Developers are increasingly relying on large language models (LLMs) for everyday computing tasks such as fixing bugs, explaining code and automating text-processing tasks like filtering logs.
However, it’s not as simple as entering or submitting a question and relying on the model to give you the answer. While humans easily understand these tasks and know exactly what they want, it is difficult to translate them into rigid computer code.
The cloud dilemma
As standard programming is often not up to the task, developers often use AI to handle jobs that are difficult to express as traditional rules. However, many powerful AI models are too massive to sit inside a laptop or a phone, so developers have to send data over the internet to paid cloud services. But doing this risks exposing private data, can break software if the AI provider updates the model and is expensive.
This headache may soon be over, though. A team of researchers from the University of Waterloo, Cornell University and Harvard University has developed a new paradigm called Program-as-Weights (PAW) that simplifies the task.
Instead of sending every single user request to a massive AI model over and over, PAW introduces a “compile once, run locally” strategy. It uses AI to act as a one-time builder that converts the developer’s plain-English instructions into a tiny custom AI add-on that can later be reused for the same task.
This can then be downloaded and plugged into a miniature AI model that lives inside a user’s laptop or phone. It also works offline. “PAW reframes the foundation model from a per-input problem solver into a tool builder,” the researchers wrote in their paper.
To see whether PAW works in practice, the researchers tested it on FuzzyBench, a database containing 10 million examples of programming tasks such as filtering logs and repairing broken JSON files. They compared a miniature AI model running a PAW tool against Qwen3-32B, an AI model with more than 50 times as many parameters. It achieved 73.78% accuracy on the researchers’ benchmark of fuzzy programming tasks, compared with the larger model’s 68.7%.
PAW is also fast. It processes about 30 tokens a second on a MacBook M3 using a quantized 430-megabyte interpreter, which means it can run efficiently on standard consumer hardware without a server.
Accessing the code
Developers who want to use PAW can do so, as the researchers released the code alongside their paper. Explaining their vision for the future, the research team stated, “We hope Program-as-Weights contributes to a future in which small LMs serve as the runtime, where large models compile and small models execute.” https://techxplore.com/news/2026-07-ai-automate-everyday-tasks.html
Recent Comments