By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

The Twenty Minute VC

Sam Altman: What Startups Will be Steamrolled by OpenAI & Where is Opportunity | E1223

Sam Altman is the CEO of OpenAI, one of the most important companies in history. OpenAI is on a mission to ensure that artificial general intelligence benefits all of humanity.

Sam AltmanguestHarry Stebbingshost

Nov 4, 202439mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Why “patching model shortcomings” is a bad startup strategy
Sam opens with a warning to founders: businesses built around fixing today’s model limitations risk becoming irrelevant as base models improve. The better strategy is to build products that become more valuable as the models get more capable.
- •OpenAI expects steep, rapid model improvement over time
- •Startups that “patch” current weaknesses may get obsoleted
- •Prefer ideas that benefit directly from stronger future models
- •Founders should align roadmaps with the direction of model progress
OpenAI’s roadmap: prioritizing reasoning models (the O-series)
Harry asks whether OpenAI’s future is smaller specialized models or ever-larger general ones. Sam emphasizes that reasoning-focused models are strategically central and should improve quickly, unlocking new capabilities in science and software.
- •Reasoning is a key strategic focus for OpenAI
- •Expectation of rapid improvement in the O-series models
- •Reasoning models could enable breakthroughs in science
- •Reasoning models should handle much more difficult coding tasks
No-code and the path to AI-built startups
Sam outlines a progression from tools that make developers more productive toward eventually enabling no-code creation. He notes that fully describing and generating an entire startup via no-code is still some way off.
- •Near-term: amplify productivity for people who can already code
- •Longer-term: high-quality no-code tools are likely
- •Today’s no-code tools are early and limited in scope
- •Full “describe a startup and build it” remains a longer-horizon goal
Where OpenAI may “steamroll”: building for model momentum, not against it
The conversation turns to how founders and investors should think about defensibility as OpenAI moves up the stack. Sam distinguishes between products that fight the model’s current constraints and those that ride the wave of improving capability to create novel services.
- •Trillions in new market cap may be created by AI-enabled products
- •OpenAI aims to reduce the need for heavy workaround engineering
- •Earlier wave of startups often bet against model improvement (e.g., GPT-3.5 era)
- •Best opportunities: products that get better as models get better (tutors, medical advisors, etc.)
Masa Son’s $9T value prediction and the economics of AI abundance
Harry asks about the plausibility of enormous AI-driven value creation versus huge capex. Sam says precision isn’t the point yet—orders of magnitude matter—and argues that agents capable of producing whole-company software could unlock massive value, especially in healthcare and education.
- •Large capex spend is expected in major tech revolutions
- •Value creation could be ‘unbelievable’ even if exact numbers vary
- •If anyone can describe and generate major software, costs fall dramatically
- •Healthcare and education highlighted as multi-trillion-dollar opportunities
Open source vs. integrated services: how the ecosystem shakes out
Sam positions open-source models as an important part of the ecosystem, alongside well-integrated APIs and services. He suggests users will choose what fits their needs rather than one approach winning universally.
- •Open source has a durable role in AI
- •Quality open-source models already exist
- •APIs/services matter for integration and reliability
- •Market will support multiple delivery modes
What an AI agent really is—and why the ‘restaurant booking’ example misses the point
Sam gives a pragmatic definition of an agent: something you can assign a long-duration task with minimal supervision. He argues the real unlock is not small convenience tasks but massively parallel work and ‘senior coworker’ style collaboration over days or weeks.
- •Agent definition: long tasks + minimal supervision
- •Common examples (booking reservations) undersell the potential
- •Agents can do massively parallel work humans cannot (e.g., call 300 places)
- •Most compelling: an agent as a smart coworker producing real work products
How AI agents could change SaaS pricing (compute as the unit of work)
Harry probes how labor-replacing agents might reshape per-seat SaaS pricing. Sam speculates pricing may shift toward allocating continuous compute (e.g., a fixed number of GPUs) rather than seats or discrete agents.
- •Uncertainty: pricing models are still unclear
- •Possible shift from per-seat to compute-based pricing
- •Continuous ‘compute allocation’ could map to ongoing problem-solving
- •Agentic workloads may demand new commercial abstractions
Are foundation models commodities? Depreciation, defensibility, and scale advantages
Sam agrees models depreciate, but pushes back on the idea they aren’t worth their training cost, noting compounding know-how and strong revenue potential at scale. He also points out that many groups training similar models may struggle without sticky products and distribution.
- •Models depreciate, but can still be highly valuable
- •Training expertise compounds—each generation improves the next
- •Too many teams may be training similar ‘me-too’ models
- •Distribution and product stickiness (e.g., ChatGPT) help amortize costs
Differentiation through reasoning and multimodality—plus rapid progress in vision
Sam explains OpenAI’s differentiation focus: reasoning first, alongside multimodal improvements and features aligned to user needs. On multimodal reasoning, he frames it as clearly possible (humans do it pre-language) and hints at fast progress in image models.
- •Reasoning is the most important near-term differentiator
- •Multimodal capabilities are a major parallel investment
- •Multimodal reasoning is plausible given human developmental evidence
- •Expectation of rapid progress in image-based models
How OpenAI breaks through: culture for ‘new and unproven,’ not just replication
Pressed on how core reasoning advances happen, Sam declines to share technical details but discusses what’s genuinely hard: consistently doing new, unproven work. He contrasts copying known successes with building a culture that repeatedly produces breakthroughs, and ties this to unlocking human potential.
- •Replication is ‘doable’ once feasibility is proven
- •Hard part: repeated innovation into unproven territory
- •OpenAI culture prioritizes doing new things, not just copying
- •Wasted human talent comes from environments that don’t enable peak output
Leadership under hypergrowth: learning to steer from 10% growth to 10× leaps
Sam reflects on how OpenAI’s explosive growth compressed learning timelines and forced rapid organizational scaling. He highlights the difficulty of keeping teams focused on step-change ‘10×’ moves while maintaining operational excellence, planning, and internal communication.
- •Hypergrowth compresses learning and organizational formation
- •Leading 10× change is fundamentally different from optimizing 10% growth
- •Requires strong internal communication and planning discipline
- •Long-lead constraints include compute buildout and even office space logistics
Hiring philosophy: don’t optimize for age—optimize for talent and stakes
On advice to hire primarily under-30 talent, Sam argues both early-career and experienced hires can be critical depending on the role and risk. He emphasizes maintaining a very high bar at any age, while also valuing society’s need to bet on high-potential inexperienced people.
- •Exceptional young talent can deliver ‘off-the-charts’ work
- •High-stakes infrastructure roles often require deep experience
- •Rigid ‘only young’ or ‘only old’ strategies are misguided
- •Key principle: high talent bar + context-appropriate experience
Choosing models, scaling laws, and the complexity of building frontier AI (plus quick-fire)
Sam acknowledges Anthropic’s strength in coding and expects developers to use multiple models, shifting from ‘models’ to ‘systems.’ He then affirms scaling progress will continue for a long time despite doubts from failed runs and unknown behaviors, discusses decision-making and supply chain worries, disputes $100B entry costs, and closes with quick-fire takes on what to build and what’s next.
- •Developers may increasingly choose ‘systems’ over single models
- •Scaling trajectory likely continues for a long time, despite setbacks
- •Hard problems include failed training runs and paradigm shifts (GPT-4 to o1)
- •Top worry is ecosystem-wide complexity (chips, power, networking, research, product, monetization)
- •Belief that entry cost for foundation models is likely < $100B; AI analogies (internet/electricity) can mislead; transistor analogy fits better
- •Quick-fire: build AI verticals (e.g., tutoring); desire for AI that understands your whole life; controllable latency/accuracy dial; product strategy is a personal weakness; 5-year view: wild tech progress, society changes less than expected

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.