AGI coding predictions for 2025

In 2024, there was much excitement about AI code assistance, and also a realization that it poses major risks. Suddenly non-experts could act like expert developers and ask AI to generate a Python script, for example. Without understanding the nuances they simply didn’t know what they didn't know. Often these users failed to ask the right questions, and the AI-generated code introduced a litany of bugs. In 2025, there will be an urgent and increasing need to have a better developer workflow for AI-assisted coding. Whether writing a script or creating a fully-fledged project, there is a need for some form of top-down oversight into how development environments are being used and created, how developers and data scientists are interacting with the tooling, and how AI interacts with the system. What kind of AI is used? What tools do users have running? It is difficult to know exactly how developers and data science teams interact with those tools and how those tools interact with the code, because the environments are typically on a user’s machine.

AI and AGI coding predictions for 2025

Next year, it will become necessary to address this. One way is to create development environments that are somewhat standardized, with approved tools or APIs that the tools automatically connect to, in order to decrease the number of bugs and increase test coverage or quality. For example, tools like DevPod can create and control environments, letting administrators add components including AI-related tools, linters, checkers — like Cursor, the AI code editor. It becomes easier to control the environments by adding internal mechanisms to validate code quality and test for issues, without stripping developers of the ability to run AI. I anticipate that it will be impossible to develop without AI even in just the next few months, because developer velocity will drop significantly. Given that, it will be interesting to see how teams adapt and build standardized workflows, with administrators controlling certain layers of the stack and adding scripts or tools to protect code quality. 

Platform engineering becomes a buffer for AI adoption

Platform engineering will continue to rise in 2025, as an emerging discipline that fits the niche of needing a layer between infrastructure — whether cloud or internal services — and the developer. As platform engineers further realize their role of taking things from the outside world and molding them in ways that make sense for engineers, this will extend into acting as a buffer for AI and other new technologies. We will increasingly see AI services and infrastructure fall under the purview of platform engineering; already, these teams are questioning how to deal with colleagues using generative AI, how to expose AI to developers, and how to reduce risk. 

Because platform teams are responsible for designing infrastructure for containerized workloads, adding and maintaining internal tools and scripts, 2025 will see their natural progression into a means of standardizing and making AI initiatives safer and more effective. Currently, there are really no teams in place to take on this task. We do not see a preponderance of dedicated “AI engineering” teams, so I believe platform engineers are the most natural candidate to expand into this use case. Companies will almost certainly create additional AI compliance roles, but for engineering initiatives we need a more structured approach. Thus, platform teams will see an influx of questions around what to do with AI — it will be their challenge to find solutions.

No, AGI is not imminent - It’s not even defined

This year brought constant speculation around artificial general intelligence (AGI), with some leaders claiming it will arrive much sooner than we might think — maybe even the next couple years. In 2025, this discussion will settle back down to Earth as the industry realizes that even the definition of AGI is so vague that people are essentially inventing it as they go. We will see models continue to improve, and improve greatly, but the big breakthrough is a ways ahead. I do not see in the near future a model that I can ask to code something for me, and it returns an output that matches or exceeds what I would have done myself. I am not a doomer; I would not say AGI won’t happen, but we are far from unleashing novice programmers on AI and being able to trust the output.  

AI is a great tool. It helps in many ways. But it also introduces friction if we want too much from it. For the time being, it remains a tool best suited to supplement the work of experts, who know what they are doing and know the right questions to ask.