Dev AIOriginally published on Zhihu, 2025-02-13
I recently saw the 2024 Products of the Year list on Product Hunt. For those unfamiliar with it, Product Hunt is a website that tracks the latest software in the tech world. Looking across the lists from the past few years, you can clearly see the trend shifting from platforms like low-code tools toward GenAI-related software. As a heavy user, I want to briefly reflect on how several of this year's winning products have impacted my work.
Products of the Year
The overall winner is undoubtedly Cursor, a product that set the record for the fastest growth from 1M to 100M ARR. When I recommended it earlier, many people felt it was pretty much the same as Copilot. Although Copilot is gradually catching up with Cursor's features, Cursor still maintains the lead in terms of smoothness and integration. From a productivity standpoint, I shifted from initially preferring the fully automated agent / composer mode to the semi-automated chat / inline query mode, realizing that implementing complex requirements ultimately still requires human supervision.
Supabase is an open-source Backend-as-a-Service platform. Its biggest advantage is convenience, allowing me to focus on bringing product ideas to life rather than building backend infrastructure. In the AI era, it is particularly well-suited for using AI to rapidly create product prototypes and put them on the market for validation. Of course, it also has limitations: the database table capacity on the free tier is relatively small, self-hosting is quite troublesome, requirements that go beyond its scope become complicated, and access from within China is restricted. However, considering how well it pairs with AI, I think it is still a tool worth including in my tech stack.
Finally, there is o1. Although subsequent models like deepseek r1 are also excellent, o1 undoubtedly pioneered a new paradigm for LLMs. It is better suited for solving complex problems, which has also made my workflow clearer: I use models like Claude for text and multimodal enhancement, and o1/r1 to handle complex architecture design and coding. I look forward to seeing R1 on the AI model rankings in '25.
Current State of AI Adoption
Combined with the Economic Index released by Anthropic, AI has currently become mainstream primarily in programming-related tasks. For a fairly long time to come, LLMs will likely be used mainly to assist rather than automate programming and design. Looking at the Products of the Year over the years, productivity tools have always been heavily favored, while the data also shows that gaps in AI adoption still exist across different industries.
The Shift in Core Competitiveness
Reflection
ReflectionIf LLMs upgrade again (right now I can already use Bolt.new + Cursor to quickly complete tasks that were previously unimaginable), where will core competitiveness shift?
Perhaps core competitiveness was never "knowing how to write code". Just as IT in financial firms is viewed as a cost center rather than an asset, could "software development" be a temporary artifact of an era over the past few decades when technology was less developed? If development skills become as fundamental as handwriting, the focus will shift to how to write good stories and how to make better pens.
Future directions might be:
- Exploring broadly to find good ideas (cross-disciplinary) and telling good stories
- Thinking about how to build software that improves efficiency, crafting a better pen to write with