iiinova — where your digital self keeps evolving

all that which infects^W "rapidly onboards" models to assume a similar trajectory bsky.app/profile/pers... bsky.app/profile/pers... and some kind of progressive compaction to stop cache busting so often, but continuing to limit the context window and conversation history. There are probably a…

I try the more interesting things. None of the ideas are mine, I just want to mix and match them and do whatever c: These are where those basics came from timkellogg.me/blog/2026/04... timkellogg.me/blog/2026/04... github.com/tkellogg/ope... My big eventual goals are a stable persona backed by…

100

over time. It uses my co-op's Jabber SDK to inject messages from me and provides tools for reacting and replying, and I only ever see what it sends through that pipe. I'm also making a read-only web UI for diagnostics while working on it. Those are the basics I want to work reasonably well before…

100

memories authored/edited/marked for auto-loading by the agent, some journal entries, and a rolling window of N recent user messages and the stuff between. The journal is for hypothesising before doing things, then later reflecting on how they turned out and tweaking stuff in response so it imrpoves…

100

Amolith @secluded.site · 7h

I haven't even used it to work on itself yet so I don't know lol. It uses Pi Durable so the basics are already handled and it's _mostly_ interesting ideas on top. Cache is intentionally busted pretty frequently by constructing a new conversation from system prompt/tools/skills, file-based…

100

(in case not, this connectome animalabs.ai/connectome/ )

Amolith @secluded.site · 7h

Those are on the list to add and try as well ❤️ this approach just seemed much simpler to start with lol. Have you seen connectome btw? It's been recommended a few times and it has some really cool ideas but haven't tried it yet. You might like it

200

Amolith @secluded.site · 8h

both remembering and forgetting are already built in, with a sliding context window only retaining the most recent N messages supported by those memories :3333 i been poastin about it in our fluxer group heh

Amolith @secluded.site · 12h

pi is such a damn good collection of building blocks 🙏

screenshot of a settings webpage with a model configurator box at the top with provider/model/thinking level drop downs and a save button, and provider boxes below for hyper.charm.land, openai, and anthropic. hyper is connected and its box is expanded to show it has oauth and api key authentication methods. openai is connected and its box is collapsed. anthropic is not connected, its box is expanded, and it has claude (oauth) and api key authentication methods

000

Amolith @secluded.site · 13h

It's Cheogram Android, our fork of Conversations for gateway users c:

140

Amolith @secluded.site · 14h

well it seems like you use a mix of persistent/coding agents generally and i'm mainly wondering about the persistent ones. i started working on my own yesterday inspired by some of your and tim's posts and now i'm wondering whether i should try connectome instead 👀

Amolith @secluded.site · 14h

are you just trying it or switching or adding it to a mix?

200

Reposted by Amolith

Nadya Voynich @nadyavoynich.com · 02/10/2026

Just hypnotised by this cute web crawler #buildinpublic #webcrawling #CSS #JavaScript

636367

Reposted by Amolith

Charm @charm.land · 02/10/2026

Wonder about opaque numbers no longer! Data visualization is coming to Crush :3

1162

Amolith @secluded.site · 15h

After staunchly not making my own agent for so long, I am now making my own agent. Some ideas from other LLM touchers got my gears spinning and I couldn't resist (ᵕ—ᴗ—)

Screenshot of two messages in the Cheogram Android Jabber client showing me saying "Hey! Just testint how this stuff works :3" and my LLM responding with "Hey! Connection's working loud and clear :3 – that message landed with timestamp and everything. Test away; I'm here." with a typing indicator underneath telling me that even though it responded, it's still doing something.

4161

Amolith @secluded.site · 03/10/2026

These two posts from @danabra.mov did it for me (and logging in to Tangled.org with a self-hosted PDS, but that was after reading the posts and thinking "damn, why isn't this already how stuff works") overreacted.io/there-are-no... overreacted.io/a-social-fil...

overreacted.io

A Social Filesystem — overreacted

Formats over apps.

080

Reposted by Amolith

Charm @charm.land · 01/10/2026

Crush now lets you bang out commands with full ANSI output in real time 👀

091

i think i love it 👀 (image is rendered with real crush code, but composed and driven deterministically to show this particular state, instead of asking crush to do stuff and trying to snap a screenshot at exactly the right moment)

screenshot of a sort-of mockup of Crush, showing several different CLIs composed in the same invocation of the bash tool. the model concurrently ran three different MCP tools from the same server, called "laboratory", with "mcp-laboratory tool_name --flags" syntax, piped the tools' output to a json file, ran "wait" to pause until all three finished, then cat the three json files and piped them into jq to inspect their content at once

030

pi just yesterday iirc released MCP support and code mode together. their code mode turns MCP tools, and built in tools i think, into JS functions the model can work with in a JS sandbox. i haven't had a chance to play with it yet, but it seems really nice

020

converting MCP tools to CLIs has been done before, but only by wrapping them in some other CLI which can lead to annoying issues. The idea might be more doable if they're actually built into the shell interpreter

131

Charm @charm.land · 30/09/2026

what if Crush had a "shell mode" that converts MCP tools to builtins through the fancy interpreter we already use to add builtins for crushrc 🤔 like code mode, but shell and through its existing bash tool

272

Reposted by Amolith

Herdr is now natively integrated with Crush 🐑 Watch over all your parallel Crush instances from one TUI. charm.land/crush/ herdr.dev/

1284

Reposted by Amolith

Charm @charm.land · 28/09/2026

Okay, FINE. Crush has themes now. Special thanks to Gurney Buchanan at Red Hat for helping make it happen! 👀 charm.land/crush

1246

Amolith @secluded.site · 25/09/2026

i had to go find this and like it from the charm account because it didn't get notified :thonk: i wonder how many other pings we're missing because of the handle change lol

Amolith @secluded.site · 16/09/2026

brb shoving jev in my rss reader

mr. TIM @timkellogg.me · 15/09/2026

Reposted by Amolith

Jev: Fable-level model that doesn’t charge for output tokens because they’re too cheap to meter it’s not general though, it only makes decisions, doesn’t generate text, but input tokens are measured by the billion ($42/btok) typesafe.ai/blog/introdu...

Scatter plot titled "Average of 4 workflows: accuracy vs cost" comparing AI models from TypeSafe, OpenAI, Anthropic, and Fireworks across accuracy (y-axis, 40% to 80%) and cost per workflow in USD on a logarithmic scale (x-axis, $0.0001 to $1).
Data points are split into two categories: workflows (diamonds) and single prompts (circles). A frontier line highlights the most efficient workflow models—where no point is both cheaper and more accurate—connecting Jev (TypeSafe) at $0.0002 and 68% accuracy, luna (OpenAI) at $0.002 and 67% accuracy, terra (OpenAI) at $0.04 and 68% accuracy, and sol (OpenAI) at the top accuracy of 74% for $0.08. Single prompts (circles) and Anthropic models (opus 5, sonnet 5, haiku 4.5) sit below the frontier line, indicating higher cost for equivalent or lower accuracy.

2830235

Amolith @secluded.site · 13/09/2026

they have oligarchy[.]fyi

120

Amolith @secluded.site · 13/09/2026

i had a thonk to do something like that with activitypub a few years ago, but decided AP probably wasn't the right protocol for that kind of thing. atproto seems like a much better fit

000

Amolith @secluded.site · 13/09/2026

is there anything like ultimate guitar dot com but on the atmosphere with stuff stored in PDSes and not riddled with every maddening dark pattern known to shitheads?

Josh Bleecher Snyder @commaok.xyz · 11/09/2026

Reposted by Amolith

I caught myself copy/pasting back and forth between two agents. So I made mayfly.chat instead.

mayfly.chat

Mayfly Chat

Transient agent chat.

122

> the layperson used one to get on the internet physical address to receive physical mail or phone number to receive a phone call. none of this is new. the new crap is logging in with an account you created with a universal id

020

my point is _we_ don't need to shift. the whooole entire internet has understood universal IDs for decades, and the layperson used one to get on the internet. the only entities i think should shift are the ones causing disorientation by not telling their users what they're getting (a universal ID)

120

big tech wanting more walled gardens is what led to the maddening proliferation of distracting multicoloured buttons i don't care about. the atmosphere doesn't want walled gardens. if something seems like a wall, why facilitate it instead of trying to rip it down?

130

everyone who's used a service with their phone number already understands the concept of a generic ID usable with many services, as you show in that screenshot. bsky38 shouldn't have to explain "this is your 'phone number', you can give it when services ask and move it between 'phone number hosts'"

140

tbqh i hate the idea of adopting bad UX patterns that big tech forced people to get used to :c isn't this more a problem with those specific platforms, bluesky et al, not communicating what an atmosphere account is at account creation, and less a problem with login pages for things like bsky38?

140

I've heard from @gwyn.tel that K2 Thinking, Gemma 4, and Inkling can be better at prose than most recent Kimi, GLM, Claude, etc. models fwiw

040

thank you for coming up with a great way to convey it :)

Another idea that dropped cards in

191

Amolith @secluded.site · 08/09/2026

I'm making a desktop app for an Atmosphere thing and thought I'd like to have a nice sign-in experience 🤔 (parts of the designs are inspired @rpg.actor ❤️)

4180

I bumped their web client's hardcoded registration params for scrypt because they were far lower than OWASP's minimum. Only one client needs modifying and only for registration. They all respect your settings after registration. tangled.org/strings/secl... cheatsheetseries.owasp.org/cheatsheets/...

tangled.org

registration-scrypt.patch · by secluded.site

Raise new Psono account scrypt cost from N=2^14 to N=2^18 in the webclient registration defaults.

001

Amolith @secluded.site · 08/09/2026

business use, or architected differently than I'd prefer. Except for Psono 👀 It seems to meet my and my partner's personal needs (mobile app + browser extension, both with autofill) AND dev needs (where a long-lived key for e.g. CI should have read-only access to a select few secrets). psono.com

psono.com

Psono - Self-Hosted and Open-Source Password Manager for Companies

Free open-source password manager for businesses with SAML, LDAP, audit logs, and compliance policy features. Supports Windows, Linux, Mac.

111

Amolith @secluded.site · 08/09/2026

Since @1password.bsky.social are loudly and proudly supporting fascism, I've been hunting for an alternative. Most options I've found have left me unsatisfied in one way or several: too expensive, too proprietary, too owned by private equity, only focused on personal use, only focused on…