Benedict Evans: Making Sense of AI and Tech Evolution

Comparing AI to Past Tech Eras

• 💡 The current AI era is likened to the mid-90s internet or mid-2000s mobile, where the technology's potential was clear but its market structure and value capture were not.
• 📌 Unlike previous eras with predictable roadmaps (e.g., iPhone capabilities, broadband deployment), the generative AI roadmap is unknown, making it hard to predict future capabilities.
• 🧠 There's a debate on whether AI is just another building block for a new platform or a fundamental change akin to computing or electricity.

Current State of AI Adoption

• 📊 Surveys show that while many have tried AI tools, a significant portion (around 50%) don't know what to do with them and don't use them again.
• 🎯 Early adopters, particularly those in tech-forward roles or with STEM degrees, are the primary users, often forcing ways to integrate AI into their workflows.
• 🔑 The challenge for AI is to break existing patterns and form new habits, similar to how users learned to leverage Google.

AI Value Chain and Product Strategy

• 🛠️ The AI landscape includes frontier model builders (a dozen organizations with significant investment) and hundreds of SAS companies using LLMs to solve specific industry problems.
• 🧩 A "fuzzy middle" exists where companies are pushed to use AI without clear guidance on practical application or product strategy, reminiscent of the "metaverse strategy" push.
• 📈 Product differentiation for LLMs is currently limited, often relying on UI/UX improvements rather than fundamentally different model capabilities.

Defensibility in the AI Era

• ⚠️ Unlike search engines, LLMs currently lack network effects; their utility doesn't inherently improve with more users, though features like "memory" aim for stickiness.
• 🚀 The value of AI products often lies in the productization, go-to-market strategy, and deep understanding of customer needs, rather than just being a "thin wrapper" around an LLM.
• 💡 Historical examples like spell-checkers becoming integrated features suggest that many current standalone AI tools may eventually be absorbed into platforms.

Probabilistic AI Systems

• 🧠 A key distinction of AI is its probabilistic nature compared to traditional deterministic software, which introduces challenges in consistent output.
• 🎯 AI excels at tasks that are easy to explain to a person with no context (like an intern), but struggles with complex tasks requiring deep context or experience.
• 🔬 Managing the probabilistic output through tooling, pre-processing, and post-processing is crucial for building reliable AI products.

Personal AI Use Cases

• 💬 The speaker, Benedict Evans, struggles to find personal use cases for LLMs beyond proofreading (which he sees as advanced spell-check).
• 🚫 He finds the "uncanny valley" effect in AI voice modes off-putting and doesn't have use cases for brainstorming or generating generic content like marketing slogans.
• 🧑‍💻 His work involves a very specific and narrow set of tasks where current AI tools are not particularly helpful, highlighting that AI utility is highly dependent on individual roles and needs.