The reason for such a bold prediction is that, with the recent release of a series of groundbreaking AI products, the sense of crisis among programmers has grown increasingly strong. The evolution of AI has led to many tasks in coding being replaceable.

Notably, Claude 3.7, which has once again pushed the boundaries of coding capabilities, and Windsurf’s Wave 4 are prime examples. Previously, OpenAI's o3 model and Cursor, which has become a classic among independent developers, are benchmark products that have led some to exclaim: programmers' jobs are at risk.

It’s worth highlighting that in the latest Agentic Coding Evaluation rankings, Sonnet 3.7 achieved a score of 67%, ranking first among all models in the evaluation of junior developer proficiency. The second and third places, GPT-4.5 and Sonnet 3.5, were also close behind, both exceeding 60%. This demonstrates the formidable programming capabilities of these models.

In reality, the two major models have slightly different focuses in programming. GPT-4.5 excels in tasks involving architecture and cross-system interactions, while Claude 3.7 Sonnet performs better in raw coding and code editing.

Simply stating scores might not intuitively convey the capabilities of these large models, so let’s use a practical example to illustrate.

Previously, the o3 model achieved an Elo score of 2727, corresponding to a ranking of 175th in a human programming test competition. This is akin to a college entrance exam where no one knows the questions in advance, so there’s no possibility of the models having practiced the questions beforehand.

This competition involved a total of 168,076 programmers worldwide, and a ranking of 175th places o3 in the top 0.1% of participants (1 - 175/168,076 = 99.9%). In other words, o3 has already outperformed 99.9% of programmers in the world in programming competitions, and Claude 3.7 Sonnet is likely to perform even better.

Not only has AI coding achieved a significant leap at the model level, but there have also been major upgrades in product development.

ByteDance's overseas version of Trae has integrated international large models like Claude-3.7-Sonnet and GPT-4o, and also possesses IDE capabilities. The goal of enabling even beginners to code is nearly within reach, further lowering the barrier to entry for programmers.

Most importantly, the overseas version of Trae is completely free, allowing users to access all its features at no cost. In contrast, Cursor requires a paid subscription priced at $20 per month.

A user abroad utilized Trae and its built-in Claude 3.7 to automatically generate code for a 3D animated globe with just a design sketch and a super simple prompt.

Some have even earned real profits from this. A tech guru overseas named Pieter Levels developed a game entirely using AI in just 3 hours. Within 13 days of its launch, the game had already earned $67,000, equivalent to nearly 500,000 RMB. Even Elon Musk gave it a thumbs up!

The capabilities of large models in coding are rapidly improving and approaching human-level performance in several aspects:

Sonnet 3.5 could output 200 lines of code, while Sonnet 3.7 can now output 1,000 to 1,500 lines, making it the first model capable of reliably generating code spanning thousands of lines.

In the future, a lot of tedious and repetitive programming won't require typing line by line, and many entry-level programmer positions will be replaced. Practicing Leetcode will also lose its significance. A friend who used to rely on Sider for coding assistance, spending over a thousand yuan annually, no longer needs it and now directly uses large models instead.

Another critical capability is Action Scaling, which enables continuous function calls and tool use, iterating based on environmental feedback until an open-ended problem is solved. This evolutionary potential should not be underestimated in the future.

Of course, some friends ask: Can AI discuss requirements with clients? Can AI argue with PMs? Can AI take the blame?

After the release of Claude 3.7, one exclaimed that his job might be at risk. His outsourcing company has already laid off some employees over the past year due to AI and other complex factors.

It’s undeniable that AI isn’t quite there yet, but products like Devin are already demonstrating certain AI Agent capabilities. Their ability to understand and execute tasks is improving rapidly, and the era where everyone can be a product manager is on the horizon.