In this livestream, the host extensively tests AI coding models like Quinn 3 Max, Kimmy K2, and Soma stealth models, highlighting their strengths, weaknesses, and practical challenges, particularly noting Quinn 3 Max’s slow speed and Kimmy K2’s promising capabilities. He also shares insights on AI development workflows, pricing, and the future of AI in software engineering, expressing optimism about ongoing advancements and encouraging community collaboration.
In this extensive livestream, the host dives deep into testing and exploring various AI coding models, with a particular focus on Quinn 3 Max, Kimmy K2, and the Soma stealth models. He begins by setting up his environment, discussing his hardware setup including the Framework desktop, and sharing his experiences with different AI coding tools like Root Code, Open Router, and Cloud Code. The host experiments with Quinn 3 Max’s coding capabilities, especially its performance in generating a Python pool game, noting that while it can produce code, its speed is painfully slow—often around one token per second—making it impractical for daily use. He also explores the impact of temperature settings on the model’s output but finds inconsistent results, suggesting that temperature might not significantly affect Quinn 3 Max’s performance.
The host contrasts Quinn 3 Max with other models like Kimmy K2 and Soma’s Sky and Dusk, highlighting that Kimmy K2 shows promise in coding and design tasks, albeit with some bugs and incomplete features, while Soma models are extremely fast but seem less capable in coding, more suited for general or planning tasks. He showcases various AI-generated projects, including portfolio websites and a web-based operating system, pointing out the strengths and weaknesses of each model in handling front-end design, tool calling, and complex coding tasks. Despite some impressive outputs, many AI-generated projects require significant manual fixes and improvements, reflecting the current limitations of these models.
Throughout the stream, the host shares insights on AI development workflows, discussing the challenges of managing large codebases with AI, the importance of tool calling support in AI agents, and the need for better integration and configuration tools to streamline AI-assisted coding. He also touches on the pricing and rate limits of different AI providers, noting that while some models like Quinn 3 Max are cheap due to slow speeds, others like Kimmy K2 and GPT5 Pro offer better performance at higher costs. The host expresses enthusiasm for emerging tools like Warp terminal and Cloud Code extensions, which enhance the coding experience by integrating AI more seamlessly into development environments.
The conversation also delves into broader reflections on AI’s role in software engineering, with the host sharing personal anecdotes from his programming career and interviews, emphasizing how AI reinvigorates his passion for coding by offloading repetitive tasks and enabling him to focus on architecture and big-picture design. He discusses the evolving landscape of AI models, speculating that the future may favor smaller, hyper-focused models tailored to specific coding domains rather than massive generalist models. The host also highlights the importance of community and collaboration in advancing AI tooling, encouraging viewers to share their experiences and contribute to open-source evaluation frameworks.
In closing, the host acknowledges the current limitations and frustrations with some AI models, particularly Quinn 3 Max’s slow performance and occasional instability, but remains optimistic about the rapid progress in AI-assisted coding. He invites viewers to join the Discord community to exchange ideas and results, emphasizing the exciting potential of AI to transform software development workflows. The stream blends technical experimentation with thoughtful commentary on AI’s impact, offering a comprehensive snapshot of the state of AI coding tools as of mid-2024.