Two years into the Generative AI revolution, the field is evolving from rapid pre-trained responses (“thinking fast”) to advanced reasoning at inference time (“thinking slow”), enabling a new class of intelligent applications. As we reflect on the second anniversary of our essay “Generative AI: A Creative New World, ” we observe significant changes within the AI ecosystem and present our forecasts for the future. The Generative AI market's foundational layer is stabilizing, characterized by major players like Microsoft/OpenAI, AWS/Anthropic, Meta, and Google/DeepMind, all equipped with substantial capital and scalable operations. While competitive dynamics remain intense, we anticipate a greater availability of cost-effective next-token predictions. With this stabilization, attention is shifting towards enhancing reasoning capabilities in AI, dubbed "System 2" thinking, which allows for more deliberate problem-solving and cognitive functions during inference. Inspired by innovations like AlphaGo, this reasoning layer is poised to redefine how AI operates, leading to new cognitive architectures and user interfaces. In our latest analysis of the AI landscape, we will address how this consolidation fuels the race to develop sophisticated reasoning capabilities and explore emerging "killer apps" with innovative cognitive tools and interactions. **The Strawberry Revolution** The key development of 2024 is OpenAI's introduction of o1, formerly known as Q* or "Strawberry. " This model not only reinforces OpenAI's leadership in model quality but also integrates genuine reasoning capabilities through inference-time compute. Unlike traditional pre-trained models that rely on vast data for basic reasoning, Strawberry enhances the model’s ability to pause and deliberate before responding. Using the example of AlphaGo’s 2016 victory over Go champion Lee Sedol, we illustrate how true AI reasoning can surpass simple pattern-matching. AlphaGo demonstrated advanced cognitive processes by actively simulating potential outcomes, allowing it to outperform human players significantly when given time to evaluate. Translating this back to LLMs, the challenge lies in creating effective scoring mechanisms for diverse tasks, such as evaluating written content, which remains problematic for existing models. Strawberry showcases strengths in logical domains but struggles with more subjective tasks due to its foundational mechanics. Deep reinforcement learning is experiencing a resurgence as it enables AI to reason more effectively, showcasing behaviors akin to human thinking, such as backtracking and innovative problem-solving strategies. The field is currently exploring various methods to refine inference-time processes and enhance reasoning capabilities. **Shifting from System 1 to System 2 Thinking** Progressing from instinctive responses (System 1) to informed reasoning (System 2) marks a crucial frontier for AI. While electrical pre-training can resolve straightforward queries rapidly, more complex issues require AI to evaluate options and reason through decisions methodically.

This deeper process is essential for tackling significant problems across disciplines, from mathematics to biology, where mere pattern recognition is insufficient. In the realm of AI applications, newly emerging models exemplify a transformative approach. Companies like Sierra are redefining customer service by resolving issues efficiently, while the advent of other agentic applications—such as Harvey (law), Glean (work assistant), and XBOW (penetration testing)—illustrates a shift in how AI delivers value and captures market opportunities. As these applications minimize delivery costs by leveraging advancements in inference, they can disrupt traditional market models, enhancing workflow efficiency and accessibility. **The SaaS Landscape Reconsidered** Concerns have arisen about whether Generative AI could threaten existing cloud companies. However, our initial stance is that foundational models will remain accessible to incumbents, as they possess existing advantages in data and distribution. Startups are better positioned to target automatable work streams rather than outright replacing established software companies. Yet, the substantial engineering and innovation needed to create compelling AI-driven solutions may suggest that we have underestimated the requirement for being "AI-native. " Just as the shift from on-premise software to SaaS transformed business models two decades ago, a similar evolution could redefine the landscape for AI applications. Day. ai embodies this potential, presenting an AI-native CRM that automatically generates tailored solutions with minimal human involvement. This exemplifies a growing trend towards automating complex processes and reshaping market expectations. **Investment Insights** From an investment standpoint, infrastructure centered on hyperscalers is less attractive for venture capitalists, dominated by major players focusing on game-theoretical strategies. The model landscape suffers from bias towards innovative models, often disregarding practical economic considerations. Conversely, developer tools and application layers attract more significant interest, with a potential resurgence of revenue generation similar to what was seen during the transition to the cloud. Investment in application-layer companies remains promising, where a wealth of opportunities can yield substantial returns and sustainability. The shift towards agentic applications is set to redefine the dynamics of the AI ecosystem, paving the way for impactful growth and innovation.