ hours
- •Blank-page anxiety and lack of onboarding suppress experimentation
- •Vague or restrictive policies (and unclear regulation) block GPT‑4 access
- •Employees hide AI use due to fear of firing, stigma, or workload increases
- •Productivity gains require incentives: expand output vs cut headcount
- 28:23 – 33:23
Jobs, inequality, and adoption: disruption is real, but AI access is unusually broad
Harry worries AI will eliminate lower-wage work first (e.g., customer service) and widen inequality. Ethan agrees disruption will be painful and not automatically offset, citing historical examples; yet he notes AI’s interface and distribution (phone + chat/voice) can spread broadly, and tech insiders may not retain the usual advantage because effective use depends on ‘human’ skills like instruction-writing and theory of mind.
- •Displacement will happen; not every job gets a clean replacement category
- •Industrial transitions create unrest; retraining has historically been weak
- •Early adoption skews wealthy/male, but access is globally widespread
- •Coders aren’t necessarily best users; AI is non-deterministic and ‘objects’
- •People skilled with humans/instructions may outperform traditional tech elites
- 33:23 – 36:10
From chatbots to multimodal agents: the next consumer interface
Ethan predicts the interface will shift from typing into chat to multimodal, conversational, always-available assistants with increasing agency. Prompting mastery may matter only in a transitional window; once voice, vision, and actions integrate into phones and workflows, adoption can accelerate as people chase time savings.
- •Multimodal (voice + vision) makes AI feel like ‘a human on call’
- •Agency (taking actions) could bypass today’s prompting complexity
- •Prompting advantage may shrink as UX improves and models scaffold users
- •Humans optimize for effort reduction—driving eventual mass adoption
- •Universities adopt faster due to social diffusion and homework fit
- 36:10 – 46:48
Startups and VCs in a radical regime: stop building ‘minor layers’ and take a future stance
Ethan argues the lean, incremental PMF playbook is poorly suited to a fast-moving, general-purpose technology wave. He calls for founders and investors to hold an explicit view of the future (how good models get, where ‘jaggedness’ remains) and to design adoption strategies around organizational reality—while highlighting the contradiction of funding narrow apps while believing AGI is near.
- •Lean/PMF methods bias toward incremental, not breakthrough innovation
- •Deep-tech style bets and imagination become more important
- •Investors should demand a coherent ‘future position’ on model progress
- •Adoption requires plans for organizational diffusion, not hope
- •Contradiction: ‘AGI in 5 years’ vs funding fragile point-solution startups
- 46:48 – 57:33
Education as the case study: AI tutoring, flipped classrooms, and why teachers remain essential
Ethan uses education to show why subject-matter and system knowledge matter: AI tutoring can be powerful, but naive deployment can worsen learning (AI doing the work). He outlines a practical future in which AI tutors support learning outside class, while class time focuses on active learning and application, with teachers orchestrating motivation, systems, and assessment.
- •AI tutoring is promising but can backfire without scaffolding
- •RCT example: better homework scores but worse tests when AI ‘does it’
- •Future model: AI-supported flipped classrooms + active learning in class
- •Teaching quality depends on known learning science (practice, testing, reflection)
- •Schools are embedded in societal systems (credentialing, childcare, unions)
- 57:33 – 1:00:00
Compute and energy: ‘currency of the future’ or another reverse salient?
Harry raises the view that compute is future currency and energy demand is a looming constraint. Ethan says this framing only fully holds under strong AGI assumptions with near-infinite demand; otherwise, energy is a manageable near-term issue that becomes a solvable ‘reverse salient’ attracting investment (e.g., nuclear buildout, efficiency gains).
- •If AGI + infinite demand: compute/energy become central constraints
- •Current per-query energy > Google search, but < human labor energy costs
- •Data centers are a modest share of total power today; runway remains
- •Constraints trigger investment and innovation to relieve bottlenecks
- •Energy could become a major profit pool if it gates intelligence supply
- 1:00:00 – 1:09:06
Politics, persuasion, and the closing quick-fire: meaning, agency, and why people bounce off AI
Ethan rejects simplistic dystopias like ‘voting for algorithms’ arriving quickly, emphasizing slow-changing human systems. He warns instead about near-term persuasion, misinformation dynamics, and longer-term risks of losing agency to AI-mediated systems; in quick-fire, he stresses the open question of why users disengage and the looming ‘meaning of work’ crisis as AI silently substitutes for parts of many roles.
- •Human systems resist overtly dystopian shifts; politics changes slowly
- •AI is already highly persuasive; marketing and politics will be affected
- •Deepfakes matter less than low-attention sharing and headline dynamics
- •Key risk: losing human agency to AI-infused systems and incentives
- •Under-asked topics: why people stop using AI; meaning-of-work crisis
