To AI or Not to AI
6 hours ago
- #LLM Challenges
- #Workflow Experiment
- #AI Development
- A two-week experiment was conducted to build an app using full AI assistance, exploring the potential and frustrations of LLM-based development workflows.
- The team built a simplified version of Facebook Ads called 'adbrew' to address specific use cases, leveraging Facebook's API.
- The development process involved defining issues, prompting the AI to implement them, refining requirements, reviewing generated code, and deploying.
- Several challenges were encountered, including insufficient context, lack of maintainability, disrupted workflow, hallucinations (incorrect assumptions by AI), and the Pareto principle (80% solution with 80% effort).
- The experiment was halted after two weeks due to messy code, loss of control, and lack of enjoyment in the process.
- Despite the challenges, the team continues to use AI for tasks like powerful search, rubber ducking, code snippets, tests, and language-related tasks.
- The team prefers local LLMs over cloud-based ones for data control but does not plan to expand AI usage beyond current applications unless technology improves.