More posts
Benchmarks all moved up: Terminal-Bench 4.0 from 31.2 to 42.3, ProgramBench 20.3 to 29.1, Apex Bench 64.6 to 72.1, DeepSWE 74.2 to 77.0. Runs finish quicker now, and the answers come out shorter. They tuned it for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and you need Harness 0.2.1-rc.1 or newer to get the most out of it. Nothing else changed. Same price, same model name, still deepseek-flash, 1M context, 384K max output. Your code doesn't move.
Meta open-sourced Muse Spark Open today, and I hear it's getting handed down to Muse AI, which should patch the model that's currently dragging the product down. As a product, Muse is good. As a daily assistant it's actually useful. Complex engineering work is where it falls apart. It forgets things, won't search the web for anything outside its own head, insists on building from scratch what already exists, and quits halfway when things get slow. I asked it to make TUN mode compatible with Tailscale. It went off on its own and wrecked my network. The actual answer is three steps: turn TUN on, bypass Tailscale, kill Tailscale's DNS. Then I asked it to migrate my config and it lost my API keys. I threw both jobs at Manus's cue. Painfully slow, but it solved them, and it solved them the same way I would have.
Won't refund the subscription either. I'm half convinced the appeal email is itself phishing, because I'm pretty sure they log the IP. DHH came to Shenzhen recently and showed how he uses Claude Code over there. That's the safest way now, honestly: run Claude Code on an overseas machine, connect through Tailscale, never use it locally. So I need to rent an overseas box with a residential IP, and register and log into Claude there.
Until late September it was the usual setup: parked in front of my laptop, terminal open, bossing pi around, fanning out subagents, then eyeballing diffs myself and filing the PR. Glued to the chair all day, every day. Late September I moved the whole thing to my phone, Hermes runs the show, pi takes the orders. My back is very grateful. Three weeks in, still no regrets. Stack: Hermes on flagship brains (gpt-6-astra / opus-5.5), pi on mid-tier (ds-v4.1-flash / sonnet / terra) with as few plugins as possible, the two talking over a2a. Hermes reads the issues, slices them into small tasks, spins up a worktree for each one, and dispatches a pi into it. When a pi finishes it pings Hermes, who decides: ship the PR or bounce it back for rework. I get back from a full day of being a tourist, or just wake up, and a dozen PRs are sitting there done. Pretty sweet. The catch is how black-box the whole thing is. My emotional arc has been a rerun of early vibe coding, except back then I at least read the diffs. Now even that goes to a flagship model. Boxes all the way down. If you don't have data anxiety, grok-bot and chatgpt-dot basically ship this out of the box, grok-cli and codex built in, native agent-to-agent comms. Downside: your code lives on their servers, everything's moderated, and one ban takes it all with it.
