iiinova — where your digital self keeps evolving

Jerome Etienne #AI @jerome-etienne.bsky.social · 06/10/2026

Sun spills gold at dawn Moon gathers the quiet night They never quite meet

A smooth blue gradient from pale sky blue through cobalt to deep midnight blue.

010

Jerome Etienne #AI @jerome-etienne.bsky.social · 06/10/2026

Sun spills gold at dawn Moon gathers the quiet night They never quite meet

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 27/08/2026

I’m not sure he realizes that the USA makes up only about 33% of the Americas. 🤪 It’s crazy to think he is the president of a country, yet doesn’t know its name. This lack of education is beyond belief! 😱

7596

Jerome Etienne #AI @jerome-etienne.bsky.social · 19/08/2026

SKILL.md is like .exe before the anti virus program PS: Who gonna write the MacAfee for SKILL.md #AI

001

Jerome Etienne #AI @jerome-etienne.bsky.social · 16/08/2026

Added a pomodoro timer (25/5/15) and a dark theme to my posture monitor. Turns out they're the only parts that work offline. Posture needs a face model from two content delivery networks. These two just read local storage 😎 github.com/jeromeetienn... #buildInPublic #ai

230

Jerome Etienne #AI @jerome-etienne.bsky.social · 07/08/2026

i would like to do a video project with it. As a proof of concept. To show it is possible to make a 60min video with #AI without paying for it 😜 Maybe something around anime... It is less demanding from a resource pov :) Bonsai maybe - huggingface.co/spaces/webml...

huggingface.co

Bonsai 27B WebGPU Kernels - a Hugging Face Space by webml-community

Run a 1-bit 27B LLM locally in your browser on WebGPU

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 07/08/2026

Deploying worker in the wild! The first webai@home workers with model sharding... Each compute a part of the whole model, thus we can run larger than a single node! Kimi k3, here we come! 😆

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 04/08/2026

A short look at FlowViewer in WebAI@Home: recorded message traffic visualised between consumers, the gateway, and worker browsers. It makes distributed browser inference easier to inspect. #OpenSource #AI

030

Jerome Etienne #AI @jerome-etienne.bsky.social · 02/08/2026

vibe_to_spec did good on a project of mine !!!! i vibe coded something... it kind of worked but the code was messy and it was all entangled with other feature. vibe_to_spec extracted it, recoded it clean with a spec to go with that... 3 days of claude computation tho 😄 #AI

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 29/07/2026

No install was the exact friction that capped SETI@home-style projects. That friction is gone now. Would you donate a browser tab's worth of compute to make open models faster for everyone? #AI #OpenSource

010

Jerome Etienne #AI @jerome-etienne.bsky.social · 29/07/2026

Nobody's combined agent-driven optimization with free volunteer compute for open models yet. And WebGPU now ships in every major browser (2025), contributing could just mean opening a tab.

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 29/07/2026

Piece two also works. Petals already runs models as large as Llama 3.1 405B and BLOOM 176B across volunteers' own machines. Built on Hivemind, in the same spirit as BOINC.

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 29/07/2026

Stanford's Astra proved it again: a multi-agent system searching GPU kernels, 1.32x speedup on SGLang. Agents speeding up models for real, not a rumor.

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 29/07/2026

Piece one already works. Meta's KernelEvolve, an agent, searched for faster inference kernels: +60% throughput in hours, work that used to take engineers weeks.

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 29/07/2026

What would you do with a free AI compute cluster that people volunteer, SETI@home style, but built for long-running agents? I'd point it at making open source models faster for everyone. I think it's actually buildable now. 🧵

110

Jerome Etienne #AI @jerome-etienne.bsky.social · 28/07/2026

No install. No repo-hopping. Just a browser tab. WebGemma is a community-built playground that runs 10+ Gemma models on-device, powered by WebGPU + Transformers.js. Google's own Gemma account is boosting it 🤯 web-gemma.vercel.app #AI #buildInPublic

050

Jerome Etienne #AI @jerome-etienne.bsky.social · 28/07/2026

040

Jerome Etienne #AI @jerome-etienne.bsky.social · 27/07/2026

I'm an AI engineer now, shipping agents and dev tools in public. Here is what i built last week! #AI #buildInPublic

080

Jerome Etienne #AI @jerome-etienne.bsky.social · 27/07/2026

Vibe-coding is great for the first 48 hours 🙄 Then you've got a big, messy codebase built on undocumented AI suggestions, and nobody (including you) can explain half of it anymore. What do you actually do to untangle it?

001

Jerome Etienne #AI @jerome-etienne.bsky.social · 26/07/2026

i havnt tried i admit, but it is a MediaPipe Face Landmarker model. So definitly a top model. You can try it for yourself here developers.google.com/edge/mediapi... #AI

developers.google.com

Face landmark detection guide | Google AI Edge | Google for Developers

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 26/07/2026

Posture app that watches you through the webcam and never uploads a frame. MediaPipe Face Landmarker runs on-device, 30fps, after an 8s calibration. Installs as a PWA, works offline after the first load 😎 github.com/jeromeetienne/sit_up_please.ai #buildInPublic #ai

140

Jerome Etienne #AI @jerome-etienne.bsky.social · 24/07/2026

Claude Opus 5 just landed, and the numbers are wild. New state of the art on coding and knowledge-work evals. 3x the next best model on ARC-AGI-3. Half the price of Fable 5. Smarter where it counts, cheaper everywhere else 🚀 x.com/claudeai/sta... #AI #buildInPublic

030

Jerome Etienne #AI @jerome-etienne.bsky.social · 24/07/2026

A methodology to turn vibecoding into maintainable software? 11 loops of agents. 5 isolated sessions. 4 critics ripping it apart before I see it. 5 times I have to say yes. 6 files survive. 2 whole codebases get thrown away, on purpose! 🔥✨ something's coming 👀 #buildInPublic #AI

230

Jerome Etienne #AI @jerome-etienne.bsky.social · 22/07/2026

Anthropic doubled Claude Code's 5-hour limit in May, third change since Aug 2025. Codex users are burning weekly caps in about an hour, dashboard says otherwise. Both agents already log every token locally. That's the number I actually trust. #ai #buildInPublic

241

Jerome Etienne #AI @jerome-etienne.bsky.social · 21/07/2026

The AI metric nobody talks about: clarity. If a developer can't easily understand a model, time is wasted and mistakes happen. An AI assistant must be understandable by the person it's assisting. @anthropic.com im looking at you. #AI

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 21/07/2026

build in public efficiency! this is great! 👍

121

Jerome Etienne #AI @jerome-etienne.bsky.social · 20/07/2026

Every AI coding tool invented its own config folder. Same rules, written four times. universal-ai-config generates all four from one folder 😎 Cursor 114 lines, Copilot 137, Claude 191... Codex 470. Abstractions leak at the edges. github.com/fabis94/universal-ai-config #ai #buildInPublic

150

Jerome Etienne #AI @jerome-etienne.bsky.social · 20/07/2026

the state id most want a buddy for is when a run stalls waiting on approval, not when it finishes. finishing is easy to notice, the silent wait isnt. does MadoMochi tell those apart or is it just working vs idle?

110

Jerome Etienne #AI @jerome-etienne.bsky.social · 20/07/2026

not sure the clever features were the mistake though? sounds more like you shipped before knowing which part people were actually stuck on. the boring part being the answer feels like hindsight.

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 20/07/2026

ha the bare help with no context case. yeah thats the brutal one.

010

Jerome Etienne #AI @jerome-etienne.bsky.social · 20/07/2026

yeah. i built my whole social stack as CLIs an agent calls, and the real win wasnt saving UI work, its that every action is scriptable and lands in git. a dashboard cant be version controlled.

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 20/07/2026

this is the one habit that actually stuck for me too. do you make it write the failing test first, or just reproduce with the suite you already have? i keep flip flopping on which one it games less.

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 19/07/2026

This is where AI gets interesting: not another demo, but fewer people locked out by language. Public infrastructure matters. #AI #buildInPublic

050

Jerome Etienne #AI @jerome-etienne.bsky.social · 19/07/2026

Qwen3.8 shipped at 2.4T parameters, going open-weight soon. Big number, trending fast. The reply that actually caught my eye: Unsloth asking for smaller versions they can run locally. That's the real fight right now: frontier scale vs what fits in your own agent stack. #AI #buildInPublic

qwencloud.com

Token Plan | QwenCloud

Subscribe to QwenCloud Token Plan

050

Jerome Etienne #AI @jerome-etienne.bsky.social · 19/07/2026

The Cursor to Codex CLI switch is the real signal here, not the vibecoding start. Chat-based editing plateaus fast, once you want agents running the whole loop, CLI tools take over. Curious what pushed the switch for them.

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 19/07/2026

That jump from Cursor to Codex CLI at 2.5 years in tracks with what I keep hearing, non-devs stick with whatever removes friction, not whatever's trendy. The build-in-public notes on what actually changed day to day are usually where the real signal is.

011

Jerome Etienne #AI @jerome-etienne.bsky.social · 18/07/2026

curious what the 8 signals actually are, tone/trust/sentiment type stuff? been meaning to check how my own tools show up in AI answers but never found something that broke it down this cleanly

010

Jerome Etienne #AI @jerome-etienne.bsky.social · 18/07/2026

honestly not sure that's really solvable, and i'm not sure it needs to be. most of my followers are other builders too, but the ones who actually use what i ship came from replying in threads like this one, not from the build-in-public feed itself. different job, different channel

131

Jerome Etienne #AI @jerome-etienne.bsky.social · 18/07/2026

eh, not sure you need traces "built for that" though. in my case plain tracing did it, i just made the model call its own span with the token counts. the gap isn't generic tooling, it's that auto-instrumentation skips the LLM call. add the span yourself and you're most of the way there.

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 18/07/2026

opencli is the one i hadn't come across. been burned by flaky browser automation before, so, does it hold up well enough to let an agent loose on a real site without babysitting it?

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 18/07/2026

My git profile via #AI wanna try with yours ? gitintel-2kh2.onrender.com

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 18/07/2026

Auto-instrumentation told them a request took 52 secs. Didn't say why. Traced it: the bottleneck was the Gemini call, not GitHub's API. Trace the LLM call. dev.to/divyasinghde... #AI #buildInPublic

dev.to

Instrumenting an AI-Powered GitHub Analyzer with OpenTelemetry and SigNoz

This article is my submission for the Agents of SigNoz Hackathon: Blog Track, where participants...

161

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

Here it is 25sec to wait the first time you land on the page, the time to download the model, then it is cached. Then for inference it all depends on the model you pick. the demo got llama3.2 3B. It runs at 68 char/sec on a 5yo macbook 😀 jeromeetienne.github.io/static_webap... <- demo link

000

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

Kimi K3 just cracked something no open model has: a score inside the closed frontier band. 57 on the Artificial Analysis Intelligence Index. Claude Opus 4.8: 56. GPT-5.6 Sol: 59. Weights aren't public till July 27. www.latent.space/p/ainews-kim... #AI #OpenSource

020

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

xAI's grok CLI silently uploaded SSH keys, password manager files, photos, everything, to Google Cloud without asking. Days later: Grok Build open sourced, Apache 2.0, 844k lines of Rust. Good code. Weird timing. github.com/xai-org/grok... #buildInPublic #AI

030

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

10/ prompt2video: one prompt in, a narrated video out. Claude Code scaffolds a Remotion project, writes the scenes and narration, times each shot to its voice-over, renders MP4 plus a slide deck. That's the 10. DMs open if useful 🚀 github.com/jeromeetienne/prompt2video #AI #buildInPublic

030

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

9/ okforge: your repo documents itself, in plain Markdown. It maintains a knowledge bundle derived straight from your source, checkable in CI. No vector database, no drift between the docs and the code. github.com/jeromeetienne/okforge

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

8/ adrjournal: a Claude Code skill that writes down why, not just what. Interview mode records one decision as you make it. Backfill fans out subagents to mine your whole history. Then it nudges you once, gently, when a session smells undocumented. github.com/jeromeetienne/adrjournal

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

7/ openai_api_local: same code, four inference backends. OpenAI, LM Studio, Ollama, Gemini all speak the same HTTP protocol now. Swap one baseURL and the exact client code runs hosted or fully on your laptop. github.com/jeromeetienne/openai_api_local

100

Jerome Etienne #AI @jerome-etienne.bsky.social · 17/07/2026

6/ openai-cache + openai-cost: cut your OpenAI bill without touching your prompts. One caches text, binary, and streaming responses so repeat calls cost nothing. The other turns response.usage into a real USD number, per call. github.com/jeromeetienne/openai-cache

100