This document is a transcript of a conversation from the #solvers channel in the "Solve It With Code" Discord community, spanning from October 15 to October 18, 2025. The discussion covers a wide range of topics centered around the "SolveIt" coding platform. A key conceptual insight shared is the idea of a SolveIt dialog functioning as a "manually created reasoning trace" for an LLM, providing verified, step-by-step context that improves problem-solving reliability over single prompts.
Members share numerous practical applications and technical solutions, including a successful use case of solving a complex brain MRI slicing problem after other AI tools failed. This sparks a sub-discussion on medical imaging formats (DICOM vs. NIfTI) and 3D printing. The community also develops and shares several advanced utilities, such as methods for bridging SolveIt to local machines and GPUs via SSH and ngrok, tools for managing dialogs, and techniques for creating custom UI enhancements. The conversation highlights a collaborative environment focused on pushing the platform's capabilities, sharing knowledge, and exploring its applications in research, workflow automation, and UI development.
Highlights from the 'Solve It With Code' Community Discussion
10/15/25, 11:43 AM - SolveIt as a "Reasoning Trace": Marius proposes that a SolveIt dialog acts as a "manually created reasoning trace" for an LLM. By building a sequence of verified steps (cells and their outputs), the user provides a reliable context for the LLM to solve complex problems, which is more robust than a one-shot prompt where the LLM's internal "reasoning" cannot be checked.
10/15/25, 6:54 PM - Bridging SolveIt to a Local Computer: David (TA) shares a successful test of connecting SolveIt to a local computer using an ngrok TCP Tunnel. This allows for executing code and interacting with local resources directly from the cloud-based SolveIt environment.
10/15/25, 8:40 PM - Solving a Complex Research Problem: Marius details successfully using SolveIt to solve a complex brain MRI problem (extracting an oblique slice from a 3D image). He notes that he had previously failed to solve it using Gemini 2.5 Pro, highlighting SolveIt's effectiveness for "algorithmic problem solving."
10/16/25, 8:28 AM - Medical Imaging Discussion: The MRI success story sparks a discussion on medical imaging. Marius recommends using the NIfTI format over DICOM and suggests libraries like ANTsPyx (for image registration) and nibabel (for working with NIfTI files).
10/16/25, 2:54 PM - Remote Execution on WSL: David (TA) demonstrates running SolveIt code on a remote Windows Subsystem for Linux (WSL) instance via an SSH connection, complete with variable persistence, further expanding local integration capabilities.
10/16/25, 8:19 PM - Running Code on a Local GPU: Inspired by the SSH tunneling work, Chris Thomas shares a method to "export and run on a local GPU." This approach keeps the user-friendly notebook interface of SolveIt while offloading compute-intensive tasks like training a fast.ai model to more powerful local hardware.
10/17/25, 2:18 AM - SolveIt Tools vs. Claude Skills: Jeremy Howard shares an article on "Claude Skills," noting that SolveIt's "tools" are a "huge superset" of this concept because they have access to a full, running Python interpreter, offering more power and flexibility.
10/17/25, 11:31 AM - LLM-Powered Workflow Automation: David (TA) showcases a Proof of Concept where the LLM uses dialoghelper functions as tools to automatically add note messages and annotate code within the dialog, demonstrating how SolveIt can be used to improve its own workflow efficiency.
10/17/25, 11:49 AM - Collaboration and Sharing: A discussion on sharing dialogs clarifies that users can "Publish" a dialog to share a read-only version or use "guest mode" on a dedicated instance to allow live, collaborative editing with others, even those without a SolveIt account.
10/17/25, 2:35 PM & 7:54 PM - Custom UI Enhancements: David (TA) shares several dialogs that customize the SolveIt UI, such as running the terminal in a side panel instead of a new window and creating a live, auto-updating counter for different message types using FastHTML.