The required review group for this topic includes AI Research & Development Strategists, Cognitive Systems Engineers, and Advanced Software Development Practitioners. The discussion centers on creating highly productive development paradigms leveraging Large Language Models (LLMs) via agentic workflows, which directly impacts these specialized fields.
Abstract:
This podcast episode from "the lid in space" features host Alessio interviewing the founding team of SolveIt (Eric, Jeremy, and Jono), discussing the evolution from their previous work ("Answer") to their current platform, SolveIt. The core thesis of SolveIt is the development of an AI platform built around an integrated research and development process, inspired by Edison's labs, that treats AI as a means to create "superhumanly productive" individuals via a "human in the loop" methodology.
The platform emphasizes extreme granularity in task decomposition, real-time feedback, and enabling deep user interaction with the underlying computational environment. Key technological features highlighted include providing each user with a persistent, provisioned Linux container (akin to a 1990s VPS) accessible via a unique URL, allowing for full software installation and deployment inside the session environment without reliance on external cloud providers like AWS.
The methodology prioritizes iterative development, step-by-step explanability, and composability, where entire dialogues or functional modules can be imported as Python libraries into other dialogues, radically lowering the friction of reusing complex AI-assisted workflows. Use cases demonstrated include complex prose and technical artifact creation (like a web application development stack), book authoring with integrated fact-checking against reader feedback, and rapid creation of custom, ephemeral software tools. The founders stress that the goal of the course and the tool is to maximize human flourishing by enabling users to build highly customized solutions economically, rather than simply outsourcing tasks to opaque AI agents.
Exploring the SolveIt Paradigm: Human-Agent Co-Development and Persistent Computing Environments
0:00:04 Introduction: Host Alessio welcomes the "Answer Gang"—Jeremy, Eric, and Jono—to discuss the transition from "Answer" to their new venture, "SolveIt," roughly two years post-inception.
0:01:20 Foundational Philosophy (Edison's Labs Model): The founders articulate a goal to maximize the utility of AI by creating an integrated research and product development structure, departing from conventional AI lab organization. They aim to leverage AI as a capability multiplier for small, highly capable teams.
0:02:49 Iterative MVP Approach: The strategy rejects building one massive super-application in favor of parallel exploration across many domains using an iterative, MVP-driven approach.
0:03:08 Human-in-the-Loop (HITL) Centrality: The core idea is an end-to-end process driven by the human agent, where problems are broken down into sub-problems, ensuring the human operator understands every step.
0:04:29 Extreme Internal Productivity: A small team (averaging 9-12 people) built a complete, in-house stack, including a web application development platform, deployment platform, DevOps, and professional services infrastructure (legal, accounting), all running on their own tools, notably without using AWS or Google Cloud.
0:06:14 Training as Compression, SolveIt as Expansion: The platform is framed as the tool for expansion—turning the latent space of the model back into tangible, valuable artifacts.
0:07:21 Smallest Iterative Steps: The core tenet, developed independently by both founders over decades, is to execute tasks in the smallest possible iterative steps, seeking immediate and accurate feedback (echoing OODA loops and the Toyota Production System).
0:08:55 Tool Integration: The platform makes it exceptionally easy to add tools (any Python function) to the LLM, transforming simple functions into immediately usable agent capabilities.
0:09:41 Return to the General Purpose Computer: The platform enables a return to hackable, general-purpose computing, contrasting with layered, restrictive cloud infrastructure interfaces.
11:52 Persistent Linux Container: Each user receives a unique URL attached to a persistent Linux Docker container, functioning as a personal VPS where software can be installed and servers run (e.g., a Discord bot was built and run entirely within an instance).
13:19 Dialogue Engineering: This concept, predating "context engineering," involves structured conversations. Users can edit, delete, or reorder messages, overcoming the context pollution issues common in standard chat interfaces by allowing precise history curation.
17:50 Jupyter Notebook Analogy: The interface supports a "code mode" where users can execute code (like a division by zero example) live and receive immediate inspection/debugging feedback, facilitated by the shared environment visibility.
18:38 Learning/Thinking Modes: Specific modes guide the AI's response: Learning Mode focuses on teaching the user step-by-step, while Thinking Mode engages in deeper reasoning before providing output.
33:25 Fact-Checking and Authoring: Eric demonstrates using the platform for book writing (Chapter 11 of The Incorruptible), loading the chapter text as a variable and using an integrated workflow to systematically check facts and incorporate structured feedback from test reader comments chapter-by-chapter.
39:47 Dialogue Composability (Library System): A key feature is turning dialogues into importable Python modules. A module created in one dialogue (e.g., reading CSV reader comments) can be imported directly into another, making true, functional component sharing possible across distinct sessions.
44:45 Scratching Own Itches: The decision for building first-party features is driven by internal necessity and discovering highly valuable workflows, such as their superior AI-driven meeting notes system.
46:18 High-Quality Content Generation: The team successfully took on the challenge of converting Andrej Karpathy's high-quality video content into an equally high-quality blog post using a detailed, human-guided, line-by-line editing process within the dialogue.
49:27 Course and Community: The founders are launching the full SolveIt course, emphasizing that the course teaches the methodology (iterative, human-driven thinking) rather than just the tool itself. They highlight the exceptional, life-changing community formed by the first cohort.