Skip to main content
株式会社オブライト

Column

Useful articles about SEO, Web Development, and IT

387 articles

AI2026-06-11
DiffusionGemma Deep Dive — Google DeepMind's June 10, 2026 Open-Weight Text-Diffusion LLM, Same Backbone as Gemma 4 26B (A4B MoE), Up to 4× Faster Than AR Counterparts, Apache 2.0, With an Honest "Quality Trails AR" Disclosure
A primary-source deep dive on DiffusionGemma (google/diffusiongemma-26B-A4B-it, 25.2B total / 3.8B active MoE), released June 10, 2026 by Google DeepMind in coordination with NVIDIA. Grounded in the official Google blog, ai.google.dev model card, Hugging Face card, and NVIDIA's blog. Where autoregressive (AR) models generate one token at a time left-to-right, diffusion language models (DLMs) denoise a 256-token canvas in parallel into final text. 15-20 tokens commit per forward pass, up to 48 denoising steps, 1,000+ tok/sec on H100, 700+ on RTX 5090, ~3.5–4× the throughput of the AR Gemma 4 counterpart. Crucially, Google openly states that quality lags AR: MMLU Pro 77.6 vs 82.6, GPQA 73.2 vs 82.3, MMMU Pro 54.3 vs 73.8. Apache 2.0, distributed via Hugging Face / Vertex AI / NVIDIA NIM — the first large-scale open-weight diffusion LLM in the industry. The column covers practical implications for Japanese enterprises (on-prem internal agents, code editing, low-latency workflows) and positioning against Mercury (Inception Labs), LLaDA, and Gemini Diffusion.
Google DeepMindGemma 4DiffusionGemma+5
AI2026-06-10
Cognition AI's FrontierCode Explained: The Next-Gen Coding AI Benchmark That Asks 'Is It Mergeable?'
On June 8, 2026, Cognition AI unveiled FrontierCode — not a product, but a coding AI evaluation benchmark. It measures not just 'does it pass tests' but 'would an OSS maintainer actually merge this?' across six axes. This article covers its differences from SWE-bench Verified, the three-tier dataset (Diamond/Main/Extended), official results with Claude Opus 4.8 leading at 13.4% on Diamond, and its relevance to Japan's rigorous code-review culture.
Cognition AIFrontierCodeSWE-bench+4
AI2026-06-10
Apple AFM Core Advanced Deep Dive — How 20B Sparse MoE Brings Frontier AI to iPhone
AFM Core Advanced, the flagship of Apple's third-generation Foundation Models announced at WWDC 2026, packs a 20B-parameter Sparse MoE with Apple's proprietary IFP technology — enabling frontier-class on-device inference on iPhone 17 Pro. This deep dive covers architectural innovations, A19 Pro specs, device requirements, and the 'fully Apple designed' controversy around Gemini distillation.
AppleAFMApple Foundation Models+5
AI2026-06-10
Google NotebookLM 'Better Research' Update: Full Breakdown — Gemini 3.5 × Antigravity Turns It Into an Active Research Agent
On June 8, 2026, Google announced a major update to NotebookLM called 'Better Research.' The AI engine was upgraded to Gemini 3.5, and the new-generation coding agent framework Antigravity was integrated. Each workspace now gets a secure VM for code execution and diverse file output. The new agentic research feature lets NotebookLM autonomously discover and analyze primary sources from the web — no pre-uploaded sources required. This article covers the official benchmarks (65%+ win rate, 4x faster than rival LLMs), pricing tiers, competitive landscape, and practical use cases for Japanese enterprises.
NotebookLMGoogleGemini 3.5+3
Software Development2026-06-10
Complete Guide to Obsidian 2026 — Turn Knowledge into Assets with Local-First PKM
Obsidian is a local-first knowledge management tool developed by Canada-based Dynalist Inc. Since commercial use became fully free in February 2025, it has become the top PKM choice for Japanese business professionals — IT consultants, engineers, researchers, and legal practitioners alike. This guide covers everything from basic Vault setup and the new 2026 Bases feature to local LLM integration via Ollama, Zettelkasten and PARA workflows, and Japanese-language caveats, so you can start today.
ObsidianPKMKnowledge Management+5
AI2026-06-10
Claude Fable 5 and Claude Mythos 5 Deep Dive — Anthropic's New "Mythos-Class" Top Tier Announced June 9, 2026, at $10/$50 per 1M Tokens, With Auto-Fallback to Opus 4.8 and Project Glasswing Gating
A grounded read of Claude Fable 5 and Claude Mythos 5, officially announced by Anthropic on June 9, 2026 (release post, Fable, Mythos). The two share an identical base model — Fable ships publicly with safeguards, Mythos is restricted to vetted partners. The release introduces a new top tier above Opus / Sonnet / Haiku — the "Mythos class". Pricing is $10 input / $50 output per 1M tokens (under half of Mythos Preview); cybersecurity, bio-chemical, and distillation-attempt queries are detected and auto-routed to Claude Opus 4.8 (firing on under 5% of sessions); Hex's core analysis benchmark hit 90% for the first time ever; and Mythos 5 itself is gated to Project Glasswing-vetted cyber defenders and biomedical researchers.
AnthropicClaudeClaude Fable 5+4
AI2026-06-10
cmux Deep Dive: Manaflow's macOS-Native Terminal for Running AI Agents in Parallel
cmux by Manaflow (YC S24) is a macOS-native terminal built specifically for running multiple AI agents in parallel. Powered by Ghostty's libghostty in Swift/AppKit, it offers vertical tabs, notification rings, an embedded browser, and a socket API. This column covers its features, install, use cases, competitive landscape, and considerations for enterprise adoption in Japan.
cmuxManaflowAI Agent+5
Software Development2026-06-10
Nix vs mise Deep Dive 2026 — Choosing the Right Dev Environment Tool
A thorough 2026 comparison of Nix and mise — the two leading dev environment managers. We pit Nix's purely functional paradigm (Nixpkgs 140k+ packages, Flakes, the Lix fork, Determinate Systems) against mise's pragmatic design (single Rust binary, asdf successor, SOPS/uv integration, 29.3k GitHub stars). Includes a decision framework for Japanese enterprises from solo developers to large-scale long-term maintenance shops.
NixmiseNixOS+5
Web Development2026-06-06
Web Dark Patterns 2026 Deep Dive — The 16 Official Categories, AI-Era Variants, EU / US / Japan Regulation, and a Self-Audit Checklist
A comprehensive primary-source guide to web dark patterns (now formally "deceptive design patterns") covering the deceptive.design 16-category taxonomy, the $2.5B Amazon-FTC settlement of 2025, EU AI Act Article 5's ban on manipulative AI, Japan's revised Specified Commercial Transactions Act (Article 12-6), the April 2025 Japan Consumer Affairs Agency 102-site field study (preselection in 45, fake hierarchy/social-proof in 39 each), the 2026 wave of AI-era patterns (CDT's 37 manipulative chatbot designs, DarkBench, GPT-4o sycophancy, 4× growth in GenAI scam pages), and a 40-item six-category self-audit checklist tuned for Japanese enterprise teams.
Dark PatternsDeceptive DesignUX Ethics+5
AI2026-06-06
Liquid AI's New Japanese-Specialized Models — LFM2.5-1.2B-JP-202606 and LFM2.5-Audio-1.5B-JP, the MIT CSAIL Spinoff's June 2026 On-Device AI Drop
On-device deep dive on Liquid AI's two Japanese-specialized models released on Hugging Face in early June 2026: LFM2.5-1.2B-JP-202606 (language, 1.17B params, 32K context) and LFM2.5-Audio-1.5B-JP (1.2B language core + 115M FastConformer encoder, 24kHz, Speech-to-Speech). Grounded in the official model cards and Liquid AI blog. Covers the Liquid Neural Network-derived architecture (16 layers = 10 LIV convolution + 6 GQA), the sub-2B-class leadership benchmarks (JMMLU 54.19 / J-MIFEval 79.08 / J-GSM8K 62.20), audio ASR with CommonVoice 8 (ja) CER 4.42 — about half of Whisper-large-v3 — but trailing Whisper on JSUT and ReazonSpeech (domain-dependent gap), the non-Apache LFM Open License v1.0, hardware support across Apple Silicon / AMD Ryzen AI / Qualcomm / NVIDIA / mobile CPU, competitive positioning vs Gemma 4 12B / Qwen 3.5 / TinySwallow / Sarashina, and adoption guidance for Japanese on-device AI, call centers, meeting notes, and in-person retail.
Liquid AILFM2.5LFM+5
AI2026-06-05
The Complete Guide to Hermes Agent & Hermes Desktop Skills and Tools — 19,932 Skill Catalog, 40+ Built-in Tools, and the Use-Case Patterns That Matter (June 2026)
A comprehensive guide to the Skills & Tools system in Nous Research's open-source agent Hermes Agent v0.15.2 (and the Hermes Desktop GUI), grounded in official docs and GitHub releases. Covers Skills (on-demand procedural docs) with the three-level Progressive Disclosure loading scheme starting at ~3k tokens, the SKILL.md format, the skills.sh catalog that exploded from 858 to 19,932 entries in v0.15.1, the standout new skills (openhands, code-wiki, web-pentest), the self-improving loop where the agent creates / patches / edits / deletes its own Skills, and Tools — 40+ built-ins like web_search, x_search, terminal, patch, browser_navigate, vision_analyze, cronjob, memory, delegate_task. Also covers MCP client + server support, the macOS Computer Use background execution that doesn't move the cursor or switch Spaces (5–20ms/event), and the 25+ messenger gateway (Slack / Discord / Telegram / Teams / WhatsApp / LINE / Feishu / WeCom and more). Ends with eight category-specific combination patterns — research, writing, data analysis, coding, customer support, social listening, internal automation, personal work — sized for Japanese enterprise practice.
Hermes AgentHermes DesktopNous Research+5
AI2026-06-04
Gemma 4 12B Deep Dive — The Encoder-Free Multimodal LLM That Runs on a 16GB Laptop Under Apache 2.0 (June 3, 2026)
A deep dive into Gemma 4 12B, released by Google DeepMind on June 3, 2026, grounded in the official announcement and Developer Guide. The standout property is encoder-free multimodal architecture — replacing the prior vision encoder (~550M parameters) with a 35M-parameter lightweight embedder plus a single matrix multiplication, and removing the 12-layer Conformer audio encoder entirely by projecting raw audio straight into the LLM's embedding space. Runs on a 16GB VRAM laptop (Copilot+ PC or Apple Silicon Mac), shipped under Apache 2.0, available through Hugging Face / Ollama / LM Studio / MLX / Vertex AI on day one. Covers the architectural rationale, the "approaches 26B MoE at less than half the memory" benchmark claim, positioning within the Gemma 4 family (E2B / E4B / 26B / 31B), competitive comparison against Llama 4 / Qwen 3.5 / Phi-5, and the fit with Japanese enterprise on-prem AI, voice workflows, and data-sovereignty requirements.
Gemma 4Gemma 4 12BGoogle DeepMind+5
123456789101112131415161718192021222324252627282930313233