Tag: LLM
All the articles with the tag "LLM".
-
Kimi K2.7 Code Lands in GitHub Copilot: After Benchmarks and Buzz, a Chinese Model Finally Wins the Channel
GitHub Copilot added Kimi K2.7 Code to its model picker on 2026-07-01 — the first open-weight and first Chinese model there. Hand-verified data, plus the "open weights + Azure hosting + off-by-default" trust structure.
-
GLM 5.2 Beats Claude on a Cyber Benchmark: Is China's 'Scenario-Specific Superiority' the New Inflection Point?
Updated:Zhipu's GLM 5.2 hit 39% F1 on Semgrep's IDOR vulnerability detection benchmark, beating every Claude Code Opus version at roughly 1/6 the per-finding cost — while still trailing about 2 points on aggregate benchmarks. Kimi and Qwen reproduce the same pattern. Is "scenario-specific superiority" the 2026 inflection point for Chinese models?
-
AI Model Comparison Mid-2026 Sequel: 52 Days After blog166, Here's What I Got Wrong in May
A line-by-line self-audit 52 days after blog166: of 5 predictions, 2 retracted (the Sonnet/Opus advice is stale, the open-source gap was underestimated), 1 still pending (Anthropic pricing), 2 held up. Plus six months of real model-routing observations from this blog and the checkpoints for the next audit.
-
ZCode on HN's Front Page: When Silicon Valley Actually Read a Chinese Dev Tool's Docs
Zhipu's official GLM-5.2 harness "ZCode" hit the Hacker News front page (155/192 when I first opened it, 264/236 now). This isn't "yet another AI IDE" — it's the first time a Chinese dev tool has been seriously evaluated by mainstream Silicon Valley engineers. This post breaks down the timeline, what ZCode actually is (a desktop app, not a CLI), the three real objections on HN, and a 30-line TS snippet for wiring up GLM Coding API yourself.
-
Two Days with Claude Fable 5: The 5 Things Every API Integrator Actually Has to Change
Anthropic shipped Fable 5 on 6/9, swapping the Opus/Sonnet/Haiku naming for Fable/Mythos. But the things that actually force every Claude API integrator to touch code aren't the names — they're the new stop_reason refusal, the forced adaptive thinking, and the "you must wire up a fallback model" architecture. After two days driving it inside Claude Code, here's the integration-side detail you need.
-
Prompt, Context, Harness, Agentic: The Four Nested Layers of LLM Apps — and Knowing Which One You're Stuck In
Prompt engineer, context engineer, harness engineer, agentic engineer — these aren't four competing job titles. They're nested layers of concern, from a single instruction to an entire autonomous system. Understand how the four layers relate, and you'll know exactly which one you're optimizing every time you get stuck.
-
AI Model Comparison, Mid-2026 Edition: Two Months After blog080, the Model Layer Has Turned Over
blog080 was written in early March 2026. Two-plus months later, GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro have all shipped, and open-source flagships GLM-5.1/Qwen 3 Coder have closed the gap to within 5-15 points of closed models. This is the May update: what changed, and how to adjust your March picks.
-
AI Agent Persistent Memory Architectures Compared: File-Based vs Vector Retrieval, Benchmarked with a blog-preflight Subagent
I hooked the same Subagent up to both Claude Code's built-in file-based memory and mem0's vector retrieval, then compared token cost, recall quality, and cross-session learning. The result: concrete thresholds for which approach fits which data scale, plus a look at procedural memory—the weakest but most promising direction.