How 80,000 companies build with AI: Products as organisms and the death of org charts | Asha Sharma

Key Concepts:

• Product as Organism
• Post-Training vs. Pre-Training
• Code-Native Interfaces
• Agentic Society
• Full-Stack Builders
• The Loop vs. The Lane
• Model Systems (Ensemble of Models)
• Seasons for Roadmap Planning
• Reinforcement Learning (RL)

1. Product as Artifact to Product as Organism

• Shift: Moving from static products (artifacts) to dynamic, learning systems (organisms).
• Details: Traditional product development involves creating an idea, solving a problem, shipping the product, and iterating with dashboards. The new paradigm focuses on products that ingest data, digest rewards models, and create outcomes.
• Example: The new IP of every single company is products that think, live, and learn.
• Mechanism: Models are tuned to specific outcomes (price, performance, quality) and improve with interactions.
• Data: Proprietary data gathered from user interactions is crucial. Synthetic data generation, rewards design, and rigorous A/B testing are also important.
• Quote: "All of a sudden these are these living organisms that just get better with the more interactions that happen."
• Connection: This concept is linked to the rise of post-training, where models are continuously optimized using data generated from their use.

2. Post-Training vs. Pre-Training

• Pre-Training: Initial training of a model on a large dataset.
• Post-Training: Fine-tuning a pre-trained model with specific data or reinforcement learning to optimize for particular tasks or outcomes.
• Economic Argument: After a model reaches 30 billion parameters, the cost of pre-training with billions of tokens becomes economically inefficient.
• Data Sources: Post-training can use proprietary data, synthetically generated data, or purchased data.
• Nathan Lambert's Study: Once a model hits 30 billion parameters, further pre-training doesn't make economic sense.
• Statistic: 50% of developers are now fine-tuning models.
• Quote: "I believe we will see just as much money spent on post-training as we will on pre-training and in the future more on post-training."
• Connection: Post-training is essential for creating "product as organism" by allowing continuous learning and adaptation.

3. Code-Native Interfaces

• Trend: Shift from graphical user interfaces (GUIs) to code-based interfaces.
• Historical Analogy: Databases moved from desktop to SQL, cloud from consoles to Terraform.
• Reasoning: Text streams connect better with Large Language Models (LLMs), facilitating composability.
• Impact: Product makers need to focus on composability and how agents will read and interact with the interface, rather than just the UI.
• Prediction: User experience (UX) will shift based on individual usage patterns, evolving automatically.
• Example: Terminals and coding environments offer better interaction with LLMs due to the text stream.
• Connection: Code-native interfaces are crucial for agentic systems and infinite scalability.

4. Agentic Society

• Definition: A future where agents are pervasive and integral to work and daily life.
• Economic Driver: Marginal cost of good output approaching zero, leading to exponential demand for productivity.
• Scaling Solution: Agents are the key to scaling productivity and output.
• Types of Agents: Embedded agents (tools, software) and embodied agents (software development reps).
• Organizational Impact: The org chart becomes the work chart, with tasks and throughput becoming more important than hierarchy. Fewer layers are needed in organizations.
• Employee Empowerment: Employees gain access to a personal agent stack, expanding their skill sets.
• Quote: "We're just starting to scratch the surface of what an agentic society actually looks like."
• Connection: Agentic society relies on code-native interfaces and continuous learning through post-training.

5. Building Successful AI Products

• Organizational Patterns:
  • Embrace AI: Everyone uses AI tools in their workflows.
  • Process Improvement: Apply AI to existing processes to make them better.
  • Growth Inflection: Use AI to improve customer experience, co-create new concepts, and expand agent capabilities.
• Pitfalls:
  • AI for AI's sake: Lack of a clear blueprint and measurement.
  • Technology Overload: Difficulty choosing from the vast number of AI tools.
• Solution: Bet on a platform or app server layer that allows swapping technologies in and out.
• Builder Mindset:
  • Full-Stack Builders: Polymaths who understand the entire product loop.
  • Focus on the Loop: Obsession with efficiency, cost, rewards system, and UI/UX.
• Quote: "It's all about the loop not not the lane here."
• Connection: Successful AI product building requires a shift from lane-based functions to loop-focused processes.

6. The Loop vs. The Lane

• Concept: Emphasizes the importance of the entire feedback loop in product development, rather than focusing on individual functional areas (lanes).
• Implication: Functions blur, feedback becomes continuous, and observability becomes the culture.
• Example: Companies like Cursor and GitHub use models fine-tuned across multiple languages for code completion. Dragon, an AI product for physicians, saw significant improvement by annotating physician-patient interactions.
• Connection: This concept is central to the "product as organism" paradigm, where continuous learning and adaptation are key.

7. Planning and Roadmaps in the AI Era

• Challenge: Rapid technological changes (e.g., new GPT releases) make traditional planning difficult.
• Approach:
  • Seasons: Define periods based on secular changes in the industry or customer needs (e.g., prototyping of AI, reasoning models, advent of agents).
  • Loose Quarterly OKRs: Set objectives for the next quarter to progress towards the overall strategy.
  • Squads: Teams operate in squads with 4-6 week goals.
  • Slack: Leave room for unplanned events and continuous disruption of the platform.
• Example: Current season is "rise of agents."
• Connection: This approach allows for flexibility and adaptation in a rapidly evolving AI landscape.

8. Reinforcement Learning (RL) and Post-Training

• RL Importance: Reinforcement learning will become a crucial product technique.
• Investment Shift: More money will be spent on post-training than pre-training.
• Economic Leverage: Adapting existing models through reinforcement learning or fine-tuning is more economically efficient than building models from scratch.
• Data Adaptation: Optimize models for specific outcomes like price, performance, or quality using various data sources (proprietary, synthetic, purchased).
• Connection: RL and post-training are essential for creating adaptive, high-performing AI products.

9. Microsoft's AI Platform

• Focus: Building a platform that enables others to create great AI products.
• Key Elements: Data residency, availability, reliability, the right selection of tools, and knowledge retrieval.
• Customer Base: 15,000 customers are building agents on the Azure service, with millions of agents running in the cloud.
• Internal Use: AI and agents are integrated into various workflows, such as summarizing live site incidents and creating prototypes.
• Quote: "The reason why I work at Microsoft is because like the whole ethos of the company is like how do I help people and businesses achieve more."

10. Leadership Lessons from Satya Nadella

• Optimism as a Renewable Resource: Nadella's ability to generate energy and renew dedication to the mission is remarkable.
• Growth Mindset: A real and important part of the company culture.
• Vibes and Mission: The importance of having a deep belief in making the world a better place.
• Connection: Nadella's leadership style emphasizes the importance of optimism, energy, and a strong sense of mission.

Synthesis/Conclusion:

The conversation with Asha Sharma highlights the transformative shifts occurring in AI product development. The move from "product as artifact" to "product as organism" necessitates a focus on continuous learning, data-driven optimization, and code-native interfaces. The rise of an agentic society requires a new organizational structure and a focus on full-stack builders who can navigate the entire product loop. Planning in the AI era demands flexibility and adaptation, with a focus on "seasons" and loose quarterly OKRs. Post-training and reinforcement learning are becoming increasingly important, offering a more efficient way to adapt models to specific use cases. Ultimately, successful AI product development requires a strong platform, a customer-centric approach, and a leadership style that emphasizes optimism and a clear sense of mission.