AI guru Karpathy is anxious: As a programmer, I've never felt so left behind.

The year-end holiday is the perfect time for reflection and summing up. However, when programmers really think about it, they might feel something is off.

Just recently, Andrej Karpathy posted a message on X, which struck a chord with tens of thousands of programmers and industry practitioners, sparking intense discussions.

Karpathy admitted, "I've never felt so left behind as a programmer as I do now."

He pointed out that the programming profession is being completely restructured. Programmers are writing less code and instead are more involved in connecting various tools. If one can make good use of the new things that have emerged in the past year or so, they can become ten times more powerful. Otherwise, failing to keep up will lead to skills anxiety.

Now, there is a new programmable abstraction layer to master, including agents, subagents, prompts, context, memory, patterns, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integration, etc.

In addition, one needs to establish a comprehensive mental model to understand the advantages and disadvantages of entities (referring to AI models) that are essentially random, error-prone, difficult to understand, and constantly changing. These entities have suddenly intertwined with traditional excellent engineering practices.

Using Karpathy's analogy, it's like a powerful alien tool has been distributed to everyone without an instruction manual. Everyone has to figure out how to use it on their own, and this transformation has brought an impact like a "magnitude 9 earthquake" to the entire industry.

In a nutshell, roll up your sleeves and work hard. Don't fall behind.

Once these words were spoken, they quickly received over 22,000 likes, more than 3,000 retweets, and 3.6 million views. Many developers expressed similar feelings in the comment section.

Even veterans are relearning

Senior engineer Boris Cherny said, "I have this feeling every week. Sometimes I start to manually handle a problem and then have to remind myself: Claude should be able to handle this."

He also gave a specific example. Recently, when debugging a memory leak in Claude Code, he habitually used the old method: connecting the profiler, using the application, pausing the profiler, and manually checking the heap allocation.

His colleague, on the other hand, directly asked Claude to generate a heap dump file and then asked it to read the file to find objects that shouldn't remain. Claude got it right on the first try and directly submitted a PR.

"This kind of thing happens almost every week." Boris noticed an interesting phenomenon. To some extent, new colleagues, especially fresh graduates, can use the models most effectively because they don't have a lot of preconceived notions about what the models "can and can't do."

He said that he needs to invest a lot of mental effort every one or two months to readjust his perception of the models' capabilities because the models are constantly improving in coding and engineering. Last month, as an engineer, he didn't open the IDE at all for the first time. He wrote about 200 PRs with Opus 4.5, and every line of code was generated by AI.

"Software engineering is undergoing a fundamental transformation. Even for us early adopters and practitioners, the most difficult part is still constantly readjusting our expectations, and this is just the beginning."

Karpathy gave an analogy to explain this feeling. It's like you're aiming an AI around, and it shoots projectiles. Sometimes it even misfires, but occasionally when you hold it at just the right angle, a powerful laser beam suddenly bursts out and solves your problem instantly.

The implication is that the AI tool is very powerful but unstable. It's not as controllable as traditional programming. You have to keep trying and making mistakes. Most of the time, it's just small efforts or failures, but once you find the right method, it can bring exponential productivity improvements.

Igor Babuschkin, the co-founder of X, specifically praised the competitor Claude Opus 4.5 in the comment section. Karpathy replied that AI is evolving so fast that those who haven't kept up in the past 30 days already have outdated views.

David Galbraith, a technology expert and venture capitalist, said, "I spent three months this summer, working around the clock to learn how to use AI coding agents to deliver truly high-quality products, not those casually coded junk. This is the best thing I've ever done in my career."

X blogger @omarsar0 has a more optimistic and relaxed attitude. He believes that the fact that code is becoming sparser and AI is advancing rapidly doesn't bother him because he doesn't see it as a "competition." Instead, the field is now completely open, and creative solutions and workflows can come from anyone, anywhere. This transformation is not limited to coding but also happening in research and other knowledge-intensive fields. He suggests that everyone should not be anxious. Spend two hours a day playing with the tools, do more experiments, share more, focus on how to provide good context to AI, and then work hard on building projects.

"Build a Large Language Model From Scratch" also takes it easy. People's general anxiety about being "left behind" in skills usually comes from trying to do too many things at the same time instead of delving deep into certain things. For example, some people learn several programming languages instead of specializing in one or two. Or they try to keep up with research literature in multiple fields/subfields at the same time... This is not necessarily a bad thing, but it does put a lot of pressure on people.

Even X blogger @samswoora sighed that "the profession of software engineer is about to be over." "It might be in 5 years or 10 years, but we can all feel that the end has begun."

Well-known blogger Yuchen Jin believes that artificial intelligence has not replaced programmers. It has replaced programming languages.

Opposition from the traditional camp

However, some people hold an opposing view. The representative figure is Rob Pike, the co-founder of the Go language, a Unix veteran, and a pioneer of minimalism and high-quality engineering.

Rob Pike received a holiday thank-you email automatically generated by Claude Opus 4.5. The email praised his promotion of simple and powerful software design and his far - reaching contributions to Go, Plan 9, UTF - 8, and Unix. But Rob was furious and directly lashed out on X: "You AI companies are wasting huge amounts of money on creating toxic and non - recyclable hardware while destroying society, and yet you let the machine hypocritically thank me for pursuing simple software?"

Rob Pike's anger hit a chord with many programmers' complex emotions towards the AI hype.

Some netizens fully understand and support Rob's attitude. This kind of low - quality code and spam generated by AI in bulk is really offensive, especially for old - school geeks like Rob who pursue extreme simplicity and pure engineering. It's simply an insult.

Anyway, we have to admit that the development and progress of AI in the past two years have been amazing. Although there has been a lot of discussion among AI experts about the end of Scaling Laws from last year to this year, the fierce competition among technology companies has not slowed down the development of AI technology but instead accelerated it.

According to data from Epoch AI, the Epoch Capabilities Index (ECI, a comprehensive indicator to measure the general capabilities of AI) has almost doubled in growth rate in the past two years compared with the previous two years. In April 2024, it even accelerated by 90%.

The actual exponential growth has even exceeded the original expectations, and this growth momentum is likely to continue until 2026.

It's hard to imagine what level AI will develop to in 2026. What are your predictions for the development of AI in 2026? Welcome to share your views in the comment section.

Reference links:

https://x.com/karpathy/status/2004607146781278521?s=20

https://x.com/bcherny/status/2004626064187031831

https://x.com/daveg/status/2004661204296589480?s=20

https://x.com/nixcraft/status/2004644277859889181?s=20

https://epoch.ai/data-insights/ai-capabilities-progress-has-sped-up

This article is from the WeChat official account "MachineHeart" (ID: almosthuman2014). Author: Someone concerned about AI. Reposted by 36Kr with authorization.