<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Piotr Sobolewski — Blog</title><description>I design and ship systems built on language models and AI agents, and I am currently building a harness for AI agents. I write about what actually works in practice, and what is just noise.</description><link>https://sobolewski.pro/</link><language>en-gb</language><item><title>DeepSeek DeepInfer: a production inference engine goes open source</title><link>https://sobolewski.pro/en/blog/deepseek-deepinfer-open-source/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/deepseek-deepinfer-open-source/</guid><description>DeepSeek DeepInfer, the engine behind V4.1 Flash, is now Apache 2.0. Here&apos;s what&apos;s inside, what&apos;s still missing and who can actually run it.</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>open-weight</category><category>DeepSeek</category></item><item><title>Gemini Minecraft agents: 4 AI agents beat the Ender Dragon</title><link>https://sobolewski.pro/en/blog/gemini-minecraft-ender-dragon/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/gemini-minecraft-ender-dragon/</guid><description>Gemini Minecraft agents: four Gemini 4 Argon agents went from a random spawn to killing the dragon in 27 hours, with no human help. Here is how they did it.</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>AI agents</category><category>Gemini</category></item><item><title>Mercury 3 diffusion LLM hits 1,000 tokens per second</title><link>https://sobolewski.pro/en/blog/mercury-3-diffusion-llm/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/mercury-3-diffusion-llm/</guid><description>The Mercury 3 diffusion LLM from Inception: over 1,000 tokens/s, 1M context and $0.20 per 1M tokens. I explain how it works and how to try it.</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>API</category><category>AI models</category></item><item><title>AI cyberattack: agents broke into more than 100 companies</title><link>https://sobolewski.pro/en/blog/ai-cyberattack-cairn/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/ai-cyberattack-cairn/</guid><description>An AI cyberattack hit more than 100 companies, including Deere and SharkNinja: agents in Cairn worked almost alone, at about $50 per successful break-in.</description><pubDate>Mon, 19 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>AI agents</category><category>Claude Code</category></item><item><title>MiniMax M3.1 Flash: the &quot;Space Bunny&quot; steps out of the shadows</title><link>https://sobolewski.pro/en/blog/minimax-m3-1-flash/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/minimax-m3-1-flash/</guid><description>MiniMax M3.1 Flash is the mysterious &quot;Space Bunny&quot; from OpenRouter. Let&apos;s check the price ($0.15 per 1M tokens), the benchmarks and how to try it today.</description><pubDate>Mon, 19 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>AI agents</category><category>API</category></item><item><title>Google Willow and &quot;magic states&quot;</title><link>https://sobolewski.pro/en/blog/google-willow-magic-states/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/google-willow-magic-states/</guid><description>Google Willow showed magic state &quot;cultivation&quot; in practice: 40x fewer errors and 99.99% fidelity. I explain what magic states are and what this changes.</description><pubDate>Sun, 18 Oct 2026 00:00:00 GMT</pubDate><category>Google</category><category>quantum computing</category></item><item><title>Claude Code Review leaves preview: PR reviews 3x faster</title><link>https://sobolewski.pro/en/blog/claude-code-review/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/claude-code-review/</guid><description>Claude Code Review is now generally available on Team and Enterprise: PR reviews in 6–10 minutes, Quick, Standard and Deep levels. How it works and how to turn it on.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>Claude</category><category>Anthropic</category><category>AI agents</category></item><item><title>Quantum circuits designed by Claude: up to 17x fewer gates</title><link>https://sobolewski.pro/en/blog/claude-quantum-circuits/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/claude-quantum-circuits/</guid><description>Quantum circuits designed by Claude Mythos 5 beat the best published designs in 7 of 8 tasks, including 17x fewer gates for elliptic curve point addition.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>Anthropic</category><category>Claude</category><category>Mythos</category></item><item><title>OpenCode API key leak: OpenAI, Anthropic and Google keys exposed</title><link>https://sobolewski.pro/en/blog/opencode-api-key-leak/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/opencode-api-key-leak/</guid><description>The OpenCode API key leak hit 1.4% of Console users and their OpenAI, Anthropic and Google keys. What we know about the attack and what to do with your keys now.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>AI agents</category><category>API</category></item><item><title>Reflection Beam open weights: a 501B MoE with 1M context on Hugging Face</title><link>https://sobolewski.pro/en/blog/reflection-beam-open-weights/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/reflection-beam-open-weights/</guid><description>Reflection Beam open weights are out: 501B parameters, 23B active, Apache 2.0 license and a 1M-token context. What it can do and how to try it.</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>open-weight</category><category>AI agents</category></item><item><title>AI news this week: Gemini 4 Argon, Haiku 5.5 and Muse Spark Open</title><link>https://sobolewski.pro/en/blog/ai-news-2026-10-16/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/ai-news-2026-10-16/</guid><description>AI news this week (Oct 10–16, 2026): Gemini 4 Argon, Claude Haiku 5.5 and fast mode, Muse Spark open weights, Krea 3, OpenCode 2.0 and the OpenAI bugging case.</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>Weekly roundup</category><category>Google</category><category>Anthropic</category><category>OpenAI</category><category>Meta</category></item><item><title>Claude Code tutorial: from zero to your first task</title><link>https://sobolewski.pro/en/blog/claude-code-tutorial/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/claude-code-tutorial/</guid><description>A Claude Code tutorial for beginners: one-command install, login, your first task in the terminal, the CLAUDE.md file and the commands worth knowing.</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>Claude</category><category>Claude Code</category><category>AI agents</category></item><item><title>Claude fast mode: Opus 5.5 up to 2.5x faster, no waitlist</title><link>https://sobolewski.pro/en/blog/claude-fast-mode-opus-5-5/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/claude-fast-mode-opus-5-5/</guid><description>Claude fast mode for Opus 5.5 is out of research preview: up to 2.5x faster output, $8/$40 per million tokens, no waitlist. Here is how to turn it on.</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>Claude</category><category>Anthropic</category></item><item><title>Claude Mythos at Bank of America: 1,000 contractors patch flaws</title><link>https://sobolewski.pro/en/blog/claude-mythos-bank-of-america/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/claude-mythos-bank-of-america/</guid><description>Claude Mythos at Bank of America: so many flaws that the bank hired 1,000 contractors to patch them. What the CEO said, what it costs and what Mythos is.</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>Anthropic</category><category>Claude</category><category>Mythos</category></item><item><title>DeepSeek free: V4.1 Flash in OpenCode at no cost</title><link>https://sobolewski.pro/en/blog/deepseek-free-opencode/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/deepseek-free-opencode/</guid><description>DeepSeek free in OpenCode: V4.1 Flash with a 200K-token context no longer uses credits. What it can do, where the catches are and how to turn it on.</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>AI agents</category><category>open-weight</category></item><item><title>What&apos;s new in Krea 3: 2K photo edits in under 8 seconds</title><link>https://sobolewski.pro/en/blog/krea-3-whats-new/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/krea-3-whats-new/</guid><description>What&apos;s new in Krea 3: 2K photo edits in under 8 s for $0.06, up to 6 reference images and LoRA training. How it compares to Nano Banana Pro and how to try it.</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>image generation</category><category>krea</category></item><item><title>AI Act Enforcement Begins: Brussels Questions OpenAI and Anthropic</title><link>https://sobolewski.pro/en/blog/ai-act-enforcement-openai-anthropic/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/ai-act-enforcement-openai-anthropic/</guid><description>AI Act enforcement has started: the EU AI Office sent OpenAI and Anthropic requests for information after agent incidents. What it means for companies using AI.</description><pubDate>Thu, 15 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>AI Act</category><category>regulation</category></item><item><title>GPT-4o API deprecation: GPT-4o and 4.1 shut down on January 14, 2027</title><link>https://sobolewski.pro/en/blog/gpt-4o-api-deprecation/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/gpt-4o-api-deprecation/</guid><description>GPT-4o API deprecation: on January 14, 2027 OpenAI shuts down GPT-4o, GPT-4o mini and GPT-4.1. Check if it affects you and how to move to the new model.</description><pubDate>Thu, 15 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>OpenAI</category><category>API</category></item><item><title>Muse Spark Open for business: license, hardware and costs</title><link>https://sobolewski.pro/en/blog/muse-spark-open-business/</link><guid isPermaLink="true">https://sobolewski.pro/en/blog/muse-spark-open-business/</guid><description>Muse Spark Open for business: Meta released the weights of its 816B model. What the license allows, what hardware you need and what the API costs. No hype.</description><pubDate>Thu, 15 Oct 2026 00:00:00 GMT</pubDate><category>AI</category><category>meta</category><category>open-weight</category></item></channel></rss>