Small language models, even the specialized ones, running locally on personal computers may after all not really take over. There is a caveat to that dream, and there are lessons for us from the software development era of Microsoft and Google.

Software runs in a partially observable and changing environment. We discovered that decoupling resource allocation from application is a necessary operating systems infrastructure to keep fixed costs on application software low. But that means for application developers, there is a platform police to run past. Users' affinity to spend time updating software is never as much as developers want. Building desktop software means fewer and fewer deployments. Leading to slower bug discovery, security, and performance issues, slower iterations from experience. Backward compatibility becomes paramount and hard to achieve in runtime that developers can not see. All those mean higher development costs.

The application and OS need another intermediate platform. A good and stable web browser on a common standard solve platform integrations, security, and packaging. OS version compatibility, memory optimization, canvas painting are completely abstracted out from web applications. Events in browsers are first class objects without really taking any additional complexity. It’s significantly harder to develop for macOS compared to other apps on the Google Chrome browser.

Decentralized access matters more than decentralized execution.  There is a great operational advantage in simply running a web server and connecting to a remote interface. It’s harder to run workloads and computations going beyond the personal computer's capabilities. Sometimes, reliable technology requires centralization, such as in instant messaging. Ironically, partial centralization is practically needed to support security in email. Its been harder to ask people to pay for the software before significantly proving its worth. That led to a pricing model where cheaper services were increasingly available on the internet. This is one of those times, access mattered more than privacy.

AI outperforms humans in interdisciplinary problems. There seems to be fundamental and practical limitations around how much humans can process information. Humans exclusively pick their professions. Most people are either lawyers, or doctors. Each of these take about ten years of training. Few try to learn another field after another but their past expertise fades away. AI is uniquely positioned to solve problems that are at the intersection of two fields that do not communicate, and those exist. In healthcare simply maintaining a complete patient history is still complicated, and conditions that require different specializations take longer time for accurate diagnosis.

People always want a better toy. We now have more compute on watches than my personal computer from high school, and the computer on the Apollo mission. I have to make a philosophical leap to make the argument. People will never run out of work. The tree of imaginative and physical abstractions of nature always grows. There’s an unbelievable compositional ability among the objects in nature and the abstract, on which the market trades forever. Unless some shit happens. All this means, as long as consumers upgrade hardware slower than companies can pool significantly larger compute and offer a better toy, it’s building locally is not an obvious choice.