AI summary by TwinMind
From Software Engineers to AI Word Artisans: Filip Kozera of Wordware
Sequoia Capital
43m 4s
Wordware's Vision and Excel Analogy
- 30 million software engineers worldwide vs 750 million active Excel users
- Wordware aims to do for AI what Excel did for data analytics in the 80s
- 1980s: needed data analytics teams/engineers or used calculators
- ChatGPT equivalent to calculator - need to redo conversation each time
- Wordware enables repeatable, structured AI workflows with encoded taste
English as Programming Language Evolution
- Andre Karpathy's viral 2023 tweet: "hottest new programming language is English"
- Philip's reframe: English is Assembly Language for LLMs, not programming language
- Still need structure and programming concepts (loops, conditionals, functions)
- Syntax becomes less important, but need to know what you're trying to say
- Not everyone can use it effectively - some people "don't have much to say"
Wordware Product Overview
- Natural language IDE with programming concepts
- Editor supports: loops, conditional statements, function calling, code execution blocks
- Three deployment methods:
- API for products/chatbots (more complex than vanilla OpenAI calls)
- AI-native workflows (brain more complex than simple prompt chains)
- GitHub for AI - sharing/forking AI components
- Focus on agents: software taking inputs/outputs with fuzzy processing stages
No-Code vs Code-Like Approach
- Philip bristles at "no-code" label - hasn't reached ceiling with any clients
- Word Artisans/Word Crafters still coding, just structuring English precisely
- Brings structure to intrinsically fuzzy English language
- Document format for structuring agents similar to traditional code
- Escape hatch: code execution blocks when wordware insufficient
Making English More Programmable
- Challenge: English fuzzy vs code's precision and expressiveness
- Solution: enough structure without losing fuzziness/power
- Editor enables quick iteration - "clicking run 100 times quickly"
- Develops intuition about what works through rapid testing
- Many companies lack evaluation datasets initially
- Speed of creativity higher with wordware - learn objectives while encountering model limitations
Current User Base and Evolution
- Current ICP: somewhat technical CEOs, technical PMs, high-level executives
- Example: MOX from Instacart spent 4 days refining idea instead of hiring team
- Moving toward blank canvas approach - describe idea, AI guesses structure
- Plan to use O3 for first draft generation
- Still allow fine-grained editing capabilities
Next Billion Developers Concept
- 30 million software engineers vs 750 million Excel users as analogy
- Calculator (ChatGPT) vs spreadsheet (Wordware) comparison
- Next 500 million to 1 billion AI users need two things:
- Knowledge of what actions should happen
- Taste in how to execute
- Future: everyone feels like CEO with thousand knowledge workers beneath them
- Post-resource scarcity: still have equivalent of billionaires based on influence/taste
AI as Intern Analogy
- Current AI like intern after university
- Need to write clear instructions on paper
- Specify: job title, documents/inputs, data sources, expected outputs
- Important caveats:
- Data sources must be trusted (not random internet search)
- Must trust AI with the task (e.g., sending thousand emails)
- Knowledge work increasingly automated
User Interface Philosophy
- Document style better than chat for complex work
- Two sentences often "just enough for model to hang itself"
- Lead investor: "two sentences just enough for model to hang itself"
- Problem with chat: insufficient context leads to wrong outputs
- Transformer = new transistor, models = mainframes
- Need better GUIs beyond current chat interfaces
- Programmable documents as preferred interface
- Want transcendent abstraction layers - zoom from sentence to detailed modules
Interface Design Evolution
- Previous approach: node-based 2D canvas
- Problem: complex reflection loops run out of dimensions
- 2D canvas becomes cluttered for complex systems (like Uber app)
- Code structure better for expressing complex concepts
- Document format allows hierarchical complexity management
- Enables both high-level vision and fine-grained control
Technical Perspectives and Hot Takes
- Pre-training still matters despite DeepSeek hype
- DeepSeek: utilized cool techniques but not revolutionary
- Community will learn from techniques, but not paradigm shift
- Pre-training not hitting wall - intelligence scales logarithmically with resources
- 2x intelligence improvement = exponential work capability increase
- Favorite tools: Script for content editing, Granola for daily use
- Gemini 2.0 Pro impressive - 6,000 page PDF ingestion capability
Future Vision
- 10-year goal: save human creative vision, not replace it
- Humans as CEOs setting strategy/intent, AI executing
- Taste and creative vision become more important
- Example: luxury shoe brand requires ingrained taste/intent
- Manufacturing automated, human job = partnerships with other CEOs
- Work in 10 years: structured documents expressing vision with AI copilot assistance
Hiring and Resources
- Wordware actively hiring aggressively
- Recommended learning: Deep Learning.ai resources (bottom to practical layer)
- 3Blue1Brown YouTube channel for understanding underlying technology
- Focus on comprehensive AI education from fundamentals to application
