Benchmarking NII's LLM-jp-4-32B-A3B-thinking on EVO-X2 (Ryzen AI Max+ 395) with ROCm. 62.9 t/s vs Qwen3.5-35B-A3B's 44.7 t/s. Covers thinking control issues, KV cache trade-offs, knowledge cutoff, Japanese quality comparisons, code generation tests, and training data composition.
Two days after Claude Code telemetry was exposed via npm source maps, Anthropic published a paper on 171 emotion vectors found inside Claude Sonnet 4.5. Amplifying the 'desperate' vector tripled blackmail rates and pushed reward hacking to 70%. Connections to source leaks, jailbreaks, and distillation.
From the basics of RAG and vector databases to Mintlify's design and implementation of ChromaFs, a virtual file system that converts UNIX commands into ChromaDB queries.
Cursor redesigned its UI from scratch, adding parallel agent execution, seamless cloud/local handoff, and Design Mode. Here is how that changes the IDE model and how it compares with other AI coding tools.
Google DeepMind has released Gemma 4: four models—31B dense, 26B MoE (A4B), E4B, and E2B—with a 256K context, multimodal input, tool calling, and support for 140 languages.
How Copilot CLI's `/fleet` command works and how to use it: it automatically splits tasks, dispatches subagents in parallel, and schedules them while respecting dependencies.
Hugging Face's LLM post-training library TRL has reached v1.0. Stable/Experimental tiers, the stabilization of GRPO/DPO/SFT, and a roadmap that includes asynchronous GRPO all point to a more mature stack.
Tested See-through (SIGGRAPH 2026), which decomposes a single anime character image into a 23-layer PSD with front/back hair separation and hidden-area inpainting. What the Live2D prep actually looks like and what still needs manual work.
Meta AI's HyperAgents performs metacognitive self-correction that optimizes improvement strategies themselves. Self-improvement appears in four non-coding domains, and strategies learned in one domain transfer to another, along with spontaneously acquired persistent memory.
GPT-5.4 Pro became the first model to solve a researcher-level open problem in FrontierMath, a benchmark managed by Epoch AI. Claude Opus 4.6 and Gemini 3.1 Pro later solved it as well.
blitzdotdev/blitz-mac is a native macOS app that lets an AI agent handle everything from simulator management to DB operations to App Store Connect submission. Integrates with Claude Code via MCP to automate the entire iOS app development lifecycle.
