Dialogue with Sam Altman: OpenAI's Strategic Choices, Computing Bets, and Underlying Thoughts on Commercialization

1. Future Trends and the Rhythm of AI Implementation: Why is Social Evolution Slower than Technology? 1. Industry Pioneer Case: Shopify CEO Tobi Lütke's Extreme Sensitivity • Deeply Involved in Code and Products: Tobi Lütke shows strong sensitivity in the AI field, always ahead of the industry by 6-8 months; as a large enterprise CEO, he personally writes code, restructures workflows, and provides extremely precise product details to OpenAI. • Actively Reshaping Company Form: He insists that "Shopify will never be a passive recipient (NPC Company)" and advocates for actively embracing agents, even attempting to rewrite Shopify with a new AI-native architecture at night. 2. Technological Disruption vs Economic and Social Inertia • Actual Delay in Disruption Cycle: Sam Altman once believed that the software industry would rapidly reshape after the launch of GPT-4, but reality shows that the speed of social evolution is much slower than pure technological development. • Barriers of Habit and Switching Costs: History repeatedly proves (e.g., Larry Ellison's discovery that installing software is easy, but changing user habits is difficult; Netflix had already mailed DVDs, but the public still preferred Blockbuster), the economic system has significant inertia, and the public tends to stick with familiar ways of working and suppliers. • Anti-Cyclical Nature of Non-AI Native Experiences: The more high-tech becomes prevalent, the more fields with real interpersonal connections, physical experiences, or emotional recognition (such as offline experiences, sports events) possess a solid moat. 3. Psychological Resistance of Habits and the Absence of "iPhone-Level Interaction" • Contradiction in Founders' Own Behavioral Inertia: Sam Altman admits to having 20 years of traditional computer operation habits (mechanically checking emails, copying and pasting, listing to-do items), and even with powerful Codex agent tools, he often finds it difficult to fully switch to a pure AI workflow due to psychological coding habits. • Currently in the "Palm PC Era Before the iPhone's Birth": Similar to the Palm Treo or Sidekick in 2003/2004, the underlying technology modules are ready, but a truly revolutionary super product that fundamentally changes user interaction interfaces has yet to emerge. 2. OpenAI's Strategic Positioning: Extreme Focus and Abandoning "Good Ideas" 1. Core Strategy: Transitioning from a "Product Company" to "Platform Infrastructure" • Integrating Core Entry Points: Deeply merging ChatGPT and Codex to create a unified personal and enterprise AGI interaction entry point, supported by powerful underlying APIs. • Covering the Full Cost-Performance Curve: • High-End Scenarios: Providing cutting-edge super intelligence for frontier scientific discoveries and complex reasoning. • Low-End Scenarios: Offering cheap, efficient, high-throughput computing power to support massive daily tasks. • Not Competing with Ecological Customers: OpenAI's goal is to become the underlying platform supporting 100 million startups and 8 billion users, rather than extending its reach into all vertical application tracks. 2. Abandoning Good Ideas, Fully Betting on Ultimate Goals (Focus & Trade-offs) • Cutting Sora and Atlas Browsers: • Sora Video Generation: Although it has innovation and entertainment value, it consumes an enormous amount of computing resources, and after weighing against the core strategy, it was decided to make way for key intelligent reasoning. • Atlas Web Browser: The product experience was excellent, but it was decisively terminated to avoid distracting top R&D talent. • First Principles Focus: Given the limited reality of computing power, talent, and resources, OpenAI will focus all its efforts on "general intelligence leading to knowledge work and scientific discovery," covering self-developed chips, infrastructure software, self-built data centers, and model pre-training. 3. Personal Energy Allocation and Computing Infrastructure Challenges • Focusing on Research and Compute: Products are built by excellent teams, while Sam Altman invests most of his energy in model research and building the computing supply chain. • The Most Expensive Infrastructure Project in Human History: The expansion of computing power crosses complex geopolitical policies, chip design, wafer foundry capacity, rack manufacturing, power and energy system scheduling, and large-scale financing structures. 3. From Investor to Research Operator: Non-Consensus Betting and Research Management 1. The Underlying Commonality of Venture Capital and Cutting-Edge Research • Dominated by the Power Law: In AI research and investment, a few non-consensus breakthroughs (such as early bets on LLM and AGI) create value that can completely overshadow all conventional projects combined. • Identifying Non-Standard Extreme Talent: Rejecting mediocre entrepreneurs/researchers who follow popular concepts, focusing on selecting "outliers" with strong independent thinking, the courage to adhere to obscure non-consensus hypotheses, and extreme conviction. 2. Breaking Conventional Startup Paths: The Darkest Moment of Not Releasing Commercial Products for 4.5 Years • Research Exploration Against YC Conventional Rules: From its establishment at the end of 2015 to around 2020 when the first commercial product was launched, the team had no real user feedback signals for 4.5 years. • Building Internal Signal Replacement Mechanisms: • Using Dota 2 Reinforcement Learning Ranking Leaderboard to build objective measurement standards. • Introducing high-standard external expert demo presentations to drive R&D breakthroughs. • The Confusion of the 2016 Apartment Cold Start: At its inception, 12 people were in Greg Brockman's apartment without even a whiteboard, gradually establishing the research rhythm from unsupervised sentiment analysis, GPT-1 to Scaling Laws. 3. The Cognitive Weight of Success and Failure Experiences • The Value of Learning from Success Far Exceeds That from Failure: • The reasons for failure are varied, and often only generalized conclusions about "perseverance" can be extracted from them. • Deeply understanding the core elements of success (such as the key grips that YC and OpenAI got right) and amplifying them through compounding is key to driving business leaps. 4. Think Tanks and Cognitive External Brains: The Influence of Peter Thiel and Paul Graham 1. Extremely Non-Linear Thinking Inspiration • When facing extremely tricky strategic bottlenecks, the main think tanks that can provide counterintuitive perspectives are Paul Graham and Peter Thiel, whose non-linear thinking can directly break through mental deadlocks. 2. Peter Thiel: Doubling Down on the "Blank Input Box" • Key Guidance During ChatGPT's Initial Confusion: Within two months of ChatGPT's launch, although there was growth, due to the lack of the information flow (Feeds), network effects, and user lock-in mechanisms that Silicon Valley valued at the time, the team considered shifting to 5-6 other directions. • The Power of Minimal Essence: Peter Thiel pointed out that this was the most powerful "blank search box" form since Google, capable of inputting everything and outputting correct results, and it was essential to double down on this core interface, which completely ended internal wavering. 3. Paul Graham: Rapid Iteration and Early Release (Iterative Deployment) • "You should release when the product makes you feel embarrassed": Pushing the initial version to market early to face real-world feedback and iterate quickly is the core gene that determines startup success rates. 5. AI Safety, Social Governance, and Human-Centricism 1. Agile Iteration is the Optimal Path to Achieve Safety (Iterative Safety) • Stepping Out of the Ivory Tower: True AI safety cannot be achieved through pure theoretical deduction in closed laboratories; models must be pushed to hundreds of millions of real users to discover hallucinations, alignment failures, and vulnerabilities in real edge scenarios, and establish a transparent review and improvement system similar to civil aviation accident investigations (FAA). • Co-evolution of Models and Society: Society needs time to adapt to technology, and technology also needs to establish robust boundaries through real interactions. 2. Beware of "Power Rent-Seeking and Centralization" Under the Guise of Safety • Firmly Opposing Anti-Human Governance Views: • Be wary of arguments that deprive the public of technology usage rights in the name of protecting humanity, concentrating superintelligence control in the hands of a few oligarchs or a single AI decision-making body. • Firmly oppose the authoritarian technological concept of "exchanging freedom and decision-making power for the elimination of diseases and cheap materials." • Empowering the Public with More Autonomy and Leverage: The ultimate significance of AI is to grant ordinary people greater creativity and freedom of action, and the future will witness the largest wave of small and micro enterprises and individual entrepreneurship in human history. 3. The Next Stage of AI: Reconstruction of Decision-Making Through Ultra-Long Context • From Model IQ to Individual Context Empowerment: Current models have significantly improved intelligence, and the next breakthrough point lies in AI's ability to digest and refine ultra-massive context (internal documents, communication records, vast papers) in seconds, becoming an indispensable high-dimensional think tank for humans when making significant decisions. • The Fundamental Connection of Humanity is Irreplaceable: No matter how advanced superintelligence develops, the human desire for real interpersonal connections, emotional resonance, and physical interactions remains a core foundation.