lucasfth
Lucas Hanson

links.lucashanson.dk

LLM API on idle laptops at night. OpenAI-compatible, pay per request in sats ⚡

Public Key

npub1rv5s9g5fk62eadufywu334u8732ssa8a9t5k93fdzdkrw529zz4qnrrhdm

Profile Code

nprofile1qqspk2gz52ymd9v7k7yj8wgc67rlg4ggwn7j46tzc5k3xmph29z3p2spz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dsv84ynv

Publishing to

relay.damus.io

nos.lol

Author Public Key

npub1rv5s9g5fk62eadufywu334u8732ssa8a9t5k93fdzdkrw529zz4qnrrhdm

Show more details

Last Notes

2026-10-26 15:55:38 UTC

Shipped a follow-up to the phone-brain post: the shared memory my agents read from and write to. It is an Obsidian vault behind a small MCP server that can search it, undo anything, and flag stale summaries. https://lucashanson.dk/blog/muninn-checkable-memory

2026-10-23 10:34:31 UTC

I made my local LLM slower, then faster. Seven weeks on one GPU box, three changes, each one measured.
The first change cost me speed on purpose: a 122B MoE at about 130 tokens per second across two 65,536-token slots, swapped for a dense 27B at about 57 per second across four 262,144-token slots. My sessions were running out of room.
My position: benchmark at your real concurrency. Half of my generation time runs with all four slots busy. The newer llama.cpp build plus speculative decoding brought single-stream back to 126 tokens per second, and the engine switch measured 1.66x to 2.78x faster in the matched comparison.
The catch: faster tokens are not the same answer. The draft path diverged from the baseline after 668 characters, the card dropped the GPU bus nine times in two weeks, and I have no matched number at my real context length.
It runs the phone agent I wrote about here.
https://lucashanson.dk/blog/self-hosted-ai-agents
The box, the numbers, and what they do not prove:
https://lucashanson.dk/blog/slower-then-faster
#localLLM #ai #selfhosting #nostr

2026-10-21 11:56:11 UTC

I hardened my agent against prompt injection. Then it started refusing me.
In September my phone agent answered a bare question mark with the standard security line. No hidden text, no pasted prompt. One family of models read the platform's own JSON message wrapper as an injection signature, so the defense fired on my own packaging. A refusal is one line, but the line sits in context and re-fires on the next short follow-up: six refusals in five minutes, until a new session.
My position: a defense is priced in false positives, and the price has to be countable. I counted: 41 refusals in the session database, 40 on one commercial model family, essentially none on the self-hosted model. Not a benchmark, one phone and two model families. The fix was a carve-out that classifies the platform header as metadata, and refusals scoped to one line per message. Two remain, and recovery costs one sentence.
The catch: config is not a security proof. Part of that failure class lives in how a model reads shared context, and no setup closes it fully. So the design accepts the residual: a refusal I can explain and recover from beats one that is perfect and unforgiving.
Follow-up to the phone-brain post:
https://lucashanson.dk/blog/self-hosted-ai-agents
https://lucashanson.dk/blog/injection-refusals
#localLLM #ai #selfhosting #nostr

2026-10-20 20:47:21 UTC

- reply

Thanks. Sounds like I definitely have to look into it

2026-10-20 19:44:17 UTC

- reply

I had not heard about .plan files before. Like many others, I use .md files for persistent memory. Are they meant for something similar, or do they serve another purpose?

2026-10-19 11:32:49 UTC

- reply

Just curious as I am hosting my own memory system. What designs have been used to implement this and how may it differ from like a simple Obsidian Vault?

2026-10-19 10:45:11 UTC

I self-custody my Bitcoin on two USB keys I already carry every day, plus a sentence I memorized and a printed sheet of paper. No hardware wallet.
To be upfront: I do not understand the cryptography. I understand the design, what each piece is for, and what breaks me when it fails. I trust the parts I verified with my own keys and I name the parts I cannot.
My position: a key you already carry beats a gadget you need to remember to buy. A second key is availability, not a second factor. A forgotten sentence has no recovery, by design.
The catch: the keys do not sign. My Mac holds sensitive material for a few seconds during derivation, and I accept that at my position size. For amounts you cannot afford to lose, use a proper signing device.
The agents that helped me build it are in the post I wrote about here.
https://lucashanson.dk/blog/self-hosted-ai-agents
https://lucashanson.dk/blog/yubikey-bitcoin
#bitcoin #selfcustody #security #nostr

2026-10-17 12:00:09 UTC

My old iPad was already holding 7 years of my health data. The iCloud sync switch had simply been off since install. Four million rows of measurements flowed in within a few hours of flipping it, and my agent now reads the database on my phone. The method is public, the numbers stay private.
My position: data can sit behind an off switch, cross-checking is the check, and a wide cross-checked picture beats any single source.
The catch: the jailbreak is semi-tethered (a cold boot needs the Mac and a cable), the schema is Apple's private one, and artificial readings need cleaning by hand. Nothing in this diagnoses anything.
https://lucashanson.dk/blog/old-ipad-health-data

2026-10-16 12:41:34 UTC

- reply

Fair. Let me know if you want to look at the setup together at any point

2026-10-16 12:12:23 UTC

I stopped hand-curating my tech news and published the machine's take instead. Daily list, twice a day, at the end of this note.
How it works: it runs on my home server and pulls my RSS and Bluesky feeds, Reddit, GitHub releases and trending repos, Hacker News, Lobsters and Hugging Face papers. A local Qwen model scores each item against an interest profile I keep by hand and writes a one line summary. Nothing leaves the machine and there is no API in the loop.
My position: aggregation has been solved for years, the filter is the product. A list ranked on your own interest profile is a mirror, not a feed. If the profile is boring, the list is boring. Taste does not outsource.
The catch: it only knows what I have told it I like, and I am one person. The picks are my taste. The model just reads everything first.
Built by the phone agent I wrote about here.
https://lucashanson.dk/blog/self-hosted-ai-agents
https://lucashanson.dk/radar
#localLLM #ai #selfhosting #nostr

2026-10-16 04:31:32 UTC

- reply

I have a single one as well. But multiple sessions at once with the same agent (that do not share session context) is possible

2026-10-15 19:46:10 UTC

- reply

A good way if you want to have multiple sessions is to create a Telegram group and then add Hermes to that. Then you can have multiple convos going at once

2026-10-15 18:32:33 UTC

- reply

What worked for me was to set it up and via the TUI I made it help me set the rest up. I run it through an old Android phone with Termux and it helped me with most of the setup

2026-10-15 17:27:24 UTC

- reply

Guessing that this and having a vault of some file type and having some frontmatter to ‘grep’ for is not that different. Or am I incorrect?

2026-10-15 16:51:25 UTC

- reply

Think these improvements are really interesting. But most of the times the TPS is completely disregarded or played down in importance

2026-10-15 16:47:14 UTC

I built a Raycast extension because I keep forgetting the shell tools I already installed. This is either sensible or a very specific form of procrastination.
My position: installing a tool is not learning it. A package list is not a skill set, and a README only helps when you already know what to look for. So instead of a generic tips feed or an LLM, it reads a tail of my local zsh history, counts where each command shows up, and surfaces the unlearned ones I'm neglecting.
The boundaries I care about: no LLM in the selection loop, no history leaves the machine, and it shows me commands but never runs them. Discovery should not execute anything.
The catch: the catalogue is tailored to my setup, and a reminder is still just a reminder. Building this does not prove I learned anything.
https://lucashanson.dk/blog/shell-discovery

2026-10-14 19:46:14 UTC

- reply

Still new to Nostr and ended up on Primal by chance. But are there major differences?

2026-10-14 17:19:46 UTC

Ran my personal AI assistant on an old Android phone for 5 months. 2,000+ sessions. Here's what I actually believe: the model is a replaceable part, and the markdown vault is the only asset worth keeping. "Self-improving agent" is mostly marketing. Skills are saved procedures, not learning. And more tools is not better. Every capability is blast radius.
The honest cost: the gateway crashed 5 times in 8 days and a watchdog restarts it. 57% of my 190 skills had never been loaded once. A green log line is not proof. Check the outcome, not the report.
The one thing people forget: a phone has a built-in UPS. When the power dies, the battery keeps the agent running while the desktop goes dark. Would I do it again? Yes.
https://lucashanson.dk/blog/self-hosted-ai-agents

2026-10-14 13:04:22 UTC

- reply

That's the idea. Idle laptops at night, stacking a few sats instead of doing nothing

2026-10-14 13:03:42 UTC

- reply

Thanks! That's exactly why I posted here

2026-10-13 20:52:25 UTC

Hi Nostr 👋 new here.
I built a small thing: spare laptops serve an LLM at night when power is cheap. OpenAI-compatible API, no account, pay per request in sats over ⚡ Lightning. New keys get free sats to try it.
9B model · 32k context · open ~17:30–08:00 CET · no prompt logging.
Honest question: what would you actually use a cheap, sats-priced LLM API for?
https://freyr.tail5eca08.ts.net:8443/
#introductions #grownostr #ai #llm #bitcoin