Metehan Ariman

@metehanariman

Engineering-first AI, tested not promised. I break down the AI tools that actually ship — for people who build. 9+ yrs in production.

metehanariman.com

Joined June 2014

192 Following

106 Followers

Replies Reposts Media

Pinned
Metehan Ariman
@metehanariman
Jul 9
OpenAI's new voice model isn't the smart one. GPT-Live runs the conversation. The second a question needs real thinking, it hands off to GPT-5.5 sitting behind it and pipes the answer back mid-sentence. Fast model in front, frontier model in the back. A teardown.
2
Metehan Ariman
@metehanariman
Aug 14
Vercel says its AI SDK factory now authors 25–35% of weekly merged PRs and closes 70–80% of issues after four weeks. Humans still approve every merge. My take: scale evidence before authority. Make agent PRs easier to verify before making review easier to skip.
Metehan Ariman
@metehanariman
Aug 13
Anthropic says users approve 97% of Claude Code prompts. In its study, people caught 13.6% of dangerous commands; auto mode caught 89%. It becomes default Aug 14. After 3 consecutive blocks or 20/session, it returns to manual approval. Better default. Repo trust still matters.
Metehan Ariman
@metehanariman
Aug 12
Docker Sandboxes gives each coding agent a microVM with only the project mounted. It can install packages and start containers without touching the host. Network and filesystem controls exist. Org-wide policy requires Docker AI Governance. Better boundary than another prompt.
Metehan Ariman
@metehanariman
Aug 11
A coding agent's deploy handoff should contain four things: exact diff, test command plus output, migration state, and an executable rollback command. If one is missing, the release is not ready. Let the agent prepare it; let CI or a human own production.