DocGPT
- 1. 1 1 DocGPT Amplifying Human Intelligence to Tackle the World's Toughest Challenges
- 2. 2 Don’t fear intelligent machines. Work with them. Elo ra ti ngs 0 850 1700 2550 3400 Human Machine Human + Machine
- 3. 3 There’s a need for better interfaces. • Copilot > Autopilot: Photoshop creates content together with user • Humans are great at setting goals and managing • Machines are fast in learning and computation • AI tools need to provide an interface to connect the strengths DALLE is great, but Photoshop AI is much better.
- 4. 4 Introducing DocGPT Copilot for learning & research: answers any question about any document Provides a familiar interface, premium PDF editor & private document cloud Creates complex content like essays, summaries, even a Bachelor’s thesis
- 5. 5 DocGPT is a collection of four different technical services. Architecture Document Cloud AI Assistant Vector Database Document Editor
- 6. Chat • Generates summary, key concepts and suggests questions for every documents • Answers any question • Adds references • Uses tools Insights • Generates complex forms of content based on 100s of templates (essays, blog posts, …) • All chat features, but longer content • Offers choices & gives user the ability to edit or regenerate any part
- 7. 7 Magic Ingredients Templates Choices Autonomous Tools Autonomous Tree of Thought Title Agenda Intro Topic for the introduction of the Essay “Why GPT-4“? Problem Solution coming soon … Size Speed Cost
- 8. 8 Templates Large object, that divides content into a sequence of customizable prompts Enables autonomous usage of tools & presentation of choices Generated by GPT-4: possibility to add 1000s of templates effortlessly
- 9. 9 Introduction ideas Are language models capable of being general problem solvers? Can LLMs be designed to engage in more conscious decision-making? Can tree of thoughts enhance problem-solving capabilities? Enter your own idea Can Tree of Thoughts Improve the Intelligence of Language Models? Choices Stay in control of what DocGPT generates Defined in templates 3 ideas, choose 1 or enter your own idea
- 10. 10 Autonomous Tools Research tools May June July Computational tools Intelligence tools • Can be triggered autonomously by DocGPT or explicitly by the user • Gather information from the internet • Solve logic problems, visualize data, etc. • Increase intelligence of AI model
- 11. 11 “Write an essay on how to predict the prices of Used Cars using AI.“ … Architecture of the Model Choose the most important metric to evaluate the performance R2 MAE MSE Predict the best performing model, given the chosen metric Random Forest Neural Network Linear Regression Continue the essay by proposing the chosen model architecture Architecture: Neural Network Model starts generating essay on user input Model decision: Start tree of thoughts Model generates sub- tasks Evaluation Evaluation Model continues generating essay Autonomous Tree of Thoughts coming soon Split problem into parts → Generate n solutions → evaluate → choose best path Exists in the form of a tool Dramatically increases logical intelligence of the AI model Artificial reflexion mechanism, end to end
- 12. Chat about a document View a document Edit a document AI can use tools Generate long-form content (Insights) Guide the AI while it’s generating content (Choices) Chat about a document View a document Edit a document AI can use tools Generate long-form content (Insights) Guide the AI while it’s generating content (Choices) DocGPT Google Tailwind Chat about a document View a document Edit a document AI can use tools Generate long-form content (Insights) Guide the AI while it’s generating content (Choices) ChatGPT + Plugins Experiment
- 13. Knowledge Workers (Lawyers, Accountants, Educators, Writers, Editors, Office Administrators, …) Students 13 A research & document interface between humans and AI has the potential to become as relevant as Microsoft Word. Market Analysis Scientific Research Community 7 million + potential users ( UNESCO ) 235 million potential users ( Universities only) 1 billion + potential users ( Forbes)
- 14. 14 • 700 users • 350% MoM growth • 2,000+ chats (all saved, evaluations included) • 350€ marketing spendings 0 175 350 525 700 Number of users February March April May 0 500 1000 1500 2000 Number of chats February March April May 0 50 100 150 200 Marke ti ng spendings in € February March April May Growth
- 15. 15 Cost for 700 users per month 0 4 8 12 16 Operational Cost • Cost for 700 users: $15 / month • Expecting reduction to $8 • Database improvements • Declining cost for LLM APIs AI Services Database Cumulative
- 16. 16 Monetization Options Subscription Pay per word Pay per PDF Sell data • Currently “pay per word“ • $1.99 per 1,000 generated words • Chats are currently free: Users only pay for insight generations • Users get 2,500 words free on registration
- 17. 17 Meet Our Team Jan Luca Sandmann CEO Katharina Matthies Head of Marketing Jan Grünke Developer