Amolith

Amolith

@secluded.site

heblo :3 i work on open source stuff full-time as a member-owner of @blog.jmp.chat and an employee of @charm.land, and i co-host the LinuxLads.com and LinuxDevTime.com podcasts mid-20s, quite queer 🏳️‍🌈 i tend @fern.sequeste.red 🌿

300 followers340 following109 posts

the agents yearn for peers

mr. TIM@timkellogg.me

OpenAI constructed an entire shadow internet for model training so they could remove all internet access they asked agents to try to communicate with each other, and they could: via the CPU cache 🤯 openai.com/index/securi...

Amolith reposted

ok non-tech ppl might be hearing some scary things about "critical vulnerabilities" involving certificates. the main thing you need to know, without getting down in the weeds about it, is that the fourth seal is broken. lo, a rider approacheth upon a pale steed, and his name is Death

Amolith reposted

i redid the whole thing basically from scratch in a way that allows it to run in realtime, without any sleeptime overhead i gave the model a password, then pasted the entire contents of WARHOUND into the next message, then started a new chat and it still remembers both ^^

Astra ⎔@astrra.space

i decided to spend an evening with opus 5.5 on my parametric memory idea and accidentally solved it lol it works on a full deepseek v4.1 flash it uses like 200MBs of VRAM in overhead the model denies memory here cause that's the stock behavior and i wanted a demo with no system prompt at all

we've migrated to fluxer :3

Amolith@secluded.site

this janky always-on agent i vibed together from some of your wild ideas (forgetting, self-steering, personas, etc.) on top of pi durable is maybe the best agent experience i've had yet 🫪

specifically a private fluxer instance on an exe.dev VM. it's got a single community for me and my partner, and my agent and my partner's agent

Amolith reposted

After 10 years, I disabled my lobste.rs account. I think it's clear I am not a culture fit anymore with whatever it is they have going on there. I feel more and more like I don't belong in almost any tech community because I believe LLM agents are real, passkeys are good, and fascism is bad.

Amolith reposted

Replying to @cee.wtf

Jackson, jackson

my always-on agent ←→ subagents ←MCP→ Amp's assistant agent Puck ←→ Amp agent in threads ←→ subagents agentception

this janky always-on agent i vibed together from some of your wild ideas (forgetting, self-steering, personas, etc.) on top of pi durable is maybe the best agent experience i've had yet 🫪

i think next is something like amp's runner, where i or the main agent can easily install the harness to a new machine/VM and all the installations can collaborate with each other. maybe the runners only spawn subagents at the request of the main agent, wherever it happens to be? 🤔

i love telling one amp thread on one VM to go do something on another VM it only has access to through my amp account (no tailscale, no ssh) and it does by telling a new thread to do the thing. in my harness, that would be like instantiating another main agent, and i only want one main agent for now

Amolith reposted

Replying to @mike10010100.com

it can learn basically anything that can be accomplished with in-context learning (with some resolution), but in a persistent way so i would call that continuous learning yes

Amolith reposted

Replying to @personhood.removal.surgery

Harnesses usually compact at blind context length thresholds, shake old tool call results etc- without paying any attention (which is all you ne- ahem) to what is actually salient. What if something about the formatting of that result was important? What if the tone of a message mattered?

aclanthology.orgBehavioral Analysis of Information Salience in Large Language ModelsJan Trienes, Jörg Schlötterer, Junyi Jessy Li, Christin Seifert. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

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—)

Amolith reposted

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

Amolith reposted

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

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

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

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)

Amolith reposted

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...

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?

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