20VC with Harry Stebbings - Kevin Scott, CTO @ Microsoft: An Evaluation of Deepseek and How We Underestimate the Chinese
The conversation highlights the current technological paradigm shift driven by AI, comparing it to past shifts like the internet and mobile. The speaker emphasizes the importance of product development over mere technical fascination, arguing that models and infrastructure are valuable only when they meet user needs through products. The discussion also touches on the role of startups and large companies in integrating AI, suggesting that both have opportunities to innovate and create value. The speaker believes that AI tools are more accessible than ever, enabling rapid experimentation and iteration.
The conversation also delves into the future of AI, particularly in software development and data usage. The speaker predicts that AI will generate most new code in the future, raising the level of abstraction in programming. They discuss the importance of high-quality data and the challenges in assessing its value. The speaker also addresses the potential of AI in healthcare, suggesting that AI models could surpass average general practitioners in diagnostics. The discussion concludes with a call for increased investment in education and the deployment of AI tools for public good.
Key Points:
- Product development is crucial; models and infrastructure must meet user needs.
- AI tools are more accessible, enabling rapid experimentation and iteration.
- AI will generate most new code, changing programming abstraction levels.
- High-quality data is essential; assessing its value remains challenging.
- AI has potential in healthcare, possibly surpassing average GPs in diagnostics.
Details:
1. 🚀 The Rise of Entrepreneurial Spirit
- The current market conditions are highly favorable for entrepreneurial activities, making it an ideal time to embrace entrepreneurial endeavors.
- Scaling laws in the market are viewed as having limitless potential for growth, encouraging innovation and expansion without traditional constraints.
- The future of product management is expected to require domain experts, indicating a growing need for specialized knowledge in product development and management.
- Future agents will focus less on transactional and session-oriented interactions, moving towards more integrated and continuous customer engagement strategies.
2. 🌟 Embracing AI's Paradigm Shift
- In moments of technological paradigm shifts, initial confusion is common, similar to early internet and mobile technology days.
- During transitions, it's crucial not to remain passive; active iteration and learning are necessary, despite potential mistakes.
- Entrepreneurial spirit is critical during paradigm shifts as it presents opportunities for innovation.
- Past lessons on product development and exploration should guide current strategies, emphasizing that models aren't products.
- Specific examples such as the shift from traditional to digital media illustrate how active engagement led to new business models and growth.
- The AI revolution mirrors the early days of the internet, where those who adapted quickly gained substantial advantages.
3. 🧩 The Essence of Models and Products in AI
- Focus on creating good products that meet user needs rather than getting lost in technical details.
- Ideas must be tested quickly to validate their effectiveness and to adjust based on real data.
- Innovation requires launching new concepts, gathering data, and iterating based on feedback.
- Models are valuable but only if connected to user needs through a product.
- Infrastructure around models, including efficient compute, is crucial for monetization.
- The majority of value in AI comes from the product itself, not just the models or infrastructure.
4. 🏢 Innovators vs. Tech Giants: Who Benefits Most?
- Startups have the advantage of integrating new technologies from scratch, providing them with flexibility and a high potential for innovation.
- Tech giants such as Microsoft and Google benefit from established distribution networks, allowing them to integrate AI effectively and maximize market reach.
- Both startups and large enterprises play crucial roles in driving value creation during technology cycles, each contributing uniquely.
- Large companies focus on enhancing their existing products with new capabilities, leveraging their substantial customer bases to introduce innovations.
- Microsoft Research embodies a drive for groundbreaking innovations, mirroring the innovative spirit of startups.
- A diverse ecosystem, with a variety of entities exploring new technologies, is essential for uncovering valuable opportunities and driving growth.
- The current AI platform transition presents unprecedented tools, infrastructure, and platforms that are both affordable and accessible, democratizing innovation.
5. 📈 Debunking Scaling Myths in AI
- Current beliefs that AI scaling laws are reaching their limits are incorrect, as AI models can still be enhanced in capability and complexity.
- Experts believe AI could scale beyond human cognitive limits, which are constrained by biological factors like neuron count and energy consumption.
- Future diminishing returns in AI scaling are expected, where increased investments might not yield proportional improvements due to escalating costs.
- Despite the anticipation of diminishing returns, no clear point has been reached yet, indicating continued potential for AI innovation.
- The idea that AI could surpass human cognitive limits underscores the transformative potential of ongoing AI advancements.
6. 🔍 Data's Role in AI Efficiency and Effectiveness
- The mix of synthetic data is increasing, and high-quality data is becoming more crucial, especially in the post-training phases of model development, compared to low-quality data.
- High-quality data combined with expert human feedback can significantly enhance the training of larger models, providing more value than generic data from the web.
- There is a lack of quantitative assessment to measure the incremental value of data tokens to model quality, leading to unfounded assertions about data value.
- There is a disconnect between perceived and actual data value in model capabilities, with models often being misused as databases rather than tools for reasoning.
- Models should be designed to reason over information rather than merely store facts, requiring different training data for reasoning capabilities.
7. 🤖 Transforming Human-Agent Interaction
- The performance of inference models continues to improve significantly year over year, optimizing their performance and reducing costs.
- Models have increased in size while API calls have become cheaper, partially due to advancements in hardware that provide around a 2X price performance improvement each generation.
- Public reactions to releases, such as the deep seek R1, provide valuable insights into market expectations and preferences, influencing future releases.
- Developers demand more options and choices, highlighting the need for more customizable solutions in AI offerings.
- There is a shift from open-source idealism to a pragmatic approach in AI development, balancing curiosity and practical decision-making.
- The future of AI will likely see both open and closed systems coexisting, similar to the search engine market where open-source and proprietary solutions thrive together.
- Infrastructure products in AI will be diverse, with people choosing between building from open-source projects or opting for established platforms like Azure or Google.
8. 💻 Software Development's AI Revolution
- AI enables non-programmers to utilize computing devices without needing to write code, lowering the barrier to entry and changing traditional software development paradigms.
- The role of engineers will shift towards building infrastructure for AI capabilities rather than anticipating user needs and coding applications.
- Agents will become a primary interface for software capabilities, with domain experts guiding their development to better serve specific industries.
- Despite skepticism about immediate adoption, AI agents are rapidly becoming essential tools, with developers showing strong attachment to them.
- The lack of 'lock-in' with AI agents encourages continuous improvement and user retention based on utility rather than exclusivity.
- AI agents currently lack robust memory, but improvements are expected within the next year, enhancing their ability to remember past interactions and user preferences.
- Future AI agents aim to handle increasingly complex tasks autonomously, akin to delegating work to a coworker.
- The skepticism surrounding AI's capabilities is countered by optimism in technological advancements and potential applications.
9. ⏱ Accelerating Innovation: Overcoming Challenges
9.1. Shift to AI-Generated Code
9.2. Implications for Programmers and Teams
10. 💡 Wisdom from Competitors and Leaders
- Large-scale technology operations often slow down decision-making processes, which can be necessary but also limit product development speed. Leaders aim to enhance these processes by overcoming infrastructural limitations.
- The infrastructure for technology has been rapidly developed, akin to 'running at a thousand miles an hour' since the release of GPT-4, highlighting the industry's need to keep pace with AI advancements.
- There is a push to reduce barriers between engineers' ambitions and their ability to implement ideas, with AI playing a key role in enhancing this capability within organizations like Microsoft.
- Technical debt in engineering teams poses significant challenges, as it requires balancing speed with precision. This debt accumulates like financial debt, accruing 'interest' and potentially causing issues if not managed.
- AI tools are seen as potential solutions to transform the zero-sum problem of technical debt into a non-zero-sum scenario, thereby reducing the need for traditional trade-offs.
- Microsoft is actively pursuing a research project aimed at eliminating technical debt at scale through AI tools, marking a significant area of innovation and excitement.
- Current AI tools have surpassed expectations, being more capable than anticipated compared to two years ago.
11. 🌍 Global AI Trends and Predictions
- Anthropic is recognized as a strong competitor in the AI landscape, with a focus on effective leadership under Dario, highlighting the importance of leadership in AI innovation.
- A strategic focus on leveraging individual and team strengths ('genius areas') rather than improving weaknesses ('idiot areas') can drive significant success in AI development.
- The speaker emphasizes the importance of delegation as a means of focusing on strengths, suggesting that successful AI leadership often involves complementing team skills rather than trying to excel in all areas.
- Delegation and strategic focus are presented as key leadership strategies that can influence AI trends by aligning team efforts with core competencies.