Dakarda AI Newsletter · 12 June 2026
Friday Edition
Claude Fable 5, DiffusionGemma, Oasis 3
Claude Fable 5 — a Mythos-class model — has debuted publicly. Google released DiffusionGemma, which generates text 4x faster. And startup Decart launched a photorealistic real-time driving simulator.
Intro · Alex
Hi, Alex here. I've tracked the major releases of the last 48 hours for you — and I can tell you this was one of those weeks where you feel the AI capability threshold has just shifted. We have the first public Mythos-class model from Anthropic — intricately named Fable 5, an open diffusion model from Google that breaks the text generation paradigm, and a simulator for training autonomous vehicles that looks like a next-gen video game. Plus two important regulatory moves in Europe and the US, and a warning from ACM computer scientists that AI agents are outpacing the law. All of this paints a picture of changes that will have a real impact on your projects in the coming months. In this edition, you'll find not just descriptions of the releases, but also their practical significance — what they change in everyday work with AI. There's also a tool of the edition and two cybersecurity stories that remind us that racing with technology comes at a cost. Let's dive in.
What's worth knowing
Anthropic releases Claude Fable 5 — the first Mythos-class model
Fable 5 surpasses previous Opus in software engineering, general knowledge, and vision. Equally important is the controversial Glasswing Program — Mythos 5 for cybersecurity experts — and the Guardians restrictions that activate on LLM-related topics. This is a qualitative leap, but with a taste of security policy.
Google DeepMind releases DiffusionGemma — a diffusion model 4x faster
Experimental 26B MoE (3.8B active) under Apache 2.0 generates text in parallel — 1000+ tokens/s on H100, fits in 18 GB VRAM after quantization. This is the first open-source model this fast that runs locally and offers a real alternative to classic transformers.
Decart launches Oasis 3 — a driving simulator for training robots and autonomous vehicles
A real-time world model generates photorealistic environments from a text prompt. It responds to user actions like a game — simulations can last for hours. It opens the door to cheap, scalable training in rare road scenarios, without physical risk.
From the tech world
Another attack on Microsoft packages — data theft using AI assistants
Cybercriminals for the second time in weeks compromised authenticated Microsoft open-source packages, installing advanced credential-stealing malware. The code activated when developers opened the packages using AI coding assistants. GitHub disabled the packages without warning — just an annotation 'terms of service violation'. A total of 73 packages with self-replicating code.
New wave of attacks on developers — SEO poisoning infects Gemini CLI and Claude Code installers
Researchers at EclecticIQ detected a cybercriminal campaign using SEO poisoning and typosquatting. Fake installation pages for AI tools (Gemini CLI, Claude Code) appear higher in search results than official ones. Upon execution — infection and data theft. Target: next-generation developers.
Tip of the day
How to test DiffusionGemma locally — step by step
DiffusionGemma is available on Hugging Face in two variants — 26B (requires ~80 GB) and 8B (requires ~18 GB). You need at least 24 GB VRAM (e.g., RTX 4090 or A6000). Google has released an official Colab notebook for quick testing. Practical step: go to the google-research/diffusion-gemma repo on Hugging Face, download the model in PyTorch format and run the `generate.py` script with the `--cache-dir` flag. For performance comparison, try the same prompt on Gemma 2 27B — you'll see a difference in generation time of up to 4x.
Tool of the issue
DiffusionGemma
Open diffusion model from Google DeepMind. Generates entire blocks of text in parallel instead of token by token. Runs on a single graphics card, available under Apache 2.0. Ideal for building applications that need fast, local text generation — without API fees.
Reading list
EU Cloud and AI Development Act (CADA) — how the cloud will change in Europe
Detailed analysis of the European Commission proposal — a four-level sovereignty framework, a shift in AI Act requirements, and the upcoming Article 50 (AI labeling) from August 2, 2026. Essential reading for companies building on cloud and AI in Europe.
ACM TechBrief #14: Agentic AI outpaces the law
The world's largest association of computer scientists warns that autonomous AI systems (browsing the web, executing code, managing files) are slipping beyond current regulations, including the EU AI Act. For agent builders — it's both an opportunity and a warning.
In a world where models leap past the threshold, the most valuable thing becomes understanding where that threshold leads.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.