iiinova — where your digital self keeps evolving

Micha the DevOp @michabbb.bsky.social · 06/10/2026

It runs on Together's serverless inference, which serves the largest share of #OpenRouter tokens for DeepSeek V4.1 Flash (40.8%), GLM 5.3 Flash (28.2%) and Kimi K3 (23.1%)

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

📊 A per-session tracker shows actual spend next to what the same session would have cost on Opus 5.5. Billing runs on the existing Together API key via serverless pay-as-you-go or credit packs, no separate contract.

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

With an Anthropic key it routes between Opus 5.5 and GLM 5.3, without one between GLM 5.3 and GLM 5.3 Flash. Routing happens once per session, so prompt caching keeps working

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🎯 "Auto" mode uses the Together AI Router: it reads each session's first task and sends quick fixes to fast, low-cost models and hard problems to frontier models.

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

⚡ Setup is one shell command: curl -fsSL link.together.ai/install | bash. The agent opens exactly as before, now connected to Together AI through your API key

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🛠️ Supported harnesses: #ClaudeCode, Claude Desktop, #Codex (ChatGPT app and CLI), #OpenCode and Pi. Settings and logins stay as they are, nothing new to learn, and switching back to the native closed models takes one command

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🧠 The reasoning: coding agents often run every task, from a one-line fix to a full rewrite, on the same premium model. Open models now handle many of the hardest coding tasks at a fraction of the per-token price, while smaller models like GLM 5.3 Flash and #DeepSeek V4.1 Flash cover everyday coding

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🔗 #TogetherAI launched Together Link, which connects the coding agent your team already uses to frontier open models such as #KimiK3 and #GLM 5.3. Same agent, same workflow, with model spend cut by over 50% according to Together # AI #opensource #LLM 🧵👇

120

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🔌 Released alongside: phpstan-strict-rules, phpstan-symfony and phpstan-phpunit 2.1.0 🌐 phpstan.org/blog/phpsta...

phpstan.org

PHPStan 2.3: Leap in Performance, Generics Improvements, Detecting Unused Variables, and More!

000

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🏗️ A single-pass inside-out engine rewrite replaces three AST traversals. $a = [$b = 1, $b + 1] now infers as array{1, 2}. New features require bleeding edge

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🧠 Closure types are now inferred from usages without PHPDocs, and static variable types from assignments instead of mixed

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🧹 Also covers array offsets never read, foreach keys and values, ++/-- results, catch variables, closure use variables and unused parameters of functions and private methods

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🔍 Unused variable detection with five core rules: variable.unused, assign.unused, assign.overwritten, assign.redundant and the data-flow based assign.unusedFlow

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🧬 Generics: bidirectional type narrowing infers template types from how objects are used after new Collection(...). Scalars stay precise (1, not int). Closes 17 issues at 1-3% cost. Funded by the #SovereignTechFund

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🗂️ New extension APIs let custom extensions describe precisely which files to invalidate. Monorepos with Composer path repositories and scanDirectories/scanFiles setups are now handled properly

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

💾 Warm result cache on #Drupal: no changes 1.6s, 10 changed files 2.8s, 10% of the project 6.2s, full run 22s. Composer dependency updates and #Symfony DI container changes no longer wipe the whole cache

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

📦 The Turbo .so/.dll files ship inside the #Composer package, so most users already run it automatically

100

Micha the DevOp @michabbb.bsky.social · 06/10/2026

⚡ PHPStan Turbo reimplements hot parts in #Cpp, forks workers instead of spawning them and adds an #OPcache optimizer pass that strips runtime type checks. It now makes PHPStan almost 60% faster

110

Micha the DevOp @michabbb.bsky.social · 06/10/2026

🚀 #PHPStan 2.3.0 is up to 7.5× faster than ten months ago: analysing #WordPress dropped from 60.5s to 7.9s #PHP #staticanalysis #opensource 🧵👇

120

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🆓 No GPU at hand? Tesseracted Labs hosts Kolibri-1 free for anyone to try over the next few days, no setup needed: tesseracted.com/kolibri-1-c... 🌐 tej.as/blog/aleph-...

tej.as

Aleph Alpha Kolibri: How the Sovereign German LLM Works

Kolibri is Aleph Alpha's open-weight German and English LLM: 78B parameters, 3.5B per token. How it works, what it's good and bad at, and when to use it.

000

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🔧 Runs on #vLLM with Aleph Alpha's inference plugin, exposing an OpenAI-compatible API. Needs 2x A100/H100 80 GB or 1x H200/B200/B300. Best fit: on-prem #RAG over long German documents like laws, contracts and manuals

101

Micha the DevOp @michabbb.bsky.social · 03/10/2026

⚠️ Weak spots: little closed-book knowledge (14.8% correct on Omniscience), multi-turn tool calling at 39.8 on BFCL, 27.7 on Terminal-Bench 2.1 vs 39.7 for Qwen3.5. The dense Qwen3.8 27B scores higher overall (80.2 EN, 79.9 DE) at ~8x the active parameters

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🤷 Trained with the Merlin-Arthur protocol to abstain: it says "I don't know" 44% of the time instead of hallucinating, vs 11.1% for Qwen3.5 35B-A3B

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🇩🇪 Reasons in German on German prompts, with 4 reasoning effort levels (none, low, medium, high). AIME 2025: 87.5 in German, 96.9 in English

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

📏 262,144 tokens of native context, validated up to 1,048,576. 40 of 50 layers use 512-token sliding-window attention, every 5th layer attends to everything. RULER at 1M tokens: 63.2 vs 57.5 for #Qwen3.5 35B-A3B

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🔤 New UniBPE tokenizer (128k vocabulary) built for German compound words: "Bundesverfassungsgericht" is 2 tokens instead of 6 with GPT-5's o200k_base. On the full German Basic Law it needed ~15% fewer tokens than GPT-5's tokenizer, and tied it in English

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🧠 50 layers, each with 384 experts plus 1 shared expert; a router sends every token to 6 of them. It computes like a 3.5B model but needs the memory of a 78B one: about 78 GB of weights in FP8

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🏗️ Trained on ~24T tokens (over a fifth German) using 768 NVIDIA B200 GPUs. Weights on Hugging Face

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🐦 #AlephAlpha released #Kolibri, an open-weight #LLM for German and English (Apache 2.0). #MixtureOfExperts: 78.1B parameters, only 3.46B active per token. Trained from scratch in Germany and Finland #SovereignAI #opensource #AI 🧵👇 🇩🇪

120

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🫠 I wasted hours figuring this out 😂 Garbage audio in, garbage transcript out. #AI #LocalAI #VoiceAI

000

Micha the DevOp @michabbb.bsky.social · 03/10/2026

✈️ Background noise kills accuracy. If your notebook sounds like it's about to take off 💨, don't be surprised when every model performs like crap 💩🤖

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

👂 Most importantly: listen to your own recording first 🎧 The model only hears what your mic captures.

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🔊 Increase the recording volume and enable any voice-boost features your OS offers. It can make a huge difference, especially with a webcam's built-in mic 📹🎤

100

Micha the DevOp @michabbb.bsky.social · 03/10/2026

🎙️ Quick tip if you're testing #STT / #SpeechToText models locally: before you think "what a stupid #LLM is this?" 🤦‍♂️🤖 — check your microphone. Your system may be recording from the wrong input device. 🧵👇 #ai

100

Micha the DevOp @michabbb.bsky.social · 30/09/2026

Everyone copied #Jev 💣 So I tested the open alternatives on long, messy real-world data One local setup matched Jev: 98.6 % / 100 %, zero false links 🏆 Which one, and why it may not be yours 👇 #LocalAI #LLM #OpenSource #ai #llm x.com/michabbb/st...

040

Micha the DevOp @michabbb.bsky.social · 28/09/2026

you can use #jev as a judge for which information inside your #knowledge #graph should be connected - it's extremely effective 😊 #ai #rag

001

Micha the DevOp @michabbb.bsky.social · 25/09/2026

🧩 First-party plugins: Blade Compiler to validate compiled output, #Statamic with 19 rules for Antlers and tags, and #NativePHP with 26 rules for native markup, theme tokens and icon labels. fortephp.com/docs/sheath...

fortephp.com

Get Started with Sheath for Laravel Blade » Forte

Install Sheath, run your first Laravel Blade lint, review its accessibility and security checks, and choose the next configuration steps.

000

Micha the DevOp @michabbb.bsky.social · 25/09/2026

📈 Existing projects adopt it gradually with inline suppressions and baselines. Output works for terminals, CI and machine-readable JSON, with custom reporters and custom rules for project-specific conventions.

101

Micha the DevOp @michabbb.bsky.social · 25/09/2026

🔧 Fixes are split into safe fixes (--fix) and dangerous fixes that need an explicit opt-in (--dry-run --dangerous previews both). The migration preset rewrites older Blade syntax in a single pass.

100

Micha the DevOp @michabbb.bsky.social · 25/09/2026

🎯 Presets: recommended (default), strict, stylistic, migration and empty. Presets can be combined and each rule gets its own severity. Unknown rule IDs or preset names fail the run, so CI never silently skips a rule.

100

Micha the DevOp @michabbb.bsky.social · 25/09/2026

⚡ Install via #Composer with composer require fortephp/sheath --dev, publish the config with php artisan vendor:publish --tag=sheath-config, then run php artisan sheath:lint on the configured paths, a directory or a single file.

100

Micha the DevOp @michabbb.bsky.social · 25/09/2026

📋 132 rules across accessibility, best practices, Blade, performance, security and SEO, 38 of them auto-fixable. Typical catches: images without alt text, buttons without an explicit type, and target="_blank" links that expose window.opener.

100

Micha the DevOp @michabbb.bsky.social · 25/09/2026

🗡️ #Sheath is a #Laravel-first linter for #Blade and #HTML templates, part of the #opensource #Forte toolkit by John Koster. It catches accessibility, security, correctness and style problems during development and in #CI. 🧵👇

130

Micha the DevOp @michabbb.bsky.social · 24/09/2026

🔐 Ships with zero credentials: bring your own free #GoogleCloud OAuth client for the YouTube Data API v3 (~5 minutes setup). Requires Python 3.9+ and is safe to point at an existing vault without overwriting content

100

Micha the DevOp @michabbb.bsky.social · 24/09/2026

♻️ No artificial caps: a channel with 500 videos in your window gets all 500. Both phases dedupe by YouTube Video ID, so re-runs never create duplicates

100

Micha the DevOp @michabbb.bsky.social · 24/09/2026

🗂️ ingest builds the vault: one note per video plus auto-created concept pages (Automation, CRM, Sales) and entity pages (Claude, GoHighLevel)

210