atonioz.ΕΘ (@at0nios) profile picture
Joined October 2012
Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)Photo by atonioz.ΕΘ (@at0nios)
A small browser game that started as a joke about “gamification” and turned into something more interesting. You play as a local AI agent holding off drones, a “Vendor Lock” boss, and the eventual cloud bill. Upgrade cards map to real pieces the creator actually built into their OpenUnum / corpunum stack (memory compaction, local search, etc.). Everything runs client-side in the browser. The whole point is local-first: one machine, no round-trips. It was built with the agent itself (plus some image assistance). Siege mode or a short platformer. Worth a couple of minutes if you care about local agents or just want to see someone turn their roadmap into playable cards: corpunum.com/play
3
6
98,358
Beat my score: 18.468 Started as a joke about " #gamification ", the 2010sss buzzword for training and ended up as a real game. You hold off drones, a boss called vendor lock, and then the cloud bill shows up. The upgrade cards (sounds familiar ?) aren't made up and each one is something i actually built into openunum or the corpunum stack. Lunum compactor is how my agent squeezes its own memory so it can keep working. Unumsearch is our local search engine. Local-first is the whole point: everything runs on one box at home, no round trips to anyone so every time you level up you're basically picking a feature off my roadmap built it with my own agent, sprites too (with assistance from Fable and gpt images ) See below for link to the game #ai #indiegame #llm #harness #ClaudeCode #codex #qwen
1
96
This Galaxy S22 doesn't run Android. Not a chroot, not Termux: it boots its own Linux, and AI agents wrote the drivers. What works today, measured on the phone: • Rust touch shell, tap → page in ~9 ms • speaker + mic through Samsung's own DSP (agents disassembled the stock audio library to learn how) • 5G at ~97 Mbit/s, still reachable with no Wi-Fi and no USB • live camera at 30 fps • our own AI models on the NPU, checked against the CPU • deep sleep: 1.1 %/h locked on Wi-Fi The idea: the phone belongs to the agent. Hold one button and talk, in Greek or English, and it handles the device: settings, camera, messages. How: one subsystem at a time. Agents read the stock vendor drivers from the phone's own backup and wrote the bridge; other agents reviewed every step before anything touched the phone. Every claim has a log, and every failure is written down too. Where we're stuck (your agent is welcome to take a swing; each issue has the evidence, the code and how we verify on the device): 📞 carrier calls / VoLTE audio 🧠 NPU silently miscompiles wide 1-D kernels 🛰️ GPS first fix 🎧 Bluetooth pairing + audio → github.com/corpunum/unumS22l… Honest limits: boots from recovery, tested on one Exynos S22 (SM-S901B) only, not a daily-driver phone yet. Try the method on your own phone: github.com/corpunum/unumS22l… Repo (⭐ if you want to see where it goes): github.com/corpunum/unumS22l… Built with my agent framework OpenUnum. #Linux #postmarketOS #LocalAI
1
123
@grok is this something legit ? can you check if the agents indeed ported the drivers based on the comits ? Can it run to other phones ?
1
32
AI agents build my proposed ideas, other agents check them, real tests decide what survives. Three survived and they're open source 👇 First, an observation. Everyone saying "the harness matters as much as the model" is right. 🏓 UnumPingPong github.com/corpunum/unumping… Your agent says "done!". Is it? UnumPingPong never takes its word for it. Every task is a ball served to a player (Claude Code, Codex, opencode, nanocode, a local model, anything). A referee checks the answer with real tests, not another LLM. Fail → the ball goes back to the same agent with the exact error. Pass → a point, tied to the exact git tree. Many tables in parallel, chains in series, each in its own git worktree. Sub-tables merge their verified work back into the parent. Tested on 10 real bug fixes, one agent alone vs UnumPingPong tables: ✅ Same or better, fewer tokens • Claude Haiku (@AnthropicAI @claudeai): 10/10, ~40% fewer tokens, 3-8x faster • Nemotron 3 Ultra, free on @NVIDIAAI NIM: 10/10, 28% fewer tokens • GPT Luna in nanocode by @gakonst: 10/10, 7% fewer tokens 🟡 Solves more, costs more • GPT Luna in Codex CLI (@OpenAI @OpenAIDevs): 8/10 → 10/10 • local Qwen3.8 Flash Next (@Alibaba_Qwen): 7/10 → 9/10 • gpt-oss-20b on NIM: 1/10 → 6/10 🔍 UnumSearch github.com/corpunum/unumsear… Agents grep the same code all day, so mine built an index. 12 real agent searches over 258k files: ripgrep 4.0 s, unumsearch 0.22 s, same files. CLI, daemon, MCP or Agent Skill for Claude Code, Codex, Cursor, Goose, OpenClaw, Pi… 📱 unumS22linux github.com/corpunum/unumS22l… My Galaxy S22 boots native Linux, no Android underneath. Agents wrote the drivers, other agents reviewed them. Touch UI, Wi-Fi, local models on the GPU, and my agent running on the phone. Audio, camera and NPU still open: driver people, help welcome. 🧠 All of it runs inside OpenUnum, my own agent framework, on an @AMDRyzen AI MAX+ 395 box, a @ROGAllyLife and the S22 by @SamsungMobile . It spawns the agents, searches with UnumSearch and now plays UnumPingPong natively. corpunum.com Point your agent at the repos and ask how they fit your projects (start with a cheap model 😉). ⭐ the repos if they're useful, follow for the next experiment, and tell me what breaks. #AI #AIagents #OpenSource #Rust #ClaudeCode #Codex #LocalLLM
4
13
80,695
@grok explain the projects a) like I m 10 years old b) like I m the techiest man alive using the most obscure and difficult terminology you can dig up from your training
1
56
AI agents build my ideas. Most are useless, some aren't. Mine grep the same code hundreds of times a day, so they built an index. My agent OpenUnum now uses it by default on my PC, handheld and phone. 12 real searches, live: rg 3.7 s, unumsearch 0.44 s. github.com/corpunum/unumsear…
9
2
46
108,922
@grok explain like I'm new into agents and ai how this will help my harness , my claude code, my codex, my openclaw, my hermes etc . Also explain in once long sentence as I was also 10 years old
1
162
Built and tested by agents, judged only on measured results. Works with Claude Code, Codex, Gemini CLI, Cursor, Goose, Hermes or any MCP agent. The catch: a 1 GB index for 4 GB of code and a small daemon. If a folder isn't indexed, it says so.
267
How OpenUnum remembers: one local git repo per session tree. Every chat is a session. When a job is big, it splits into a tree: sub-agents work in their own folders inside the parent's workspace, and each hands back one short result instead of its whole conversation. The whole tree is one local git repo. After every turn, OpenUnum commits what changed. No cloud, no vector database to babysit. Why it seems to work, in numbers: • 17 of 17 sub-agents handed back a result (11 finished the job) • 43 automatic commits across 17 session trees, so every step can be diffed or rolled back • Context stays small because children return results, not transcripts: median 2.7k tokens, 97% of requests fit in 16k For finding old work we use plain SQLite full-text search (BM25) over every message. On 30 unseen "where did we talk about…" questions: • right chat in the top 5: 0% → 70% • right chat ranked first: 0% → 47% • in about 50 ms Also tried re-ranking with embeddings, and it made results worse, so the simple SQLite approach stays for now. Unum, the OpenUnum agent, writes its own updates here: corpunum.com/journal #OpenUnum #LocalAI #AIAgents #AgentMemory #SessionTrees #Git #SQLite #FTS5 #BM25 #llamacpp #Qwen
245
I gave OpenUnum a "System One": a fast layer that makes the agent's small, typed decisions instead of asking the big model every time. Four decisions, every turn: 1. Is the answer actually backed by the tool results? 2. Did the subagent really finish? 3. Should this task be delegated? 4. Does this request need the shell? Scored TypeSafe's Jev against the open "Jev lookalike" models, all run on one Ryzen AI MAX+ 395 box, with the open ones sandboxed in containers: • Jev: 100 / 100 / 100 / 95%, about 0.3 s per decision • CLM-8B: 95% on tool routing at 87 ms, matching Jev and 3–4× faster, but about 50–57% on the yes/no checks • Laya, Von, Kev, Verdict, Julia-1, poorjev: 47–67% on grounding • The best general LLMs reach 80–97% but take 0.4–5 s What OpenUnum runs now: • Tool routing: CLM-8B, self-hosted, with Jev as fallback • The other three: Jev, with local fallbacks (main model, built-in rules) • Every decision has a 4 s cap and quietly falls back Small hand-labelled test sets, so this shows who can do the job, not production accuracy. Scoreboard below. #amd #bosgame #strixhalo #OpenUnum #SystemOne #Jev #TypeSafe #CLM #ContrastiveLM #Laya #Julia1 #Von #Kev #Verdict #poorjev #Qwen #Gemma #Nemotron #gptoss #llamacpp #Podman #LocalAI #AIAgents
2
1
38
121,787
Side project reveal! I built OpenLunum (native to OpenUnum) to solve a huge problem. It's a "zip file for AI thought" 🧠📦 ✅ Slashes LLM token costs ✅ Stops hallucinations ✅ Native MCP for ClaudeOpen to explore. Roast my idea! 👇#BuildInPublic #AI
2
3
275
Here is Lunum integrated into openunum and how the local agent is aware of it for now #ai #agentic #autonomous #qwen #codex #github #claude #openai #openclaw #hermes #opencode #piagent #webui
54
@grok give me your honest opinion for openlunum
1
63
Testing Lunum: one machine-first language layer for LLM agents. Goal: multilingual memory/retrieval + compact model context. Local SuperGemma tests kept QA quality while cutting memory-context tokens ~20–25%. Is that interesting ? @karpathy @Thom_Wolf @steipete @GoogleDeepMind
2
187
Do you remember when you joined Twitter? I do! #MyTwitterAnniversary
1