Why AI Isn’t Ready to Be a Real Coder? AI’s coding evolution hinges on collaboration and trust (AI column)
Interesting. article from 26 In August from IEEE Spectrum about the current state of AI tools, which, according to the authors, are rapidly progressing, but they are far from the bottom autonomy. They are useful as assistants, but can not autonomously solve problems with a large context, complex logic and a long planning horizon. As a result, now we should strive not to replace the developer, but to build effective human interaction.> AI.
Interestingly, this article is based on the March whitepaper.Challenges and Paths Towards AI for Software Engineering", presented on ICML-2025 (Cornell, MIT CSAIL, Stanford, UC Berkeley, UPenn). In which
- A taxonomy of AI tasks in development is given (not only code generation, but also testing, analysis, refactoring, maintenance, etc.);
- Listed bottlenecks: evaluation and benchmarks, effective use of tools, collaboration with people, long-term planning, huge context, semantic understanding of codebases, low-resource languages / rare libraries, API / version updates, high logical complexity;
- Proposed ways forward: better data collection and curation, RL environments for code, rapid adaptation to specific codebases, training models to collaborate with people, semantically informed search/retrival, deep integration with tools and development processes
And in article "Why AI Isn’t Ready to Be a Real CoderIn addition to relying on this article, the ideas of other scientists who shared their thoughts are given.
- Armando Solar-Lezama (MIT CSAIL) He said that the current interfaces and interaction with AI are still far from working with a living colleague.
- Koushik Sen (UC Berkeley) Talk about the difficulties of finding and correcting complex defects (For example, problems with memory security) in large codebases.
- Shreya Kumar (University of Notre Dame) She raised the question of the cost of "prompt engineering": sometimes it's easier to write code than to explain it to AI.
- Abhik Roychoudhury (National University of Singapore) expressed the criticality of capturing user intent when creating software (Architects and analysts usually talk about this in the format of problem space and solution space, as well as functional and non-functional requirements.). Abhik also talked about the role of “agent” AI and the emerging issue of trust.
To summarize, here are the main conclusions of the article. **1. Full autonomy isn't here yet.**Models often “hallucinate” the causes of bugs, give irrelevant fixes, and don’t keep a long context/plan well. You need a mandatory human-in-the-loop. **2. Breakthrough will require not only large models, but the right organization of work.**New interaction interfaces, the ability of the model to express uncertainty and proactively clarify requirements, explicit capture of user intent. 3. Agential and evolutionary approaches provide encouraging signalsBut the issue of trust and verification will remain central. This article mentions AlphaEvolve, which I am talking about. told earlier
#AI #Engineering #Software #Metrics #Devops #DevEx #Architecture