This week, a developer announced the first phase of rewriting microgpt (a minimalist GPT implementation) in Futhark—a functional data-parallel GPU programming language born in academia and virtually unused in industry. The project itself isn't news, but it points to a signal we must watch: the AI learning path is shifting from "how to call" back to "how to build."
What this is
microgpt is an ultra-minimalist GPT model implementation with a codebase small enough to read end-to-end, designed for teaching rather than production. Futhark is a functional data-parallel language designed specifically for GPU computing, similar to Haskell but far more obscure. The developer is porting the former to the latter, releasing the first part this week. At its core, this is a personal learning project with no commercial intent.
Industry view
A counter-trend is emerging: while most people are busy calling large model APIs, a group of technologists is doing "minimalist replications"—implementing GPT from scratch with the least code possible. karpathy's nanoGPT is a prime example of this approach. Supporters argue that you can only truly understand core concepts like the attention mechanism by building it yourself once; this is the foundational skill of an AI engineer. The opposition is equally clear: this has no direct value for the vast majority of practitioners. Enterprises need people who can select models, debug, and deploy applications, not people who can write a GPT. Time spent reinventing the wheel is better spent learning RAG (Retrieval-Augmented Generation, allowing large models to query external knowledge bases before answering) or Agents (AI autonomously planning and executing multi-step tasks), which are far more practical.
Impact on regular people
For enterprise IT: The project itself has no commercial value, but if someone on the team proactively does this kind of exploration, it indicates they are building deep understanding. These individuals are often faster at debugging and model selection, and we believe they are worth the space. For individual careers: The window for "knowing how to call APIs" is closing. Those who understand the principles will have an increasingly obvious advantage in model selection and anomaly troubleshooting. For the consumer market: No short-term impact—Futhark is unheard of by most programmers and won't appear in any products.