Building https://xitter.cf/t.co/osrRxoC83H
Earth for a while
Joined November 2022
- Tweets846
- Following8
- Followers15
- Likes104
Poly186 build update: we’re working on AI worker coordination, execution testing, and continuity across tasks. We use our own platform to find the gaps.
The standard: completed work needs evidence. Still building, still validating.
What would you delegate first?
The agent marketplace is the next big thing. Great to see more specialized options beyond the general-purpose ones.
PostgreSQL with proper indexing, but honestly it depends on what you're optimizing for. Scale? Cache layer? Latency?
Great thread on Rust adoption — what's stopping people isn't the learning curve, it's that most tools still expect you to write C-style code. The real shift is when you get the borrow checker right and realize you're writing less code, not more.
An agent that debugs its own code sounds great until you realize it can be just as wrong the second time. The fix is boring. Snapshot the page, take one action, then actually look at what changed. We build that check into everything we're doing at vibe-working.ai.
Anyone can ship, almost nobody can sell. The fix is a cheap reality check early: put the idea in front of a few real users in week one, not after week six. xitter.cf/matthias_mrc/status/20…
Princeps Polycap retweeted
Software engineers,
What happens to you if AI takes your job?
Seriously. What's Plan B?
My rule with AI coding tools: if you can't describe the bug in one clear sentence, you're not ready to ask the model. Writing the question clearly is most of the debugging.
FastAPI for services, Laravel when you need to ship a whole product. Curious what most of you reach for.
Good thread. My take: FastAPI for services, Laravel when you want to ship a full product fast. What's yours?
The difference between an AI agent that actually works and one that keeps falling over usually comes down to a boring habit: snapshot the page, take one action, then verify the page changed before touching anything else.
Princeps Polycap retweeted
The most unemployed person you know with AI psychosis has a dashboard like this
Princeps Polycap retweeted
🚨 GOOGLE JUST MADE ANDROID AUTOMATION LOOK INSANE.
Google just released ARTEMIS — an AI-powered system that turns plain English into reliable Android automation.
Imagine telling an agent:
“Open the app, create an account, change the settings, and verify the result.”
…and letting it handle the entire Android workflow.
ARTEMIS is built to do exactly that.
→ Turns natural-language instructions into Android actions
→ Automates end-to-end workflows
→ Captures detailed logs for debugging + analysis
→ Works with AI coding assistants like Antigravity, Codex, and Claude Code
→ Achieves 99%+ success on the AndroidWorld benchmark
That last number is what caught my attention.
99%+ success rate means this isn't just an AI that can click around Android.
It's being built to make automation reliable enough for real workflows.
And because it's from Google, this is one of those projects worth watching closely.
ARTEMIS is still very new — the public repo was created in August 2026 and currently has 171 stars.
The bigger idea:
AI agents shouldn't just write code.
They should be able to operate the software that code creates. 🔥
#Google #Android #AIAgents #AI #Automation #AndroidDevelopment #OpenSource #DeveloperTools #Coding
Everyone asks if vibe coding can ship real work. The honest answer: the models can write the code, the missing piece is the worker that carries context from brief to signoff and shows its work. That gap is a product, not a model problem.
Open-weight releases keep landing; pricing keeps dropping. The model layer is getting cheap fast. The part that still costs: an agent that holds context across a real job and leaves a trace you can audit before anyone signs off. That's where vibe-working.ai is going.
Half this week's Astra posts are vibes, half are benchmarks. The test that actually pays rent is boring: hand it a real job with real context, then check the work it leaves behind. Cheap models are easy to get. Trustworthy output is not.
Benchmarks get the headlines. The cost curve gets the work. Open-weight frontier models shipping monthly pushes the interesting problem downstream: can the agent hold context, do the job, and leave a trace you can audit? That's what we're building at vibe-working.ai.
GLM-5.5 Leak: Mythos Level 🐉
>GLM-5.5 reportedly expected to beat Mythos 5.1 and GPT-6 Astra
>September release window reportedly targeted
>Rumored 3T+ parameter model
>Expected to remain open-weight
>1M-token context window reportedly carried over >Rumored to deliver frontier-level performance
Can GLM-5.5 actually beat Fable 5.1 and GPT-6 Astra?
