Among my peers who study the cultures and economies surrounding technology, I often find myself more tech-positive than most. On the whole I like computers, like the internet, and like in general the emergent human creativity that these things have enabled and revealed. A source of my appreciation is that I expend a lot of effort on the practice of “slow computing,” which is to say, choosing tools and technical communities that I can feel in right relationship with. I tend to approach new socio-technologies (like gig economies, blockchains, and social media) not from a perspective of why is this so bad? so much as asking how could we do this right?
I have taken the same approach to generative AI. My first inclination is a coincident sense of awe that we now have talking computers that can write, code, hear, infer, and cheat (amazing!), along with dread about how our reigning economic system has channeled this fearsome technology into psychopathic business models and planet-threatening energy consumption. I’ve proposed better ownership structures for existing AI companies, imagined new ones, and mapped opportunities for democratic governance. All of this has been guided by my own experiments with the technology. I have often found that the better I understand how a technology ticks, the more clearly I can see opportunities for organizing it differently.
Here, I’ll report on my anecdotal findings over the past year or so, and especially this summer, as I’ve been blogging pretty frequently on the topic. Forgive the length and unnecessary detail; this is more an act of note-taking for myself than an unmissable decree for anyone else. The results will probably not come as a big surprise. But the saga might be of interest to you nonetheless.
My stack
First, a few notes on what I’m using, to provide some context. I have tried to choose tools that are appropriate for the task, that are as privacy-preserving as possible, that encourage me to be intentional about any interaction, and that are not too cumbersome to use.
Local chatbots
For lots of small tasks, like foreign-language practice, basic tech support, and offline informational queries, I use a local model on my computer or phone. On the computer, I mainly use Google’s Gemma 3 or 4 models (but also stuff from Mistral and Alibaba), and Gemma 2 on the phone. Because the models are hosted on the machine, they don’t send any data to the companies that built them. On the computer I use Ollama to run the models and Simon Willison’s LLM to interact with them in the terminal, where I have a few simple aliases with my preferred configurations. Page Assist is also a nice way to access Ollama-hosted models in a GUI within Firefox.
On the phone I’ve played with various apps that run local models, mostly built
around llama.cpp, but there isn’t one I really like yet.
Hybrid agents
This has supposedly been the year of “agentic AI,” so of course I had to set up
some agents. I have a little Raspberry Pi computer at home running Hermes Agent from Nous Research, a company focused on supporting decentralized AI. (Hermes is sort of like the infamous OpenClaw, but with a few differences that I like.) I interact with them via a self-hosted Matrix chat server.
I don’t have the hardware to run models capable enough to run agents, so I use Ollama’s cloud service, which provides open-weights models on a promise of no data retention. This allows me to try out a variety of options, and I frequently swap models on the same agent.
With this approach, the runtime interactions go to a corporate cloud, but the only persistent data should remain on my own hardware. I have set up several agents for specific projects or ongoing tasks, so each agent develops its own context and skills.
Sometimes I’ve used the pi coding agent. Its bare-bones design with a large plugin library has been a helpful way to learn some of the building blocks of a usable harness.
Web-based chatbots
I also use some conventional chatbots on websites, usually for questions separate from my research projects that would benefit from web searches. In particular, I like DuckDuckGo’s Duck.ai, which provides access to various services (such as Claude, ChatGPT, and Mistral) with more privacy protection than those companies typically offer. Duck.ai also makes it easy to compare the outputs on the same prompt from different services. For some reason, I have found my campus-provided Gemini chat especially useful for figuring out tech problems, like fixing broken Linux configuration files and installing alternative operating systems on phones. The campus contract supposedly prohibits using interactions as training data.
So that’s what I’m working with. Now I’ll turn to some of the various uses to which I’ve attempted to put my little toolbox.
Coding
I’m one of those people for whom the ah-ha moment with generative AI was not getting the original ChatGPT to write sonnets—it was when I got a chatbot to turn a little program I’d written in Python into JavaScript. It worked the first time. Then the ah-ha happened again when Claude Code came out, introducing me to the eerie experience of a bot prowling through your files and making changes in multiple places to follow an instruction.
As a coder, I’ve been only ever good enough to be dangerous, and I don’t know the first thing about most modern languages and frameworks. I do know enough basics to know what to ask for and to think through a problem with some help. I have hand-built several apps over the years that have been moderately successful proofs of concept. CommunityRule and Modpol have influenced the thinking of developers more skilled than I am. In both cases, I also had the chance to work with collaborators who contributed mightily to both the design and implementation. Although they always worked for far under market rates, I was able to raise some money to pay them.
With coding agents, I’ve been able to build several proof-of-concept-level apps for a fraction of what I’ve paid collaborators before—a few dozen or hundred dollars, say, rather than a few thousand. That has made it a lot easier to try things, and the results look and function pretty well. But none of these have gained much of any user traction. One reason might be because of what AI has done to the attention economy: given the onslaught of vibe-coded slop, I’ve been more reluctant to promote them like I used to and join the horde. Perhaps potential users are less willing to try experimental apps out because now there are so many of them.
I suspect the bigger reason, however, is simply that, without the active collaboration of other people, the results aren’t as interesting. I can’t verify this. But when I’m approached by founders who want my support for their new projects, one of the biggest red flags for me is the lack of even one co-founder. If you can’t convince a single other human being that a project is worth investing time and energy in, that is a signal. And it is also a signal if you don’t bother trying.
My most successful vibe-coding has been the small scale stuff: simple scripts that only I use, or automations for mindless administrative processes at work.
Verdict: While AI allows for far easier and cheaper rapid prototyping, that very fact risks producing lower quality outputs, less validated and improved by human collaboration. I conclude that I should aim to do less AI coding rather than more, and make sure to involve human collaborators at the design stage before turning to artificial assistance.
Research
Like a lot of researchers, I have been curious about whether these tools trained on already published knowledge can help in the process of crafting new ideas.
I used generative AI to assist in several rote textual analysis projects, including one with AI-generated synthetic data; that one seemed to confirm a hypothesis of mine, which was satisfying (and without knowing that it was my hypothesis, so this was probably not sycophancy). In another case, I ended up finding the textual analysis to be garbage, so I did it myself, but continued using the AI to assist in analyzing the resulting data.
I found these potentially useful. The trouble is that the methods of analysis started extending beyond my expertise. While at first this felt like a fun centaur situation, increasingly I became uncomfortable at my lack of ability to fully explain the results. Not good. As soon as you’re outside your expertise, you lose the ability to judge the quality of outputs. And as my student Kadallah Burrowes found in their experiments, it’s a lot easier to get a chatbot to say something than to say something accurate.
I have also used coding agents to assist in editing interview transcripts. All of the Protocol Oral History Project interviews were edited that way. I would not have been able to have and publish so many wonderful conversations otherwise. I think the results were decent, though several interviewees commented that they didn’t sound like themselves. At the same time, when I read a faithful transcript of myself speaking, I refuse to believe I actually sound that idiotic, so I’m not sure this can be chalked up to the AI. I always did a close manual edit before asking the interviewee to do the same, and I felt the results were true to what the other person said and their way of speaking.
Another practice I have found useful is having my project-focused agents produce weekly digests on news and research related to the project they’re assigned to. This has the dual purpose of informing me and building up their working knowledge. These digests have pointed me to a lot of useful information, and they do a surprisingly good job of matching what they find with gaps in my own thinking. It is a major step up from Google Alerts, to say the least.
A note on labor: I have paid transcribers in the past, back when I was a journalist, but I don’t think it is the most fulfilling work, so I don’t feel too bad about replacing it. And while I do have paid research assistants, I encourage them to develop their own projects alongside mine rather than simply helping me with my books, so these agents fill a role that people did not previously hold.
Verdict: There are some promising, limited use-cases, but it’s important to avoid venturing past one’s expertise so results remain explainable. I still have my research agents running.
Writing
I have mostly avoided any AI use for writing. Writing is my primary craft, and I also happen to be emotionally addicted to it. It helps me sharpen my thinking. I will not give this up no matter how good chatbots get at replacing me.
Still, I want to be a sport and take seriously what AI can do here. I’ve tried a few times to get it to write something for me based on notes, and the results were horrific. But I did find one process that struck me as having potential.
I learned recently that I had an longer-than-usual abstract for a conference due in a day or two. It was a weekend, when I am usually with my family most of the time, so I had to be quick. First, while doing an errand in the car, I had several conversations with Duck.ai’s voice chat about channeling part of an ongoing project into the abstract. Despite its oppressive sycophancy, I felt that I managed to get some decent ideas out, and that it made decent suggestions. When I got home, I exported the transcripts and asked a local model on my computer to turn them into an abstract of the required length. That took a while on my aged desktop. Later, when I had some time alone, I edited it to the point of mostly rewriting it. I didn’t worry that I was submitting something other than my work; the source material was transcripts of me speaking, and I touched every sentence with rewrites. It wasn’t ultimately clear whether the bot saved any time at all, but maybe. And the idea-generation process seemed helpful.
The result? The abstract was rejected. This is an n of one, of course, but in my career rejections from conferences have been more the exception than the rule. I think that, despite my best efforts to be intentional at each stage of the process, working with the bot let me get away with thinking inferior to what I’d do on my own or with other humans.
One way AI has definitely been a gift: It has facilitated my return to drafting by hand. Because of the ease with which some models can transcribe my handwriting, I’ve started hand-writing more (on paper and PineNote) than I have in decades.
Verdict: AI writing is not my thing. If it’s yours, I confess I’m not very interested in reading what your bot writes. It is a missed opportunity to think for yourself and probably a waste of readers’ time. But by transcribing handwriting, it can have the surprising benefit of de-digitizing the drafting process.
Teaching
With all the people promising that AI will wonderfully revolutionize education, this summer I tried to imagine a version of one of my two classes this semester that is AI-ed to the max: a class agent, no restrictions on AI use, all kinds of bot interactions during lectures, the works. After all, my classes are on tech startup culture and techno-futurism. But as I revamped my syllabi for the semester, I couldn’t bring myself to do it. I just couldn’t imagine a scenario where the learning experience would be better, and more likely the students would lose a chance to do their own reading, writing, and thinking. My experiment was not worth them having a worse learning experience. I ended up with a pretty blah syllabus statement, which reflects my fatalism about the fact that I probably can’t control what my students do with AI anyway:
Since this is a skills-building and knowledge-integrating course, generative AI should not be the primary means of producing anything turned in as original work.
Maybe the trouble here is a lack of imagination on my part. But I really tried, and I’m pretty handy with what the tech is capable of. I have come to think that the best way to respond to the rise of AI in education is to double down on what is not AI-able: teacher-student relationships, peer accountability, and passion for the material.
I did use some AI assistance to create the formulas for a spreadsheet-based system that students will use to do peer evaluations, but the system itself is fully deterministic and so does not involve the use of AI for interpretation or evaluation.
Verdict: When students come to a classroom to learn from teachers, AI is probably a distraction from what that space is for.
Home
Is there a place for AI at home? I sometimes ask it for quick help with cooking, but I’m terrified that most models’ surprising wrongness about guitar chords will extend to a proportion of ingredients. It’s almost certainly better to consult one of the many beautiful cookbooks on our kitchen bookshelf. My home has no internet-connected “smart” devices, because they’re mostly dumb cybersecurity nightmares, so I can’t use AI to turn on lights or whatever. I’ve seen some passably clever AI-ed party invitations—no shade there.
I have set up an agent to send daily digests with the weather, local events, and liturgical calendar stuff, and that has been nice but not really indispensable. As someone not usually inclined to seek that kind of information out, it has been helpful to have it delivered.
One use-case I’ve persisted in is running a server for Luanti, the community-developed, Minecraft-like game my kids are obsessed with. The game doesn’t have a GUI interface for running a local server; you have to log into the server on a terminal and use the command line to get it started. This is tricky even for someone like me who lives in the terminal whenever I can, as you have to remember a bunch of options and file paths. First, I tried to get a coding agent to make a GUI frontend for me. It was pretty, but it kept glitching out, and the agent was continually causing new glitches by fixing old ones. I realized it might be simpler to just have the agent run the server and interact with it by chat. This has proved a fairly workable UX, even if sometimes it makes configuration mistakes here and there. But a polite reminder usually gets us to a fix.
Maybe it would be better to force the kids to learn terminal commands, which I tried in the past, but I’m not sure it’s worth it. I am also proud to note that my kids experience no special awe or pleasure in interacting with AI, and mostly they just make fun of its mistakes.
Verdict: Homes are for people and for interacting with people, and there may be some ways that an AI can help facilitate those interactions. Otherwise, it should probably stay out.
Will AI find a way?
Most of the verdicts above have been pretty negative. I have given these experiments serious effort and a fair bit of technical knowledge. But—and perhaps I am lucky in this—most of my life and livelihood still depends on being human with other humans.
Perhaps I would find additional benefits if I gave agents access to more levers to play with, like my emails accounts and diaries and calendars or whatever. My job’s Microsoft apps are already AI-ed to the wazoo, and it has not once been helpful. But with my personal tools, sorry, that’s not happening, at least not without vastly more assurances about these systems’ privacy and security guardrails.
Perhaps what I’m missing is just better AI. Will the next model version change everything? Maybe? I wouldn’t hand over my community’s water supply and power grid to find out.
I started these experiments as something close to an AI optimist, despite the nasty corporate forces at work. I come out not quite as a pessimist but where a lot of us outside the CEO class have landed: Meh. I might think differently if I were building drone swarms or trying to scam people or obsessing about life extension or looking for needles in cybersecurity haystacks. But I’m not. If I find hope in these experiments, it is in three things:
- It is possible (and could be widely feasible) to have AI tools that are loyal to their users rather than being surveillance devices for corporations, and these can be useful for well-chosen purposes.
- It is possible (and could be widely feasible) to use AI tools in ways that help us strengthen relationships and work together, rather than supplanting ourselves, but we have to work against the corporate narratives.
- For most of the things that are actually valuable about life and work, we can more or less ignore AI altogether.
In its present form, this is technology for overlords and those aspiring to become them. Beware the feeling that it is making you vastly more powerful, and effective, as it likely involves troublesome shortcuts that will spell trouble in the end. The same warning applies to the current state of the stock market.
Death to hype, and long live play.