Collective Social

Track what you love. Share what matters.

@brittanyellich.com
Collective Social — Track what you love. Share what matters.

About

A media collection and media club app. It allows you to create and share lists of media like books, movies, and TV Shows, for you and for your media club.

Reviews

No one has reviewed Collective Social yet. Yours could be the first.Write the first review
Working with clayI originally wrote this essay as part of The Pragmatic Engineer Essay Contest and am sharing it now! I joined GitHub from a small medical education startup in May 2022, and it ended up being a very interesting company to work at during the rise of AI. At the time, GitHub Copilot felt revolutionary and was gaining traction among some developers, but it was still mostly about code completions. A lot of what it could do still resembled the tools that came before it, like ReSharper and IntelliSense that already worked as autocomplete. I worked there for four years, and personally moved from AI skeptic to enthusiast throughout 2025. I still remember reading a blog post from Thomas Dohmke in early 2025, then the CEO of GitHub, about the agentic future, and thinking it felt like he was describing a hyper-futuristic science fiction dystopia. It felt unreal that engineers would be ceding all of their development decisions to AI tools any time soon. Yet a year later, it was real life. During that time I also got to do a lot of dogfooding of the GitHub Copilot coding agent, which could autonomously create full pull requests from a GitHub issue. The coding agent was released to public preview in May of 2025, but internally I got to start using it earlier than that. This was the first thing that I used that made me start understanding the potential of AI. Suddenly I had extra capability to get all of the little bugs completed that went to our backlog to die, and that felt incredibly powerful. In mid 2025 I joined an internal group of "AI Advocates", hoping to find other folks interested in AI and how the technology could change software development and beyond. We had a Slack channel where folks would share a few things they found interesting, but it was hard to gain traction on any specific advocacy efforts. There were a lot of skeptics internally around the move of GitHub to focus on Copilot, and I understood why. Many had joined during the big 2021-2022 zero-interest rate period hiring boom, when GitHub was still primarily the home for developers and open source, and they didn’t necessarily join for the transition to AI company. GitHub was an interesting place to learn to use AI. The model agnostic approach to GitHub Copilot meant I had a lot of opportunity to experiment with outcomes from different models. Working in billing I had a lot of understanding around how much Copilot I used compared to other people at my organization and beyond. And I had access to dogfood a lot of tools before they were available to the public, which meant I had a lot to say about them in blog posts and conference talks. By the end of 2025, I was fully AI-pilled, and very interested in learning how other developers were using AI to expand their development capabilities. Discovering atproto Something else was happening at the end of 2025. I started experimenting with a technology called AT Protocol, or atproto. I largely dropped out of most public social media in 2015, after being frustrated with constantly being shown ads and the “social” part of social media disappearing. I was no longer seeing updates from my friends, and was seeing more rage bait and folks yelling at each other online. It stopped being fun for me then. In 2022, I read The Chaos Machine by Max Fisher after it was recommended by my book club at GitHub (where I met Erika Eggemeyer and Bethany Janos, my cohosts for our podcast, Overcommitted!). It provides some evidence to the patterns I had noticed online. Most social media companies get revenue through advertising, which means they are incentivized to show as many advertisements as possible to their users. In order to keep people on the platforms, they’ve done years of A/B testing to determine what keeps people scrolling, developing common UX patterns we see now like infinite scrolling and push notifications.  They’ve also developed algorithms that optimize showing content that keeps people engaged for as long as possible. Most people don’t engage with content that makes them happy, but do often engage with content that makes them angry. The additional likes and comments on rage-inducing posts means that many social media algorithms are built in such a way that they continuously show you more and more rage-inducing content, because being enraged keeps people commenting and reading for longer. This has had serious mental health effects, particularly for teenagers and young adults, throughout much of the world. When switching into tech from occupational health and safety I was told to get on “developer Twitter”, since that’s where everyone in tech networks, but every time I re-joined Twitter it was a terrible experience. It reminded me of the same problems I was experiencing on social media in the past, where I was constantly shown ads and content I was not interested in seeing. I joined Bluesky in November of 2024, with a large wave of other users after the US presidential election, and it felt different. The developer ecosystem was easy to fit into and actually fun to participate in. We were having real conversations about what was changing in tech in ways I wasn’t able to find elsewhere online. When I learned about atproto in 2025, I realized how it was designed to address the problems of misaligned incentives from large tech companies specifically, and I was hooked. All of my side projects and public speaking efforts switched to atproto at the end of 2025, and I became deeply integrated into the atproto community. It has felt like the one pocket of hope in tech right now. While much of tech discourse has been drawn into political arguments of whether AI will save the world or doom it, the atproto community is working on building an internet that democratizes access to your data and changes the incentive structure on how software applications behave. I initially discovered atproto primarily through a social media lens, but I didn’t realize until recently how ideal atproto is for AI and agentic work. It turns out that having your own personal cloud that you own and can use across many applications can be the perfect substrate for context storage by agents. Becoming a Product Engineer When I left GitHub in May of 2026, going to Bluesky to work on AI for the Atmosphere felt like an unbelievably perfect fit. I was delighted by the opportunity to work on something bigger than myself, and to go to a smaller company of very AI-engaged engineers to have a much larger impact when compared to being one of thirty engineers working on billing at GitHub. I began a role as a product engineer. I have always been a very product-minded engineer and passionate about being good at more than one thing. With AI advancement where it is, now is a great time to expand your skillset so you aren’t dependent on a single thing that can be replaced overnight by AI. Despite much doomerism and science fiction about the future of AI, no one actually knows what the future is going to look like. AI could easily have major breakthroughs in any specific skill set, and if it's the one you have become an expert in you will likely be out of a job.  It's not enough to build great software anymore. You have to also get involved in infrastructure, or product, or marketing, or sales. You need more than one skill to be competitive in a world where any one of those skills can be easily made redundant.  Having re-invented my career once already, I knew that having a diverse background only made me a better software engineer, and I figured the same could be said about becoming a product engineer full-time. I knew I was ready to move to a smaller team, because it was becoming increasingly difficult to work in an environment with many other people in a large old codebase, and I also hypothesized that being a generalist was going to be the winning strategy of staying competitive in the world of AI.  I'm now working full-time on the exploration team at Bluesky, building Attie, an AI tool for the Atmosphere, as well as exploring new ideas for how the Atmosphere can expand. This consists of keeping up with the latest on AI and agents, some infrastructure work, product and feature planning, and even communications and community management. It is a lot of different hats, which feels like the exact right move to stay competitive in a changing world.  Bluesky is small. There are 45 million people in the Atmosphere, the ecosystem of apps built on atproto, but up until the beginning of this year there were only around 25 people working on it. Now there's around 50. While it has grown quite a bit as myself and other folks have joined, it still feels like a fairly small company with a lot to do, which means a great opportunity to learn to do a lot of different things. Using AI to connect the dots Getting up to speed in a new area with no background in it is now easier than it has ever been. I have never worked for a social media company before, and Trust and Safety (T&S) work was completely new to me before starting this role. T&S is a small, but very important, part of my job. It means preventing folks from using our tool for harassment, illegal content, or just generally doing things they shouldn't be able to do.  I started working on Attie right as another engineer was leaving. Bluesky has a few folks dedicated specifically to trust and safety, but in general it is a small, scrappy, and extremely busy team, much like the rest of Bluesky. I had some time from the head of T&S, a couple of meetings to gather context and understand what was at stake, but otherwise had to use AI to get up to speed as quickly as possible. Having the breadcrumbs from folks who knew T&S inside and out was incredibly valuable, as it helped me get started and know what to research and what the largest T&S risks were for an AI tool that generated content based on social context. Then I relied on learning with AI, generating data to experiment, and testing the new T&S methods I put in place to understand how…Oct 25, 2026Atproto Is the Context Window You OwnI don't think it's a coincidence that atproto is taking off at the same time AI is. Well, okay, it kind of is a coincidence. I don't think anybody planned these things to work out at the same time. But it's a lucky alignment that I think is worth savoring: AI is maturing at exactly the moment the atproto ecosystem is getting diverse enough for your data to really mean something. AI is the tool, and your atproto data is the substrate. Quick disclosure before we go further: I'm an engineer at Bluesky, and I spend my days building Attie, an AI tool built on atproto. This post is about what I've learned while doing that. It will probably sound a bit like an advertisement, and I suppose it is (please check out Attie if it interests you, it's in beta and we love feedback!!), but these are also my own thoughts that have been evolving over time as I’ve become one of the biggest Attie users. The context you're accumulating without trying Anyone who has used AI tools seriously knows that they're only as good as the context you give them. You can spend a lot of effort assembling that context by hand, or, in the case of your atproto data, you can realize you've been accumulating it all along. Every post I write, every blog entry I publish, every like I leave behind lives in my PDS (Personal Data Server, the folder where your atproto data lives) as structured data. And because there are so many different shapes of atproto apps now, my PDS has insight into a lot more than my social feed: games I play, hobbies I'm interested in, what I've liked and found interesting, who I interact with. All of it also sits in a wide graph of other people connected to the same thing. That's a fundamentally richer starting point than a traditional algorithm that looks at "things you like and how to show you more of that." A recommendation engine sees engagement signals and infers a shadow of you, colored by what they think they can advertise to you. My PDS contains things I actually said, actually made, actually did, in schemas any tool can read. Who gets to build on it Meta has rich context about you, too. Even richer, in some cases, given I started a Facebook account when I was 16 years old. Facebook has known about your hobbies and your social graph for twenty years now! The difference between that and atproto is who gets to build on it. Right now, Meta is the only one that gets to know your context, and it uses that knowledge to figure out how to manipulate you into buying things from advertisers. That's literally their whole business. On atproto, anyone can build tools on that data. Any new tool can access your context and build something highly tailored to you and your interests on day one. There's no cold start, no moat, and even better: you can build those tools yourself. You can customize your tooling to fit your exact needs, interests, and wants, instead of waiting for a platform to decide your niche is worth the efforts of an engineering team. What this actually feels like This is exactly what we're building with Attie: AI personalization that you want, that serves you and your community. Here's what that looks like, as I've become one of Attie's most enthusiastic users. I've used Attie to assemble pages summarizing events I had major FOMO about, gathering all the relevant links and conversations in one place so I could catch up like I'd been there. That only works because the event, the people who attended it, and everything they said about it all live in the same open graph that I'm living on. I've used it to pull together the links and conversations around specific areas I care about. I've used it to more deeply connect to the Fiber Arts community, which my PDS knows about because I've actually posted about my crochet projects, not because an algorithm watched me linger on yarn websites. And I've used it to find the people to follow and what's happening each week in the atproto community, which amounts to building my own "what's happening on atproto" newsletter that Attie updates for me every week. Notice what none of these are: none of them is "show me more posts." Every one is an artifact I asked for, in a shape I wanted, pulled from my connections and my interests. That's the shape of personalization I want more of, and it's a shape that engagement-optimized feeds structurally can't offer. But everything is public The obvious problem with all this openness is that "anyone can build on your data" cuts both ways. Open data is open to the tools you love and to the scrapers you don't. It also means there isn't really a way to have the deep and meaningful conversations that you can only have in private spaces on the protocol, and that shapes a lot of the ways the data shows up. Something I haven't mentioned here is the ways I could have been using Attie to show me parenting content, because that's something I rarely talk about in public. I don't share things about my kids on the public internet, other than the fact that they exist, because I don't think everyone needs to know about those things. That context doesn't exist on the protocol at all, which means no tool can serve me with it, no matter how much I'd want one to. Right now atproto is open by default, but it won't be forever. Atproto spaces are here, in alpha on the protocol, which means real control over who you consent to read your data. And my bet is that this makes the whole graph richer. People are likely to put more into a store they control (I know I would). The data will continue to get even more rich as atproto spaces are released, and new apps are going to evolve to leverage them to build incredible new experiences for individuals and communities. Go build something There's a second half to the lucky alignment I opened with. AI isn't just the tool that makes this context useful, it's also what makes building on it reachable! You can describe the tool you wish existed and have a coding agent help you make it real, in an afternoon, without a platform's permission and without an engineering team behind you. The substrate is here and it's yours. Atproto is the context window you own, go see what you can make for you and your community.Aug 21, 2026AI as a Tool for the Mental LoadEarlier this year I wanted to plan a family vacation. For anyone who has ever planned a trip for young kids, you know that that is actually a lot of work. Planning meant researching, booking, and preparation. It meant hunting for places to stay, remembering to buy the plane tickets at the right time, building a packing list. It meant prepping my kids in advance to get them excited about the trip and assuage their fears about doing something new, so they could actually enjoy the journey. It meant finding someone to take out the trash while we were gone. It meant getting my dog boarded, and getting my dog's vaccines updated so he could be boarded. It meant remembering to take time off and prepping everyone at work for the time I'd be gone. Every one of those things is yet another task that lives in my head in anticipation of the trip, on top of everything else happening in my life, working full-time and raising three young kids. This time, though, all of it got easier, because I could brain-dump everything that needed to get done in a single session with Claude and got back checklists, packing lists, calendar entries for when each thing needed to happen, research on which car company to book with and where to rent a stroller and how to prep the kids for the trip. The result was that I found myself actually looking forward to the trip, instead of seeing it as yet another burden of being a parent trying to build these core memories for my kids. The thing that made the difference was AI. And what it did for me had basically nothing to do with productivity. The mental load is the tracking If the phrase "mental load" rings a bell, it's probably because of Emma's 2017 comic "You Should've Asked", which put a name to something a lot of us (particularly women and default parents) had been living without a term for for many years. The comic shows that the exhausting part of running a household was never the chores themselves... It's the noticing, the anticipating, and being the one who remembers what needs to be done and when. The load isn't the actual doing of the work. It's all the mental effort that comes with knowing the work that needs to be done and making sure it happens. You can hand someone a task. Handing someone the knowledge about which tasks need to be done and when is much harder. "Just share the load with your partner" The feminist answer to the mental load has always been redistribution: partners should carry the tracking too, not wait to be assigned tasks. I agree with that, but also that in itself is a lot of work! I've been in a lot of community spaces where we share our frustration with the mental load of existing during this particularly difficult time, and here's what I've watched happen: folks patiently trying to get these points across to their partners, repeatedly, for a VERY long time, and either nothing happens, or they end up having to do more work to educate their partners, which is itself adding to their own mental load. I'm not against redistribution. I'm saying the people carrying the mental load are tired of waiting for it to be redistributed and being the educator of that redistribution. A tool showed up that actually helps with the tracking and knowledge-keeping now, today, without needing to be convinced, reminded, or managed. A tool that can help with creating tasks for multiple people, with whatever reminders need to be built-in for each person, and to alleviate that burden from the individual traditionally holding it. And people should use it! What this looks like on an ordinary week The vacation planning was a big one for me, but the place I feel this most is the most ordinary thing imaginable: meal planning. Good god is it exhausting. Constantly remembering which foods my kids like, which ones they don't like right now, remembering to introduce new things, making sure we have a grocery order scheduled (yes I order my groceries to either be delivered or for pick-up and rarely go in the store and yes I realize that comes with an awful lot of privilege), actually ordering the groceries, and adding all the other things we need to the list throughout the week. Now I have a tool that does that for me: Claude creates a meal plan for me each week of meals that my kids actually like, while also introducing new things occasionally and remembering which days of the week we have events going on (like gymnastics class) and need a quicker option. Google Assistant allows me to quickly say out loud "Hey Google, add hand soap to the grocery list" from anywhere in my house. A tool I made by describing what I wanted to Claude, often called vibe-coding, adds all of the things to a grocery list each week, including any items that were added to my Google list for the week, for Instacart shopping on Sunday and sends me a reminder to actually look over the plan and make the order, so I don't have to remember that weekly chore. The remembering happens somewhere that isn't just in my head, and this network of tools makes it happen. One of the objections I hear often about AI is that I'm losing something by handing the thinking over to a robot, that ceding your thought processes to AI means you'll get worse at thinking over time. Two thoughts on that: First, I don't think the skill actually disappears. I've had a calculator available for my entire adult life and have been out of school for over a decade, and I'm confident I could still bang out some calculus if I absolutely had to. Would it take me longer than it would have if I had been doing math regularly? Absolutely. But I know I could still figure it out. Second: I actually don't mind losing some of these skills. Taking up my brain space with "which of three children currently accepts bell peppers" is not a skill I would miss if I lost it. Not every capacity your brain gets conscripted into is one worth defending. Yeah, my brain may physically change if I no longer need to remember each of these things, but actually that doesn't sound too bad at all. Where the capacity went Even better than filling my brain with the worry and stress of everyday life, when the tracking moved out of my head, the space it left behind didn't stay empty. I started writing blog posts again early last year (you can see the uptick right on this blog). I started my podcast, Overcommitted, around the same time. That timing wasn't a coincidence: it happened when I realized I had more creative energy because I wasn't spending it on the menial tasks of just surviving. I have this boundless creative energy that would otherwise get squandered on the day-to-day logistics of existing, and for the first time in my adult life, a meaningful amount of it is going somewhere I chose for it to go. And it's not just about output or productivity like the AI industry would have you believe from their poor attempts at marketing. My mom is a heavy Claude user now. I've watched her relationship with food and exercise change over the years, and lately I've watched her use Claude to learn about the importance of protein and building muscle mass as you age, about balance control and aging healthily. I've watched her shift from thinking about food and exercise in relation to body image toward becoming healthier and aging gracefully without losing her physical self in the process. She chats with Claude about what she's eating, gets advice on better choices, organizes her workouts. She's told me herself how much of a difference it has made in her life to have that as a tool and how she really appreciates the positivity that the tool brings to her day-to-day, and how it's super accepting when things don't work out as planned, which is actually incredibly cool. That's a person using a tool to organize the human thoughts and feelings and messiness that go with living, not becoming some 10x productivity expert. AI tools have so much more potential for good in the world than I think that they are given credit for. Multiple things can be true I want to be clear about what I'm not saying. The way this technology was developed is problematic in ways that are genuinely hard to wrap my head around, and the questions about its footprint are real. I'm not asking anyone to stop caring about those things. I care about them. We should continue to care about them and attempt to mitigate those impacts in whatever ways we can. But discounting the tool entirely because of them oversimplifies it. Two things can be true at once. Technological advancement often comes with a very steep and ugly price. But I think that the people most burdened by the mental load are being told that the first tool that actually meaningfully helps with it is morally off-limits. I think that's a loss for those people, and for humans as a whole as their energy stays squandered on survival instead of whatever they're passionate about. That's not some win against the technology industry, it's a loss for everyone. Folks trying to survive today, the ones sandwiched between raising kids and caring for aging parents, working full time in an increasingly expensive world, holding entire households in their heads, deserve to know that a tool exists that actually helps with the mental load of it all. This isn't just a productivity tool, it's a life tool, one that can help you organize the messiness of being a human in a way that hasn't ever been possible before. You're allowed to use the things that help you, and now you don't need technical skills to create a system that can alleviate your own mental load. You just need a bit of creativity and patience to explore this new toolset, and discover how it can help you organize all of the things that are currently living only in your head.Jul 30, 2026

On Bluesky

All mentions
ATProto PDX
@pdx.atproto.camp ·

📣📣 TUESDAY, DECEMBER 1 at @codetv.dev studio! Join us for a Hack & Chill, where we'll play with the spaces public beta 🎉 @brittanyellich.com will do a demo of how spaces works, then we'll hack away - with your own laptop, or you can pair program with buddies! RSVP here:

ATProto PDX: Spaces Hack & ChillIt's time to join your fellow idealists, builders, and shitposters for another meetup about the atmosphere! Together, we’re doing our part to create a more social web so the internet belongs to us, Th...atmo.rsvp