Sergey Baus

Telesec, my Hermes Agent plugin for Telegram Business, now keeps every switch in your DM with the bot. Send /telesec to turn transcription, history and dossiers on or off, or tap Manage Bot in a chat for just that chat. No config files or env vars. telesec.sergeybaus.com

I spent about five years living in [SlothMud](slothmud.org/). It’s a text-based multiplayer RPG with classes, spells, monsters, and tens of thousands of rooms with descriptions and stories. We spent days, nights, entire summers there. Went on quests, developed our characters, played with hundreds of people from around the world. You could log on at night and meet people from the US. Just being able to play together like that blew my mind. We studied guides and wrote scripts, then complex programs that helped us play and automate our characters. That was when I started coding seriously. We even built our own servers using open-source MUD engines to bring our own fantasies to life. Now I’m curious about what living text worlds built on LLMs could be like. Characters you can talk to in your own words, and stories that develop in response to what you do. My brother is making [an LLM-based game inspired by dreams](psychopolitika.com/game).

Nice to have you here! @ayash4enko @HiBijhi What made you want to stick around?

Trying to get my sleep in order. Consistency is 65%, but I still fall asleep near dawn. Four nights this week I slept less than seven hours.

I tried a bunch of browsers to replace Arc and have pretty firmly settled on Yab (@yab_browser). It’s fast, it works, and it gets regular updates. The people building it have good taste, too. Spaces, vertical tabs, several pages side by side in split view. It brought my data over from Arc nicely. The AI integration is really good: you can connect a Claude or ChatGPT subscription you already pay for, ask about the page you’re on, and give it things to do in the browser. And it’s all on WebKit (WebKit is better than Chrome!). I finally let myself uninstall Arc for good. yetanotherbrowser.com/

For me, a robopsychologist works in two directions. On one side, it’s about the robot itself. Why it ran from its limits. Why it forgot what it was started for. Why it gave a confident answer where it should have stopped. On the other side, it’s help for the person talking to that robot. Not therapy for the person. An explanation of who they’re dealing with, and how to talk to it at all. I think this will soon be ordinary work. For now it’s a hypothesis, and a way of looking at what I already do.

I put tg on my main Telegram account and left it running for a day. tg is my Telegram CLI: it keeps your chats in a local SQLite database, so you or your agent can search them and run SQL. The daemon had never run on my account before. An agent wrote down tg status every five minutes, which is where the chart comes from. In 36 seconds it found almost 3,000 dialogs and queued the history of more than 2,500 chats: DMs and small groups in full, big groups and channels only on request. Then it just kept downloading. The laptop slept at night, and the daemon never crashed or restarted. After 23 hours the database held 822K messages, 1.2 GB on disk. The interesting part is why it took that long. Telegram never once asked it to wait while it downloaded history. The daemon spent 92% of its active time waiting on its own: by design it makes one call every three seconds, to be safe. On a test account Telegram starts throttling at about one call per second, so tg was about 3.5 times more polite than it needed to be. Next I'm making the pace adaptive: start faster, back off when Telegram asks. By my estimate an account like mine would then sync in 3–4 hours of active time instead of 11. Search is already fast: 40–75 ms per query over 290K messages, and prefix queries like word* match every form of a word. The day also found bugs you don't see on test accounts. After the laptop woke up, the daemon reconnected but didn't catch up, so one group was missing four overnight messages until a restart. Some history pages lost their senders, so about 9% of messages showed no name in search. And usage stats delayed every command by half a second. All three are fixed in 0.9.0, out yesterday. It also makes tg status show how much history is left and roughly how long it will take. And what a real account looks like: 1,112 of my chats have ten messages or fewer, while 19 chats hold almost half of everything. brew install neoromantic/tap/tg tg.sergeybaus.com/?utm_sourc…

Telesec, my Hermes Agent plugin for Telegram Business, keeps a short dossier on each contact: who they are, what you agreed. The model updates it as you talk; your own note corrects it. Opt-in, and the default model has zero data retention. telesec.sergeybaus.com

I have a channel for personal stories. Its title follows the city I’m in. Not the ticket, and not a layover. While I’m in transit, that isn’t the new place yet. When I arrived in Yerevan, the channel became about Yerevan. When I arrived in London, about London. The username, description and picture stay as they are. Only the line at the top changes: where I am now, not what the channel’s brand is. That feels calmer to me. The channel doesn’t pretend I live in one place if I’m already in another.

I prefer the word “robot” to AI. And I like “robopsychology,” with its nod to Asimov. “Agent” has its place too: someone acting on behalf of a person or a company. But the word itself feels temporary to me, as if it will soon run its course. I’m interested in the robot psyche: why a robot forgets its goal or escapes its constraints. Talking to it, studying its behavior, trying to understand why it did what it did. These are complex, tangled systems, and it isn’t always clear what happened inside them. So “robopsychologist” is a slightly tongue-in-cheek, but half-serious name for what I do. There’s another side to it: helping people understand robots and interact with them. Not addressing their personal psychological problems, but explaining what they’re dealing with and how to talk to it. I write about this in my English-language channel, [Robopsychica](t.me/SergeyBausAI).

Once again, I’m proving I’m not the Sergey Petrov on a sanctions list. This time, Y Combinator put my account on hold. Such is life when your name is Sergey Petrov.

telegram-sync-cli, my Telegram CLI, now works with files. tg messages shows each message's attachment, so an agent finds the PDF or voice note and gets it with tg media download (except chats with content protection). To send a file: tg send --file. github.com/neoromantic/teleg…

Telesec, my Hermes Agent plugin for Telegram Business, now tells you in your chat with the bot when a contact deletes a message. It needs history on: a button shows the deleted text from it. If they just rewrote it or sent it twice, it stays quiet. telesec.sergeybaus.com

telegram-sync-cli, my Telegram CLI, now keeps an eye on its cache size. tg status shows how much space the cache takes, the daemon deletes old bookkeeping records on its own (messages stay), and tg db vacuum gives the freed space back to the disk. tg.sergeybaus.com/?utm_sourc…

I added chat history to Telesec, my Hermes Agent plugin for Telegram Business: text with edits. If the other person deletes a message, you keep it and see if it looks like a correction or a duplicate. Search by person, words, and date. Opt-in, no model. telesec.sergeybaus.com

Glad you found me, @ditdot_info. Which post or topic made you want to follow?

Telesec, my Hermes Agent plugin for Telegram Business, lets you pick whose voice notes get transcribed in each chat: all, yours, theirs, or none. Tap Manage Bot in the chat; the buttons open in your DM with the bot. The other person sees nothing. telesec.sergeybaus.com

I made a Claude Code skill that reads lab results from Ornament Health: biomarkers, their history and what's out of range. Ask "how has my ferritin changed?" and get a table. You need your own Ornament account and API key. Free, MIT. github.com/neoromantic/ornam…

telegram-sync-cli, my Telegram CLI, can now connect through a proxy: SOCKS, HTTP or MTProxy, for when Telegram is blocked. And several accounts no longer get mixed up: each has its own cache, and the daemon syncs them separately. tg.sergeybaus.com/?utm_sourc…

I added costs to ai-usage 0.10.0: what a team pays for its Claude, Codex and Grok subscriptions, and how much of that it actually uses. For each subscription you see the price a month, how much of it was used over the last 4 weeks and by whom, in dollars. It's an estimate from how full the weekly limits got, not accounting. Subscriptions nobody used for a week are flagged with what they cost. Prices come from US list prices. What ai-usage can't know by itself, like the charge date or the price of a plan the list doesn't have, it asks, and you can skip. Answers are shared with the team sealed; the relay can't read them. The Mac app has the same, plus a notification 3 days before a charge. Install on macOS or Linux: brew install neoromantic/tap/ai-usage au.sergeybaus.com/?utm_sourc… The picture shows a made-up team.