The first 40 months of the AI era
(lzon.ca)Written in March 2026, this article reflects on 40 months into the AI era since ChatGPT's launch, detailing the practical utility and limitations of AI coding tools. The author, while impressed by AI's content and code generation capabilities, points out its struggles with maintaining context in complex projects, often requiring human intervention and questioning its real-world efficiency. The piece particularly highlights the potential of new computer input methods like Claude Code, yet underscores the need for further technological advancement for full utilization.
- 1AI coding tools are useful for initial code generation and simple, repetitive tasks, but they clearly show limitations in complex or iterative development projects, where AI often loses context and requires significant human intervention and code replacement.
- 2Integrated natural language interfaces like Claude Code are establishing themselves as a 'new form of computer input and control' alongside keyboards, mice, and CLIs, demonstrating the potential for revolutionary changes in developer productivity and AI-human collaboration.
- 3The author presents a critical perspective on AI's practical utility, specifically questioning how much AI tasks actually save human time and effort, ultimately emphasizing the need for commoditized, local LLMs and on-device AI.
This article, written from the future perspective of March 2026, provides a sober assessment of AI technology after a substantial 40 months since ChatGPT's launch. Rather than mere praise, it frankly exposes the trials, errors, and limitations encountered when using AI tools in real-world development, clearly demonstrating that AI is still a 'tool' requiring human intervention and judgment, not simply 'magic.' In particular, the tendency for AI to produce impressive initial outputs but then 'lose the plot' when projects become complex or require iterative adjustments is a realistic problem that many developers can relate to.
The core context of the article is the impact of generative AI, specifically large language models (LLMs), on development work. As the author notes, LLMs can now shorten a developer's 'research loop,' moving beyond simple code snippet searches to potentially replacing StackOverflow for certain tasks. Furthermore, AI support in integrated development environments like Claude Code suggests an evolution beyond simple chatbot interfaces into a 'new form of computer input and control,' alongside keyboards, mice, and command lines. This implies that AI has the potential to fundamentally change the human-computer interaction paradigm itself, not just serve as a functional assistant.
These developments have significant implications for the industry and startups. First, in terms of developer productivity, AI excels at repetitive tasks and initial scaffolding but still faces limitations in high-level design, complex bug resolution, or long-term architectural planning. Thus, AI will be re-defined as an 'augmenting' assistant rather than a replacement for developers. Second, the 'new form of input' offers immense opportunities for startups to create innovative development tools, from AI-powered IDEs and code editors to even operating systems. Third, the author's yearning for 'local LLMs' and 'commoditized AI' foreshadows the growth of edge AI, on-device AI, or specialized smaller LLM markets that overcome the limitations of cloud-based AI.
For Korean startups, the implications are clear. First, rather than generic AI services, they should focus on developing AI solutions for specific, 'unambiguously good, useful' niche areas within particular industries or development workflows. Examples could include AI coding assistants specialized in specific languages or frameworks, or tools for automated generation of Korean-specific development documentation. Second, there's a need to innovate user interfaces (UI/UX) through AI. Intuitive natural language-based control can maximize development convenience, allowing Korean startups to leverage their strengths in design and user experience by integrating them into AI tool development. Third, investment is needed in cultivating developer talent skilled in 'complex problem-solving' and 'creative architectural design,' areas where AI still falls short. Finally, exploring opportunities to use open-source LLMs or develop lightweight on-device models can address the demands of the domestic market sensitive to data sovereignty and security.
This article offers a realistic perspective on AI, rather than blind optimism, providing crucial insights that startup founders should heed. It clearly establishes that AI is a tool to 'support' rather than 'replace' coding tasks, prompting deep consideration of where that support is most effective. Beyond merely adopting AI, the true opportunity lies in creating tools that help AI 'not lose context' for developers—that is, tools that optimize AI-human collaboration. For instance, startups focusing on efficiently injecting domain knowledge into AI, or on interfaces that help human developers quickly and accurately verify and correct AI responses, could gain significant competitive advantage.
Furthermore, the author's yearning for 'local LLMs' and 'commoditized AI' holds significant implications. This reveals a market need to address the limitations of cloud-based AI services, specifically regarding cost, speed, and data privacy. Korean startups can align with this trend by exploring opportunities in on-device AI, edge computing-based AI solutions, or private LLM solutions fine-tuned with specific company data to enhance security and efficiency. Not just innovation in AI technology itself, but also services and solutions that 'optimize' AI technology for real-world corporate and individual environments will lead the market going forward.
댓글
아직 댓글이 없습니다. 첫 댓글을 남겨보세요.