The Conversation the Fediverse Refuses to Have

When Twitter got taken over by Musk a couple of years back, something akin to an intellectual exodus of sorts happened, and lots of communities left Twitter for other, less ideologically captured alternative platforms that aligned more closely with how they perceive the world. Don't get me wrong here. Many parts of the platform and the people driving it are what can only be called fascist in nature, and I fully agree that resistance is the only correct answer, even if we might disagree on the exact form this resistance should take. Today Bluesky is often described as the leftist form of Twitter and is still perceived as the stronghold of anti-AI ideologies, but people on the Fediverse are at least equally opposed to it. What makes this particularly interesting is the gap in technical literacy between the two. Bluesky has a rather broad user base, while the ActivityPub ecosystem, Mastodon and its sibling platforms, are home to some of the most technically capable people across the broader internet. People doing programming language research or performance-sensitive work in game development, the kind of deeply principled craftspeople whose departure from Twitter marked the end of what once made it genuinely valuable. You would expect this level of technical depth to produce a more nuanced engagement with LLMs. Instead, the criticism voiced across both platforms is largely indistinguishable, and the very people you would expect to apply careful technical analysis are relying on ideological reflexes instead.

While I agree with lots of the criticism about the negative externalities and the way this technology is forcefully being pushed onto our industry by people in roles of power, I think dismissing it in its entirety solely on ideological grounds is the wrong answer and, with regards to the goals of the people that voice such criticism, a quite unproductive one at that. While much has been said on the purported merits of LLMs and their value as tools in software engineering, there is another angle that has not yet been addressed enough: how LLMs democratize access to technical education.

It is true that one can use LLMs in ways that stand in opposition to the principles of engineering. But even in their current state, they are a remarkably effective tool for studying technical fields. The usual caveats regarding their probabilistic output still apply, but this does not invalidate everything else that they bring to the table, and this is where the criticism voiced by these technically able people breaks down. Using natural language to communicate with an LLM can feel close to talking through a more knowledgeable version of oneself. The interactive nature and the targeted feedback that they provide are not replaceable by more traditional and established forms of media like a textbook. To the contrary, the two work rather well together. Being able to ask questions about the contents of a textbook without end enables everyone, regardless of their background, to study technical topics in depth.

Unfortunately, that this aligns with the values of accessibility and equality so central to these communities seems to go entirely unexamined. The most obvious objection here is of economic nature, i.e. cost. But open-source models lag behind their commercial counterparts by roughly half a year at most, and that gap is narrowing. Advances in optimization techniques paired with increasingly capable consumer hardware mean that running a highly competent model locally is something people are already doing today. The economic barrier, while genuine, is eroding faster than most critics seem to acknowledge.