Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model

Source: Simon Willison's Weblog | Tags: coding, model, performance, quantization | Published: 2026-04-22 | Novelty: 45%

Qwen3.6-27B showcases outstanding coding performance with a significantly smaller footprint at just 55.6GB, surpassing its predecessor Qwen3.5-397B-A17B on major benchmarks. The model was tested using the unsloth/Qwen3.6-27B-GGUF:Q4_K_M quantized version (16.8GB) with llama-server, generating detailed SVGs efficiently despite its reduced size.

It's all about the angle: Your photos, re-composed

Source: The latest research from Google | Tags: 3d, ai, google, photos, recomposition | Published: 2026-04-22 | Novelty: 42%

The new Auto frame feature in Google Photos uses machine learning and generative AI to re-compose photos from a different perspective after they are taken. By estimating the 3D scene, camera parameters, and faces' positions, it intelligently generates missing content to create a more flattering view of subjects, especially useful for portraits and reducing wide-angle distortions.

Microsoft is so fucking stupid

Source: §kuthus | Tags: gaming, microsoft, os, technology | Published: 2026-04-21 | Novelty: 37%

The article criticizes Microsoft for failing to capitalize on its strengths and instead tarnishing its reputation, particularly in gaming and operating systems. The author calls out the company for defending rather than addressing these issues, suggesting a significant decline from what was once a dominant brand with a strong consumer presence.

Where's the raccoon with the ham radio? (ChatGPT Images 2.0)

Source: Simon Willison's Weblog | Tags: ai, chatgpt, ham-radio, image-generation, racoon | Published: 2026-04-21 | Novelty: 35%

The article compares the performance of different AI models including ChatGPT Images 2.0, Claude Opus 4.7, Nano Banana 2, and Pro in generating a Where's Waldo-style image featuring a raccoon with a ham radio. The new model, gpt-image-2, successfully produced an image where a raccoon holding a ham radio was easily noticeable (bottom left), contrasting sharply with earlier models like ChatGPT Image 1 which struggled to generate such detailed images.

Is Claude Code going to cost $100/month? Probably not - it's all very confusing

Source: Simon Willison's Weblog | Tags: anthropic, claudocode, pricing | Published: 2026-04-22 | Novelty: 35%

Anthropic quietly updated their pricing page to make Claude Code exclusive to the $100/month and$200/month Max plans, but later reversed this change after significant backlash. Amol Avasare, Anthropic’s Head of Growth, clarified that this was a test for about 2% of new prosumer signups, though the update is no longer visible to the public.

Gemma 4 VLA Demo on Jetson Orin Nano Super

Source: Hugging Face - Blog | Tags: gemma4, llamacpp, nvidia, vla | Published: 2026-04-22 | Novelty: 34%

The article details a VLA demo for Gemma 4 using the NVIDIA Jetson Orin Nano Super, focusing on building and running llama.cpp locally to leverage its full performance. Notable points include using Q4_K_M quantization for optimal resource utilization, downloading specific models via Hugging Face, and configuring environment variables like LLAMA_URL, MIC_DEVICE, SPK_DEVICE, and VOICE.

Quoting Bobby Holley

Source: Simon Willison's Weblog | Tags: anthropic, cybersecurity, firefox, vulnerability | Published: 2026-04-22 | Novelty: 31%

The article highlights the integration of Anthropic's Claude Mythos Preview into Firefox, addressing 271 identified vulnerabilities. CTO Bobby Holley emphasizes the importance of dedicated focus and the potential for defenders to gain an upper hand in cybersecurity efforts. The release marks a significant step forward in software security.

Changes to GitHub Copilot Individual plans

Source: Simon Willison's Weblog | Tags: copilot, github, pricing, usage-limits | Published: 2026-04-22 | Novelty: 28%

GitHub has announced changes to the GitHub Copilot Individual plans, including tightening usage limits and pausing signups for individual plans. The new pricing scheme introduces token-based usage limits on a per-session and weekly basis due to increased compute demands from agentic workflows. Notably, Claude Opus 4.7 is restricted to the more expensive $39/month 'Pro+' plan, while previous Opus models are dropped entirely.