Greetings from a flight to my littlest brother’s bachelor party. Hopefully I will not talk about AI at all this weekend.

I re-started this newsletter in part because of a discussion with a friend about how he thought use of LLMs was going to impact open source tooling—something I wanted to understand more deeply. This layered on top of a lot of my longstanding concerns about the economic and regulatory dynamics of open. 

Increasingly, though, I’m wondering to what extent LLMs are going to change open’s social dynamics. Not the obvious ones like “are there too many pull requests”, but deeper ones. A couple of things in particular are nudging me in this direction this week.

• Sycophancy reducing collaboration: Some interesting new research suggests that one effect of sycophantic LLMs is that it makes people worse at collaborating. That’s got very scary implications for all of society, of course, but it seems particularly likely to afflict open. Full paper; thread:

• Skills gap? In a long piece on how he is using agents right now, Simon Willison made a throw-away comment that LLM coding tools are now very good at understanding complex code bases. Just a few hours earlier, I’d read an email from a new maintainer lamenting that “as you know, learning a new codebase is hard”. If a (laudable!) commitment to purely open tools puts smart, hard-working open developers at a disadvantage… that would be a shame.

This is going to be the slowest-developing, and hardest-to-predict, type of change in the next few years. But of course also the most important.

Open experience & culture

How LLMs are changing the culture and lived experience of open source communities

• Dan Davies has an interesting thought experiment on what we do and don’t trust LLMs to do, with a focus on middle managers. I wouldn’t normally compare open source developers to middle managers, but FOSS often asks developers to wear many different hats—code, marketing, documentation, funding, etc. “Which of those do even the most LLM-forward developers trust LLMs to do” would be a very interesting survey to run.
• LLM translation is a “wicked problem” for smaller languages on Wikipedia: This is a well-researched article on how machine translation can be very, very problematic for smaller languages on Wikipedia—in part because people think they’re helping by translating stuff… but the translations are very bad. Meta point: the impact of LLMs on “open” is going to vary widely; it’s not even consistent between Wikipedia’s languages, much less open more broadly.

Infrastructure for open development

How LLMs are changing the technical infrastructure and tooling of open development

• Symposium, a tool for agent-centric coding workflows: This is a cool-looking, open source, experiment in what UX might look like for organizing coding around AI agents. No idea if this is The One, but it’s pretty clear we’re in a time where there could be a potential paradigm shift akin to the jump from text editors to IDEs—and it’ll be interesting to see if open is a leader, fast follower, or completely misses the boat.
• Curl gets an AI assistant: Daniel Stenberg of curl has been a pretty aggressive critic of LLM-generated pull requests, and with good reasons. But someone used curl’s code as a testbed for LLM-aided security vetting… and so far Daniel has integrated almost 50 changes as a result. That’s a real acid test, passed.
• The Return of Micropayments: Cloudflare and Coinbase, among others, are trying to bring micropayments back, in part as a response to LLM scraping. Will be interesting to see both (1) if this works in the non-code context and (2) if it does, what that might signal for reuse of code.

Access and barriers

How AI is changing who can participate in open source and what barriers remain

• Some recently-published analysis suggests that local models are only currently lagging frontier models by a year or so. (Coincidentally, here’s an announcement of “frontier-grade performance on everyday devices"—under the Linux Foundation’s model weights license.) Not clear that this would have huge impact on FOSS, exactly, but small, powerful, local models would have wide-ranging impacts on exactly the power differentials FOSS often is aiming to counteract.
• “Code written by academics” has, to be generous, a mixed reputation, and academics would be the first to admit that the code is often a side-effect of their real skills and passions. Google prototyped using LLMs to help academics write code across a variety of domains, and had good results (though of course take those with a grain of salt). Interesting to see if this will help (by improving quality) or hurt (by increasing volume) the reproduction crisis in science, and how it changes the intersection of academic code and open code. Related: some LLMs are now good enough to do replications of academic data analysis, for values of replication that include the question “is it really a replication if it is just as opaque to the human reviewer as the original paper”?

Power and Centralization

• The EU thinks it can get to digital sovereignty by… focusing on the top of the stack, AI. Suffice to say I am skeptical; if the bottom of the stack is still controlled by the US or China, you’re not going to get control of the top of the stack. Worse, I worry that this is going to be a recurring theme: governments and investors distracted from FOSS infrastructure by the potential impact of AI. 
• In related news:

Outside the bubble

• The UN says that everyone in Asia/Pacific (4+B people) will have access to electricity by 2030. But this is just a start: billions don’t have access to WEIRD economy levels of electricity, and any discussion of the future of global power usage that doesn’t discuss this is an incomplete one. (via Fix The News)