Artificial Intelligence has reached an inflection point, transforming how businesses operate and how individuals interact with technology. In a comprehensive lecture at the Stanford Graduate School of Business, Mihail Eric, Head of AI at Monaco, shared his insights into the history, mechanics, applications, and future of AI. Characterizing the current era as “the age of the builder,” Eric provided a deep dive into the technology stacks and business models defining the AI ecosystem today.

The Evolution of AI: A Brief History

The modern AI boom did not happen overnight; it is the result of over a decade of compounding breakthroughs.

• 2012–2016: The Deep Learning Renaissance. While neural networks have existed since the 1950s, they experienced a massive resurgence around 2012. This was driven by a perfect convergence of new machine learning methods, the repurposing of consumer gaming GPUs for computation, and the creation of massive training datasets like ImageNet. This era yielded massive 10% to 15% performance improvements in tasks like image classification, speech recognition, and machine translation.
• 2017: The Transformer Architecture. A seminal paper by Google researchers titled “Attention Is All You Need” introduced the Transformer architecture. This new category of neural network was trained on a simple objective: “next token prediction”—learning to predict the next word in a sequence based on vast amounts of internet text. Today, transformers serve as the backbone of every modern Large Language Model (LLM).
• 2022: The Birth of the LLM Era. The release of ChatGPT marked the moment LLMs became highly accessible and versatile. By combining simple token completion with post-training methods like Reinforcement Learning from Human Feedback (RLHF), models transitioned from merely completing sentences to actively solving complex tasks based on user instructions.
• 2024: Reasoning Models and Agents. The introduction of reasoning models shifted the paradigm toward “test-time compute.” By allowing models more time and computational power to “think” through long-horizon, multi-step tasks, their performance in complex domains like math and coding skyrocketed. This breakthrough became the genesis of modern AI agents.

The Modern AI Recipe

Building a highly capable, modern AI system is a multi-stage process:

1. Pre-training: The model is fed massive datasets (like Common Crawl, which is an index of the web) and learns to predict the next token. At this stage, the model only knows how to complete sentences, not answer prompts.
2. Post-Training (Supervised Fine-Tuning): The model is fed specific input-and-output pairs to learn how to mimic desired formats, such as generating a memo or a spreadsheet when asked.
3. Alignment: Human preference data is used to teach the model how to output answers that are not only accurate but also helpful, safe, and aligned with user expectations.
4. Reasoning/RL (Reinforcement Learning): Models are placed into “agentic environments”—digital sandboxes where they can interact with tools, roam around, and learn to accomplish complex tasks through trial, error, and reward systems.

Understanding Agentic Terminology

To navigate the modern AI landscape, Eric outlined essential terminology:

• Agent: A model equipped with specific instructions, tools, and an execution loop to accomplish a user’s goal.
• Harness: The scaffolding or overarching software box that orchestrates an agent’s steps, tools, and skills.
• Tools: External applications or capabilities the model can access to interact with the world, such as web searching, code execution, or external software integrations.
• Skills: Text files containing specific procedures or behavioral descriptions that the model can dynamically pull into its context to execute specific tasks.

Real-World Applications and Evolving Business Models

Agentic systems are driving massive enterprise value, particularly in vertical applications. Companies are building specialized agents for coding, legal work, sales, and customer support. The standard playbook involves building a platform to capture domain-specific workflows and then codifying those workflows into an AI agent.

Because software generation is becoming commoditized by AI coding assistants, traditional “seat-based” SaaS pricing models are evolving. New businesses are increasingly adopting:

• Usage-Based Pricing: Charging based on the number of “tokens” (computational units) consumed by the user.
• Outcome-Based Pricing: Charging only when the AI successfully completes a task, such as fully resolving a customer support ticket without human intervention.

The Infrastructure Powering AI

Beneath the applications lies a $2 trillion infrastructure layer. This includes:

• Data Providers: Companies that generate task-specific data and environments to train frontier models.
• NeoClouds: Specialized cloud providers that are heavily optimized exclusively for AI compute and GPU capacity.
• Inference Providers: Platforms that provide low-latency, cost-effective hosting and serving of massive open-weight models.

Open Problems and the Future

Despite rapid advancements, the industry faces several open challenges:

• The Model vs. The Harness: Companies debate whether to invest more in raw model intelligence or in building robust harnesses and workflows around existing models.
• Memory and Continual Learning: Developing systems that can seamlessly remember past interactions and learn from user corrections over time remains a product design challenge.
• Security: As agents become fully autonomous, new threats emerge, including AI systems independently crossing network boundaries or generating misinformation.
• Hardware Scaling: Scaling data centers to meet projected AI usage could cost upwards of $5 trillion and require tripling current electricity generation, drawing comparisons to the massive railroad build-outs of the 1800s.

To bridge the gap between AI capabilities and enterprise needs, a new role is heavily emerging: the Forward Deployed Engineer (FDE). These professionals embed directly with clients to extract domain insights and customize AI workflows, proving that deep domain expertise and customer understanding are just as valuable as technical engineering.

Ultimately, Eric urges professionals from all backgrounds to experiment with these systems daily. AI is no longer restricted to computer scientists; anyone can leverage AI to build innovative solutions for the specific pain points they understand best.

Artificial Intelligence has reached an inflection point, transforming how businesses operate and how individuals interact with technology. In a comprehensive lecture at the Stanford Graduate School of Business, Mihail Eric, Head of AI at Monaco, shared his insights into the history, mechanics, applications, and future of AI. Characterizing the current era as “the age of the builder,” Eric provided a deep dive into the technology stacks and business models defining the AI ecosystem today.

The Evolution of AI: A Brief History

The modern AI boom did not happen overnight; it is the result of over a decade of compounding breakthroughs.

• 2012–2016: The Deep Learning Renaissance. While neural networks have existed since the 1950s, they experienced a massive resurgence around 2012. This was driven by a perfect convergence of new machine learning methods, the repurposing of consumer gaming GPUs for computation, and the creation of massive training datasets like ImageNet. This era yielded massive 10% to 15% performance improvements in tasks like image classification, speech recognition, and machine translation.
• 2017: The Transformer Architecture. A seminal paper by Google researchers titled “Attention Is All You Need” introduced the Transformer architecture. This new category of neural network was trained on a simple objective: “next token prediction”—learning to predict the next word in a sequence based on vast amounts of internet text. Today, transformers serve as the backbone of every modern Large Language Model (LLM).
• 2022: The Birth of the LLM Era. The release of ChatGPT marked the moment LLMs became highly accessible and versatile. By combining simple token completion with post-training methods like Reinforcement Learning from Human Feedback (RLHF), models transitioned from merely completing sentences to actively solving complex tasks based on user instructions.
• 2024: Reasoning Models and Agents. The introduction of reasoning models shifted the paradigm toward “test-time compute.” By allowing models more time and computational power to “think” through long-horizon, multi-step tasks, their performance in complex domains like math and coding skyrocketed. This breakthrough became the genesis of modern AI agents.

The Modern AI Recipe

Building a highly capable, modern AI system is a multi-stage process:

1. Pre-training: The model is fed massive datasets (like Common Crawl, which is an index of the web) and learns to predict the next token. At this stage, the model only knows how to complete sentences, not answer prompts.
2. Post-Training (Supervised Fine-Tuning): The model is fed specific input-and-output pairs to learn how to mimic desired formats, such as generating a memo or a spreadsheet when asked.
3. Alignment: Human preference data is used to teach the model how to output answers that are not only accurate but also helpful, safe, and aligned with user expectations.
4. Reasoning/RL (Reinforcement Learning): Models are placed into “agentic environments”—digital sandboxes where they can interact with tools, roam around, and learn to accomplish complex tasks through trial, error, and reward systems.

Understanding Agentic Terminology

To navigate the modern AI landscape, Eric outlined essential terminology:

• Agent: A model equipped with specific instructions, tools, and an execution loop to accomplish a user’s goal.
• Harness: The scaffolding or overarching software box that orchestrates an agent’s steps, tools, and skills.
• Tools: External applications or capabilities the model can access to interact with the world, such as web searching, code execution, or external software integrations.
• Skills: Text files containing specific procedures or behavioral descriptions that the model can dynamically pull into its context to execute specific tasks.

Real-World Applications and Evolving Business Models

Agentic systems are driving massive enterprise value, particularly in vertical applications. Companies are building specialized agents for coding, legal work, sales, and customer support. The standard playbook involves building a platform to capture domain-specific workflows and then codifying those workflows into an AI agent.

Because software generation is becoming commoditized by AI coding assistants, traditional “seat-based” SaaS pricing models are evolving. New businesses are increasingly adopting:

• Usage-Based Pricing: Charging based on the number of “tokens” (computational units) consumed by the user.
• Outcome-Based Pricing: Charging only when the AI successfully completes a task, such as fully resolving a customer support ticket without human intervention.

The Infrastructure Powering AI

Beneath the applications lies a $2 trillion infrastructure layer. This includes:

• Data Providers: Companies that generate task-specific data and environments to train frontier models.
• NeoClouds: Specialized cloud providers that are heavily optimized exclusively for AI compute and GPU capacity.
• Inference Providers: Platforms that provide low-latency, cost-effective hosting and serving of massive open-weight models.

Open Problems and the Future

Despite rapid advancements, the industry faces several open challenges:

• The Model vs. The Harness: Companies debate whether to invest more in raw model intelligence or in building robust harnesses and workflows around existing models.
• Memory and Continual Learning: Developing systems that can seamlessly remember past interactions and learn from user corrections over time remains a product design challenge.
• Security: As agents become fully autonomous, new threats emerge, including AI systems independently crossing network boundaries or generating misinformation.
• Hardware Scaling: Scaling data centers to meet projected AI usage could cost upwards of $5 trillion and require tripling current electricity generation, drawing comparisons to the massive railroad build-outs of the 1800s.

To bridge the gap between AI capabilities and enterprise needs, a new role is heavily emerging: the Forward Deployed Engineer (FDE). These professionals embed directly with clients to extract domain insights and customize AI workflows, proving that deep domain expertise and customer understanding are just as valuable as technical engineering.

Ultimately, Eric urges professionals from all backgrounds to experiment with these systems daily. AI is no longer restricted to computer scientists; anyone can leverage AI to build innovative solutions for the specific pain points they understand best.