David Luan: Why Nvidia Will Enter the Model Space & Models Will Enter the Chip Space | E1169

Shift in AI Strategy at OpenAI:

• OpenAI (and DeepMind) recognized that the post-Transformer phase would focus on solving major unsolved scientific problems rather than just writing research papers.
• This led to a cultural shift from loose, curiosity-driven research federations to large, focused teams targeting specific real-world problems (e.g., robot hand control, beating humans at video games, scaling GPT).
• The paradigm moved from "building small rockets" to an "Apollo project" style approach with a singular, high-stakes objective.

Model Scaling and Compute Economics:

• Diminishing returns to compute are not imminent; the industry is transitioning from simple base model scaling to a new wave of improving model intelligence.
• Historical Scaling: For base language models, doubling the compute consistently and predictably improves intelligence, following a logarithmic curve rather than a straight line.
• New Scaling Method: A second, emerging method involves models collecting their own data via simulated environments (e.g., theorem provers, Jupyter notebooks) to learn through trial, error, and reflection (RL-like loops).
• This "simulation/synthetic data" path absorbs significant compute and is the new critical path as base model scaling becomes prohibitively expensive (billion-dollar training runs).

Reasoning and Capabilities:

• Pure unsupervised learning on internet data (LLM scaling) cannot inherently discover new knowledge or solve novel problems; it can only mimic human behavior.
• Solving reasoning requires giving models access to interactive environments (like theorem provers) and human feedback loops to compose existing thoughts into new ones.
• Minimum Viable Capabilities: Specific capabilities (e.g., three-digit arithmetic) emerge unpredictably at specific model scales; engineers act more like "gardeners" than traditional programmers.
• Memory: Short-term working memory (context length) is improving (e.g., millions of tokens), but long-term memory for user preferences is the responsibility of application developers, not the model provider.

Market Dynamics and Consolidation:

• Provider Consolidation: The speaker predicts a steady state of only 5–7 major LLM providers due to the immense capital requirements and the necessity of vertical integration.
• Vertical Integration:
  • Tier-one cloud providers must win to avoid obsolescence, as models will become the base computing primitive.
  • Chip makers (like Nvidia) face pressure to move up the stack into models, while model providers (like Google with TPUs) move down the stack to own chips for cost advantages.
  • Apple holds a unique advantage in running private, edge-based models but will likely rely on third-party "big brains" (like OpenAI) for complex reasoning tasks, treating the model as a "swappable" component.
• Independent Model Sellers: Pure-play model sellers to developers face a short window to build economic flywheels before commoditization forces them to become first-party efforts of a major cloud.

Adapt (The Speaker's Company) & Agent Architecture:

• Business Model: Adapt is building a vertically integrated enterprise agent stack, not just selling foundation models.
• Product Strategy: The goal is to create a "system of record for workflows" where any employee can teach the AI arbitrary tasks by showing examples, handling the extreme variability of enterprise edge cases.
• Differentiation from RPA:
  • RPA: Follows rigid, pre-defined "yellow lines" (high-volume, identical tasks); akin to factory robots.
  • Agents: Constantly think, re-evaluate, and plan to achieve high-level goals; akin to full self-driving.
• Disruption: Agents disrupt RPA business models by replacing the need for months-long process engineering with natural language instruction and observation.

Pricing and Organizational Structure:

• Pricing: The speaker challenges the "price per work" model for knowledge work, arguing that AI agents should function as co-pilots/teammates augmenting human creativity rather than replacing jobs on a transactional basis.
• Talent Stack Collapse: AI will enable humans to become generalists, supervising cohorts of AI specialists, effectively collapsing the traditional multi-person talent stack (e.g., one person acting as PM, designer, and engineer).

Enterprise Adoption and Hype Cycles:

• Adoption Curve: Enterprise AI adoption is currently in the "experimental budget" phase; widespread core budget integration will take a long time due to legacy systems (mainframes) and workflow inertia.
• Comparison to Autonomous Driving: Unlike self-driving cars which faced a "99.9%" reliability plateau, AI progress is characterized by visible, periodic scientific breakthroughs that continuously expand capabilities.
• Economic Winners: While AI services/consulting firms will lead in revenue initially, the long-term economic winners will be companies that productize successful use cases into repeatable software, not just the services providers or the model makers.

Risks and Safety:

• Regulatory Capture: The speaker fears regulations will be shaped by incumbent players to raise barriers to entry, concentrating power among the few large model providers.
• Open vs. Closed: Open source will likely lag closed models due to resource constraints, but remains vital for the broader field to keep up with incumbents.
• Safety Concerns: Misuse is already occurring (e.g., automated vulnerability scanning); AGI safety is difficult to reason about as an "infinity" problem.
• Human-Computer Interaction (HCI): The final frontier for AGI is defining the right interface for human-AI collaboration, moving beyond chat to richer, shared-context interactions.

Forward-Looking Statements:

• Agent Evolution (5 Years): Agents will function like non-invasive brain-computer interfaces, allowing humans to operate at a higher level of abstraction and reason beyond current cognitive limits.
• Threat to Agent Vision: The most probable reason this vision fails is "walled gardens" created by incumbents, preventing agents from bridging domains and spanning the full spectrum of work.
• Market Size: The addressable market for agents is estimated to be 1,000x to 10,000x larger than the current RPA market.
• Speciation: Chatbots and agents will speciate into two distinct product lines: creative/therapeutic chat interfaces vs. reliable, goal-oriented work agents.