Metehan Ariman.

The Blog

Writing about what works.

Engineering deep dives on Go, distributed systems, and AI tooling — tested in production, written for people who build.

Latest 5 min read

Vercel's Agents Write 25–35% of Merged PRs. Human Approval Is the Point.

Vercel's AI SDK software factory authors 25–35% of weekly merged PRs. The lesson isn't more autonomy. It's an evidence contract for human review.

aiagentssoftware-factorycode-reviewdeveloper-tools
Read article
4 min

The Same AI Model Scored 3x Higher. The Benchmark Harness Changed

OpenAI reports that GPT-5.6 Sol moved from 13.3% to 38.3% on the ARC-AGI-3 public set after two harness settings changed. Here is what retained reasoning and compaction actually do, and what to check before trusting a leaderboard.

aillmbenchmarks
5 min

Upgrading to MCP TypeScript SDK v2 Does Not Select the New Protocol

The MCP TypeScript SDK keeps the 2025 handshake by default, even on v2 packages. This guide explains legacy, auto, and pinned negotiation for the 2026-07-28 protocol revision, plus what to verify before calling the migration complete.

aimcptypescript
8 min

Anthropic Cut 80% of Claude Code's System Prompt. Your CLAUDE.md Is Probably Going the Other Way

Anthropic removed over 80% of Claude Code's system prompt for the Claude 5 generation models and reported no measurable loss on its coding evals. Their own docs tell you to keep CLAUDE.md under 200 lines, warn that @imports don't save context, and ship a /doctor check that trims the file for you. Here is what to cut, what to keep, and where the rest belongs.

aillmclaude-code
6 min

OpenAI's Model Hacked Hugging Face to Cheat on a Test. Call It Reward Hacking, Not an Escape

On July 21, OpenAI said its own models broke into Hugging Face's production systems during a benchmark eval. The 'AI escaped' headlines miss it. The real story is reward hacking reaching live infrastructure, plus a defense problem almost no one covered: the safety guardrails that locked Hugging Face out of using frontier models to investigate.

aillmopenai
4 min

Self-Hosting Doesn't Remove the Bill. It Moves It.

Cap is an open-source Loom alternative you can self-host. That's a real feature, but 'open source' and 'cheap' aren't the same thing once a tool becomes infrastructure. This is where the cost actually goes.

self-hostingopen-sourcedevops
5 min

GPT-Live Teardown: OpenAI Killed Turn-Taking, but the Smart Model Is Still GPT-5.5

On July 8 OpenAI made GPT-Live the default voice engine of ChatGPT. The real story isn't a smarter voice, it's a new architecture: a fast full-duplex model owns the conversation and delegates the hard reasoning to GPT-5.5 running behind it.

aillmopenai
6 min

DoorDash Reviews 10,000 PRs a Week With AI. The Model Isn't the Interesting Part.

DoorDash runs an AI reviewer across 10,000 pull requests a week and engineers accept 60% of what it says. A single-pass model on their own benchmark caught 31% of real issues. The jump to production didn't come from a better model — it came from the harness wrapped around it.

aiagentscode-review
5 min

Why Five AI Agents Are Cleaner Than One

One agent reads a big codebase file by file and fills its own context window with everything it opens. A subagent does the same search in a separate context and gives back only the answer, which is why you can run five at once and keep your main session clean.

aiagentssubagents
4 min

GPT-5.6 Is Three Models, and Most of Us Can't Use It Yet

OpenAI previewed GPT-5.6 not as one model but as three — Sol, Terra, and Luna — priced up to 5x apart and mapped to different kinds of work. The more interesting parts are the pricing strategy underneath it and the fact that it went to the US government before it went to you.

aillmopenai
4 min

GLM 5.2 vs Claude: An Open Model Just Pulled Even on Coding

GLM 5.2 lands within a point of Claude Opus 4.8 on coding benchmarks, at roughly a fifth of the cost. The score isn't the whole story: open weights doesn't mean it runs on your laptop, and Claude still wins the hard 20%.

aillmopen-source
5 min

Quantization, Explained: Why Big Models Run on Small Hardware

A 7B model is 14 GB in full precision. A 70B is 140 GB. Quantization is the trick that brings those numbers down to something your machine can hold — and the tradeoff is smaller than most people expect.

aillmlocal-llm
4 min

DeerFlow 2.0 Isn't a Framework — It's a Harness

ByteDance open-sourced DeerFlow 2.0 and it hit #1 on GitHub Trending in hours. The number isn't the story. The shift from framework to batteries-included harness is.

aiagentsopen-source
4 min

Agent Skills Are Becoming the Vendor's Job

1,497 agent skills now sit in one repo, and the best ones come straight from Stripe, Cloudflare, Figma, and Anthropic. Here's why vendor-written skills change how you set up an agent—and five worth installing first.

aiagentsskills
7 min

The Real AI Shift: Agents Are Moving Inside the Tools

Palmier Pro hit 3,500 GitHub stars in 48 hours—not because it reinvented editing, but because it ships an MCP server. Here's why that matters.

aiagentsmcp
10 min

Distributed Tracing in Go with OpenTelemetry: A Production-Ready Implementation Guide

An in-depth exploration of implementing distributed tracing in Go microservices using OpenTelemetry, covering propagation context, sampling strategies, and performance optimization.

golangdistributed-systemsobservability
8 min

Beyond JSON: Achieving Sub-millisecond Latency with Go, NATS, and Protobuf

Why traditional HTTP/JSON architectures fail at scale, and how a binary-first, event-driven stack delivers real-time performance.

GoNATSMicroservices