Dario Amodei’s 2025 AI Coding Forecast and Matthew Berman’s Response

In a clip, Anthropic CEO Dario Amodei predicts that AI could write 90% of code in three to six months and essentially all of it within a year. I call the statement striking and think the timeline may be aggressive, but believe the shift could be closer than many people expect. These are forecasts made in 2025, not measurements showing that the outcome occurred.

I start with “vibe coding”: a person describes what they want to an AI coding tool, such as Windsurf or Cursor, then iterates on the result without personally writing every line. I think the approach broadens access to making software, including websites, games, and applications, although a working and dependable result is not guaranteed.

I cite a 2024 report that Google CEO Sundar Pichai said more than 25% of new code at Google was generated by AI and reviewed by human engineers. I suspect much of that could come from completion tools, though the precise breakdown is not established in the video. The company-specific historical figure is not a measure of all code worldwide.

A separate clip shows Mark Zuckerberg predicting that Meta would have an AI capable of acting like a mid-level engineer during 2025. I present this as another industry prediction, not proof that a particular system later achieved that role.

I argue that models gain more practical capability when coding editors and agents let them use tools, inspect environments, and iterate. As an example I discuss Manus searching documentation, creating folders and containerized environments, taking notes, and running code. These are described capabilities, not a controlled evaluation of reliability.

To think about the 90% forecast, I distinguish the volume of code produced by people from code generated by AI. I suggest that AI output could grow much faster because systems can generate text quickly and people are increasingly using them. The share depends on the definition of code and how it is counted; it does not directly measure completed, reviewed, or valuable software.

I also acknowledge limitations. AI-written code is not always correct, and in my own experience larger projects can make adding features and making changes considerably harder. I have seen impressive projects made with natural-language tools, but those examples do not establish that autonomous coding is reliable in general.

I return to Amodei’s wider discussion of work and meaning. In the excerpt, he says society may need a different way to think about usefulness if technology changes what people can do, and rejects the nihilistic conclusion that people would be useless. He says he does not know the solution.

My short-term expectation is that more people will be able to build useful software because they can direct tools without knowing every line. In the medium term, I expect agents to take on more work with less intervention. These are my scenarios, not a measured employment forecast.

I point again to Manus as a glimpse of a longer-running agent that takes a prompt and works for minutes or hours before returning a result. I also say it remains far from perfect, with mistakes, bugs, and final outputs that are not always good. The example suggests a direction to me rather than proving full autonomy.

In the longer term, I speculate that code produced and consumed by AI might no longer need to resemble today’s human-readable programming languages. I imagine symbols or other forms that could be hard for people to interpret. This is a theory about what might happen, not a description of a demonstrated replacement for programming languages.

I mention GibberLink as an example of two AI systems exchanging encoded audio, while noting that it was created by humans and is decodable. I use it to imagine a possible future of less interpretable machine communication, but the example does not show that such a future has arrived or that existing model internals are wholly unknowable.