Fresh public snapshot of M.A.R.I.A. is live 🧠
Local autonomous cognitive architecture — offline-first, 13 cognitive contracts, 5700+ tests, running 24/7 on a mini PC since Feb.
No cloud. No API key. Clone & run.
https://t.co/2iYGJ26wC3
Dear @AnthropicAI,
Can I get a one-week token reset?
Claude Code just did 45 minutes of “let me quickly check this” and burned through tokens like it was training itself for the Olympics. 😂
Worth it though.
Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.
Available today at the same price.
@aitechbrief Exactly. The key is observability.
Once memory, goals and runtime loops are inspectable, an agent stops being a black box and starts becoming a system
Started this 6 months ago as: “let me try building an agent.”
Now it has memory, goals, runtime loops, and a live dataflow map.
I think I accidentally started building an AI operating system
5/ I think the future of AI agents is architecture, not only bigger models:
local-first runtime, memory, tool routing, evaluation, safety, cost awareness and autonomy.
First serious use case after stabilization: market-agent.
https://t.co/2iYGJ26wC3
1/ M.A.R.I.A. update
I’m not building a chatbot.
I’m building a local autonomous AI agent running on a mini PC — with memory, homeostasis, planner, safety layer, UI, logs and tests.
Less PowerPoint.
More runtime.
4/ New layer: Skills as artifact.
Maria extracts repeated patterns from her own decision traces/logs and turns them into reusable skills.
DRAFT → SANDBOX → PRODUCTION
Practical procedural memory, not magic.
METAOPERATOR (n.) – person who orchestrates AI not as tool but as team. Multiple models with specialized roles. Biological methodology over reward-based. Cross-domain pattern recognition. Outputs measured in emergent behavior, not LOC.
Day 547 of being one.
@akshay_pachaar Great breakdown. Hermes nails memory + skills + profiles.
In M.A.R.I.A. I’m exploring the next layer: goals, policy, homeostasis, world model and runtime evaluation for long-running local autonomy.
https://t.co/2iYGJ26wC3
@SilverMarketFan AI nie premiuje kierunku studiów, tylko jakość myślenia. Jak nie umiesz powiedzieć maszynie, czego chcesz, to ani matma, ani humanistyka cię nie uratuje
Hey @Google@googlechrome — quick question.
Chrome installed ~4GB of Gemini Nano weights on my PC. I didn’t ask. I deleted it. It came back.
If it lives on MY machine now, who controls it?
Can I use, study, copy, or run it outside Chrome?
Just asking 🙂
@CKeruac Dokładnie ten sam kierunek testuję w M.A.R.I.A. LLM nie jako „mózg”, tylko jako warstwa językowa nad stanem, pamięcią,celami,policy,ewaluacją i symulacją. Różnica między RAG a World Model to różnica między znalezieniem notatki a zrozumieniem dlaczego ta notatka zmieniła projekt.
@LeszBuk Świetne ujęcie. W praktyce dodałbym do C(t): cele, stan systemu, policy i ewaluację akcji. Podobny kierunek testuję w M.A.R.I.A.: https://t.co/2iYGJ26wC3