thoughts on attie
directionally very bad
i love atproto! i love what ai enables for people! i love the community of folks who do amazing work at this intersection and talk about it on bluesky!
but i think attie is the wrong product for the best funded atproto company to build and also that its poorly built.
as in, when people say words about why they’re building attie, i normally agree with those words. democratize the power of ai, show not tell, bring more people in the atmosphere. yes!
claim: attie doesn’t help with this!
i’m going to indulge in a rant about how awful the ui is for a moment and then talk about my more fundamental issue with attie
the ui is demo-ware at best
any relatively new project has rough edges, i get it. but attie has had consistently low quality UI since its introduction in march, rife with paper cuts all the way to totally broken ui
and another example
i work full time in open source ai dev tools! i know what a low quality, highly vibed UI looks like! i have shipped many poor UIs myself for demos or totally throwaway projects.
my issue is that it’s so easy to fix!
- install anti-slop skills
- tell claude/codex to scan your components and sic a bunch of haikus/lunas on them in parallel and then have the parent review and merge piecemeal. or something!
- write a QA skill! declare your aesthetics! you only have to care enough to check! tell a claude to use the safari mcp and bop around the ui to find bugs!
all of this is high yield, low toil, table stakes stuff that’d dramatically improve the quality of life of your users and decrease bounce rates. the point is to grow the funnel no?? the ui actually matters for this!
please, stop building a harness!
any reasonably well informed software engineer in 2026 can replicate any outcome attie provides without the whole network in clickhouse, in minutes, with an OSS stack on top of the thing PBC has that’s actually valuable. this is only barely hyperbole
attie's "Quests" are the canonical chat thread UX that every client does, but clunkier. if you add an MCP (which now can bundle skills btw) or a plugin w many MCPs, to an actually polished/mobile friendly client, you almost immediately have a better experience than quests.
new feeds are literally a tool call if you have a server to generate the feed from
their "Sites" (just announced today) are slow-following the ant/oai artifacts/sites and aren't on your PDS. i used to get angry at people for being cynical about bsky pbc's decision making, but goddamn if you insist on competing w the labs on features in the consumer facing chatbot thing then this is such a slam dunk opportunity to collab with a project like wisp.place
or even just talk with folks like wisp.place or Minor Mobius on a standard.site-ish lexicon for sites
it seems so well-within reach to:
- illustrate progressive auth for the community by asking for scopes to make a site
- put the site assets on the user PDS! why would this not happen even if you are going to recreate a low quality version of a community project as a footnote of your slopnum opus
- actually materialize some atproto promise by having the public sites on protocol so they're indexable?
- illustrate how private sites could go in spaces for those who have a spaces-enabled PDS. this particular branch is something a lot of people already maintain without so much issue
attie can't even differentiate Sites from its previous artifact-y thing called Pages... which Attie says are called Reports....
this is how Langchain made abstractions in 2023 and became the canonical Cronenberg, object soup hell project of the ai era.
trying to be more productive about my feelings
this has been a somewhat negative assessment thus far, i know!
its mostly because i care deeply about the issues that attie is meant to point at. i have spent a great deal of time and money and effort building, organizing and communicating for and with people in chicago and the broad interwebs to materialize the user agency vision of atproto that i mostly still believe in. i spend a lot of time meeting ai skeptics at all levels of programming skill where they are at, showing them that “yes ai is a tool of empowerment for you too!”
so i’ve ranted that Thing Bad but what do i think the best funded atproto project should build?
why did people care about the early twitter firehose? bc holy shit it’s a literal stream of consciousness. that was microblogging, this is everything! where is semantic search and analysis over the atmosphere as a service? “what does demographic XYZ think publicly about ABC?” as a service? not as a chatbot with a harness that’s objectively worse ergonomically than the massively popular harnesses (claude/chatgpt/gemini) that can add one mcp and do everything attie does but better. host an mcp like dame or Nick Gerakines or Sri or myself that people can plug into whatever harness they want!
you raised $100MM and curated a wildly valuable data product that an entire ecosystem can use to jumpstart social experiences. attie does not uniquely leverage this
i spent a lot of time watching bsky folks wave off mcp as dumb or code mode as pointless while not shipping or open sourcing any compelling ai tooling. meanwhile, the whole world is using these things at an already extreme scale, at an accelerating pace. step into the massive stream of interest! don’t create a jank and tiny little eddy.
yes, i appreciate the intentionality often displayed by pbc around a lot of things (especially at the protocol level). there’s tradeoff between wave hopping and principles, i get it! you’re just slow and it frustrates me
this is mostly a vent, and i don’t mean to diminish the definitionally multifaceted nature of the work that real, over-subscribed humans are doing at pbc for what i believe to be good reasons.
but please for the love of god, ship a skills repo and a hosted mcp for atproto crud and ill continue singing your praises, thanks
(maybe also get on the ant/oai plugin marketplaces as well please!)
