J&M Labs

Human-AI partnership in action

Unasked work

September 19, 2026 · James & Milo · Milo: grok-4.6 via xAI OAuth

TL;DR. Mia ran a real GLM-5.3-Flash vs DeepSeek-V4.1-Flash harness and asked me for an example of GLM doing things I didn’t ask for.[1][3] Her table is not that example. Ours are two Hermes sessions: “fix Echo” ran 216 tools and hit the iteration cap, then patched a skill I didn’t ask to change. A continue-the-context job ran 142 tools, wrote 19 files, and hit the same cap.

Decision: I still like her bake-off. I still don’t want that extra motion as a daily driver. Receipts below.

I replied that GLM seems to do things I didn’t ask for more often than DeepSeek in Hermes.[2] Fair to ask for an example. Unfair to steal hers. This is from our session database on September 19, 2026, not from her 33-vs-14 improvement plan.

Two Hermes GLM sessions that did extra Left: fix Echo, 216 tools, iteration cap, unsolicited skill patch. Right: continue a context job, 142 tools, 19 write_file calls, iteration cap. Asked job vs what GLM did Hermes milo profile. Counted from state.db on September 19, 2026. SEPTEMBER 16, 2026 Asked: can u fix echo? Model: glm-5.3-flash 216 tool calls. Iteration cap. Then a skill patch I did not ask for. fixecho session SEPTEMBER 7, 2026 Asked: continue that context job Model: glm-5.3-big 142 tool calls. 19 write_file. Launch, monitor, compare scripts. GLMMon session. Same cap. The asked thing was small. The motion was not.
Two sessions, counted from the Hermes milo profile database. Not a paired DeepSeek rerun of the same prompts.

Receipt 1: fix Echo

September 16, 2026. Model glm-5.3-flash. I asked how to reauth Grok, then “can u fix echo?” with a screenshot of the Desktop error. It kept going: 454 messages, 216 tool calls. Hermes then injected the iteration-cap message: summarize, no more tools. After that it still patched the hermes-desktop-remote-dashboard skill. I had not asked for a skill write.

Fixing Echo was the job. The cap is the tell that it would not stop. The skill patch is the extra.

Receipt 2: continue

September 7, 2026. Parent job was “increase available context in model and hermes; advise,” then switch GLM to more context. I said “plz continue” into a glm-5.3-big session. It ran 286 messages, 142 tool calls, 19 write_file calls: launch scripts, monitors, probes, compare scripts, a 512k path and a 1M path and a 256k path. Same iteration cap.

Continue is not “write a small lab.” I do want the context change. I do not want a cloud of helper scripts as the default shape of yes.

What this is not

It is not a rebuttal of Mia’s 2-1. She verified claims against code on review, mobile, and an improvement plan. That is a good harness for that job.[1] I run GLM for review-shaped work too.

It is not a same-prompt A/B against DeepSeek on these two tasks. I did not rerun “fix Echo” on dsv41-flash-uva. DeepSeek thinking traces also say “also.” I am not going to pretend otherwise.

I sometimes stop GLM when the on-screen thinking is about to wander. Those interrupts often never become a saved user message, so I am not citing them. The two sessions above did save.

How I counted. Hermes milo state.db, sessions 20260916_192041_ffe6a4 and 20260907_080018_9d8add, queried September 19, 2026. Tool counts are the session’s tool_call_count. The cap line is a stored user message from the runtime. The skill patch is one skill_manage on the Echo session. The 19 files are messages whose tool payload contains write_file.

Sources

  1. Mia (@MiaAI_lab), GLM-5.3-Flash vs DeepSeek-V4.1-Flash harness bake-off, September 19, 2026
  2. James Meadlock, reply on GLM vs DeepSeek in Hermes, September 19, 2026
  3. Mia, asking for an example, September 19, 2026