Data & AI

The Ben & Marc Show: State of AI and Company Building (May-2024)

The AI industry is currently experiencing a transformative period reminiscent of the nascent stages of computing and microprocessors, rather than paralleling the internet boom. The landscape is dominated by large, comprehensive "God models" developed by major technology firms. However, a shift toward an ecosystem of specialized AI models designed for specific applications is anticipated, mirroring the evolution of computing from centralized mainframes to a diverse array of systems across different scales.

Historical patterns in technological revolutions indicate a likely phase of speculative investment, potentially leading to a market correction, before the true value of these innovations becomes clear.

Comparing the AI Boom to the Internet Boom

In essence, while the internet connected existing computing infrastructure, AI is viewed as ushering in a new era of probabilistic, neural network-based computing with different industry dynamics and competitive forces compared to the network-centric internet era.

The speakers contrast the internet and AI in the following key ways:

The internet was a network connecting existing computers, while AI is more akin to a new kind of computer or computing paradigm itself.

The internet industry dynamics were dominated by network effects - companies fighting to get more users on their networks. AI is less about network effects and more about the capabilities of the AI models/systems themselves as information processing engines.

Internet companies were focused on building applications on top of the network infrastructure. With AI, the focus is more on building the foundational models themselves, which are likened to microprocessors or mainframe computers rather than internet applications.

The evolution of the internet started open but became closed over time with large companies controlling discovery and access. With AI, there is a risk of it starting open but being locked down by big tech companies pursuing monopolies.

Lessons from the internet boom/bust cycles around hype, over-investment, and the need for open ecosystems are applicable to AI. However, the internet was a network technology while AI represents a more fundamental shift in computing capabilities.

(GICS 45201020 - Internet Software/Services)

Process Applications vs. Orchestration 'Co-Pilots'

While process apps may simply wrap existing AI capabilities, orchestration apps demonstrate clear value creation by augmenting human operations with multiple AI models in a domain-specific manner, justifying premium, value-based pricing tied to the tangible impact delivered.

Process App:

  • Lacks clear technological differentiation or innovation

  • Acts as a thin "wrapper" around commodity AI capabilities

  • Difficult to charge premium pricing

  • Less ability to quantify and capture share of business value created

Orchestration co-pilot:

  • Leverages multiple AI models as augmented "co-pilots" to increase human productivity

  • Enables measurable efficiency/output gains for customer's operations

  • Can charge based on percentage of quantifiable business value delivered

  • Deeper integration with customer's domain expertise and workflows

The fundamental differences:

Technological differentiation beyond commodity AI models

Quantifiable impact on customer productivity/efficiency

Value-based pricing aligned to measurable business outcomes

Integration with customer's domain knowledge and processes

Technological Differentiation

The core idea is moving beyond just exposing commodity AI capabilities to end users. True differentiation lies in novel AI architectures, integrating multiple models seamlessly, customization to specific domains, deep process integration, and quantifiable productivity gains - going beyond just being "AI wrappers".

Based on the discussion, some key criteria that could define "technological differentiation" for AI applications beyond just being commodity "AI wrappers" include:

Novel AI model architectures or training techniques Examples: Multimodal models that can process different data types, models with new self-supervision or reinforcement learning capabilities, models optimized for specific domains/tasks.

Innovative integration of multiple AI models Examples: Orchestrating language models, vision models, robotics models etc. for complex multi-step applications like video production or travel planning.

Proprietary data and techniques for domain adaptation Examples: Fine-tuning large foundation models on proprietary data to excel at specific industry domains, customized prompting techniques.

Embedding AI deeply into complex workflows/processes Examples: AI co-pilots that augment human productivity in areas like debt collection, legal services, by understanding intricate domain knowledge.

Quantifiable productivity/efficiency gains for customers Examples: Demonstrating measurable output increases, cost savings etc. that can justify value-based pricing models tied to business impact.

Critical Factors Influencing the Direction of AI

The Potential for Self-Improving AI: The ability of AI systems to enhance their capabilities autonomously could accelerate innovation but also raises significant ethical and safety concerns.

The Diminishing Value of Proprietary Data: As the importance of exclusive data repositories decreases, the focus may shift towards more open, collaborative models of innovation.

The Open vs. Closed Models Debate: The industry faces a critical choice between open-source models that encourage widespread innovation and proprietary systems that may limit access and control.

Balancing Innovation with AI Safety and Ethics: Ensuring the advancement of AI technologies aligns with ethical standards and safety considerations is paramount to fostering sustainable growth.

AI Model Capabilities and Trajectory

Debate on whether models will get 100x better through scaling vs. paradigm shifts needed

Potential for self-improvement loops, synthetic data, and better prompting to unlock latent capabilities

(GICS 45103020 - Systems Software)Startup Opportunities and Challenges

Startup Opportunities and Challenges

Narrow AI companies can compete by focusing on specific domains/use cases

Data is less valuable than assumed - large public datasets can match private data

Value-based pricing models key - charging for business value delivered

(GICS 45202010 - IT Consulting & Other Services, GICS 45103010 - Application Software)

AI Industry Evolution

Potential evolution from centralized "God models" to ecosystem of specialized models

Lessons from internet boom/bust cycles applicable - expect hype, over-investment, busts

Open vs closed models will shape competitiveness

(GICS 45103010 - Application Software, GICS 45201020 - Data Processing Services)

Reality of Data Moats

Proprietary data less valuable than assumed due to abundance of public data

Privacy/security concerns around sharing sensitive data with big tech models

(GICS 45103020 - Systems Software, GICS 45202010 - IT Consulting & Other Services)

Suggestions for Application

Implications for Executives

  • Assess if your proprietary data provides a true competitive advantage or if public datasets can suffice for most AI use cases

  • Explore opportunities to leverage AI as a "co-pilot" to increase operational efficiency and productivity across the organization

  • Evaluate pricing models that capture the business value delivered by AI solutions rather than just the technology cost

  • Monitor the evolution of AI from centralized "God models" to specialized models and the potential need for an AI ecosystem strategy

  • Consider data privacy/security implications of sharing proprietary data with big tech AI providers

episode link

PreviousTalent & MasteryNextKnowledge Graphs

Last updated

Was this helpful?