Abstract:

This document is a transcript of a multi-day conversation from the "Solve It With Code" community Discord server. The discussion centers on the use, philosophy, and future development of the SolveIt platform, an AI-assisted coding environment. Key topics include the optimal paradigm for human-AI interaction, which founder Jeremy Howard describes as a blend of natural language and high-level code. Community members provide extensive user feedback, suggesting features like an outliner-style notebook interface and improved dialogue management. Several users showcase impressive weekend projects built with SolveIt, including AI agents and hackathon entries, detailing their development process and choice of tools. The conversation also features technical deep dives into using libraries like Lisette with local models via Ollama, debates on data tools like Polars versus DuckDB, and philosophical discussions on the nature of the current AI boom and development workflows. The transcript captures a vibrant, collaborative community actively shaping a tool while sharing knowledge and supporting each other's learning and projects.

Highlights from the 'Solve It With Code' Community Discussion

• 10/11/25, 7:25 PM The SolveIt Philosophy: Jeremy Howard posits that the most effective way to interact with AI is through a combination of a high-level programming language (like Python) and natural language, mirroring how human teams collaborate. This hybrid approach is the core design principle of the SolveIt platform.
• 10/12/25, 5:44 AM User Feedback on Workflow: A user named Radek provides feedback on the SolveIt user experience, requesting features for easier management of dialogues, such as quick deletion and better naming conventions for duplicates. This would better support a "lateral thinking" approach by allowing users to explore multiple problem-solving paths simultaneously without clutter.
• 10/12/25, 12:16 PM Feature Request for Outliner View: User haya 隼 suggests implementing an outliner format for SolveIt notebooks to create a more hierarchical and flexible document structure, similar to OmniOutliner. The discussion explores how this would differ from the existing markdown heading folding feature.
• 10/13/25, 12:59 AM Community Project Showcase: User c5huracan shares two impressive weekend projects built using SolveIt: an MVP for a Personal Finance AI Agent and an agentic Crypto Twitter bot that uses Playwright to automate posting. This highlights the platform's utility for rapid prototyping.
• 10/13/25, 12:21 PM Hackathon Experiences: User Pol Avec shares a detailed account of using SolveIt in a hackathon. They found it highly effective for solo development but noted challenges for team collaboration, particularly when integrating with front-end frameworks, leading them to use React for projects requiring shared styling work.
• 10/13/25, 3:16 PM Technical Discussion on Local Models: A technical Q&A unfolds around using the Lisette library with local models via Ollama. Key takeaways include the correct syntax to use (ollama_chat/<model_name>) and troubleshooting tool-calling capabilities with different models like Qwen3.
• 10/13/25, 9:23 PM Philosophical View on the AI Boom: User haya 隼 offers a unique perspective, suggesting the current AI investment boom is driven by a "defensive (fear)" of China's rise rather than offensive greed. They frame the pursuit of AGI as a "deus ex machina" the West hopes will solve its systemic problems.
• 10/15/25, 10:09 AM AI as a Compiler vs. an Interactive Partner: Sparked by a Geohot blog post, the community discusses the "AI is a compiler" metaphor. They contrast this with the SolveIt approach, which treats AI as an interactive partner in an exploratory, Jupyter-style workflow rather than a tool you simply "recompile" prompts with.
• 10/15/25, 8:24 PM Speculation on Apple's AI Hardware: User haya 隼 predicts that Apple's upcoming M5 chips will offer a massive leap in on-device AI performance, making Mac Pro machines powerful tools for local inference and ML development, and boosting the relevance of frameworks like MLX.
• 10/16/25, 1:52 AM Troubleshooting on the Plash Platform: Community members discuss how to run cron jobs on Pla.sh, the hosting service for SolveIt apps. While a solution using setup.sh is proposed, a user later reports that browser automation with Playwright fails in the Plash environment.
• 10/16/25, 10:37 AM Data Tools Debate: Polars vs. DuckDB: The community debates the merits of replacing dataframe libraries like Polars with DuckDB. The consensus, guided by Jeremy Howard, is that these tools serve different purposes; DuckDB is a columnar database ideal for analytics but is not a suitable replacement for transactional databases like SQLite or for all dataframe use cases.
• 10/17/25, 1:54 PM Advanced AI Agent Architecture with Prolog: Pol Avec shares an idea from a podcast about building more robust AI agents by combining LLMs with Prolog. In this architecture, the LLM translates natural language to formal logic for a Prolog knowledge base, which then handles deterministic reasoning and state management, preventing common agent failures.
• 10/17/25, 4:44 PM Using SolveIt for Accelerated Learning: Members share methods for using SolveIt to learn from educational materials. They discuss programmatically fetching YouTube video transcripts (including from fast.ai courses and Andrej Karpathy's videos) and using them as context for interactive learning sessions within a SolveIt dialogue.