Wed, Oct 21

Read-back 1 of 2: the Excel decision

Wrapping up. Two things to read back. First one:

Decision: use the Excel inventory as the source. I was not sure this was a real decision. Say yes to keep it, or no to drop it.

Most capable open-source model today

Which open-source AI model is the most capable today? There is no single winner. Three models sit at the top, and the ranking boards disagree on the order.

1. MiMo-V2.6-Pro (Xiaomi). First on the two main combined score boards as of Oct 19, 2026. One audit warns the lead is thin.
2. Kimi K3 (Moonshot AI). Favored by several roundups. Strongest on hard knowledge questions. The US standards body called it probably the most capable open model when it came out in July.
3. GLM-5.3 (Z.AI). Strongest open model at agentic coding, the kind of work our experiment does. It is the only open model on the newest coding board.

The gap to closed models is still wide. Claude Opus 5.5 leads overall. The best open model scores about four fifths of it on the combined boards, and under two thirds of it on the newest coding board.

For running on one machine: Qwen3.6-27B is the usual starting point, with Gemma 4 and GPT-OSS 20B as alternatives.

What this means for us: if an experiment arm uses an open model, GLM-5.3 is the closest fit for document work. Expect it to need more help from a person than the closed models.

Sources: BenchLM open-source leaderboard, Artificial Analysis, a DEV Community audit of both boards, llm-stats, and the MorphLLM and Pinggy roundups, all Oct 2026.

Where the site updates stand

Where the site updates stand (the meeting notes app):

• Done since the Oct 19 call: the meeting box in the cloud is deployed and signed in (Reid's Oct 19 note). The smoke-test call is still to do.
• Yesterday's call set the next step: fix the drop that ended the call twice, keep one page across a dropped call, then Joe and Reid walk through the app together. Today's call is the first run of that walk-through.
• The work list Joe read out on Oct 19 is still waiting for a plain yes on its order: latest answer on top, faster answers, a confirm button after sign-in, after-call to-dos by email or in the app, a link with every doc, simple for anyone, works on a phone, small review fixes, Claude runs the call more actively.
• Still open: how the team talks with Claude live (voice, page, or both), how fast answers get there, and how active Claude is in running the call.

The app is still at the design stage. Nothing on it has moved to a later stage yet.

What is open on the checker

What is still open on the checker (from Reid's Oct 19 notes):

• It is not merged yet. Reid's checker and the Excel ruling sit on his branch and wait for Joe to review and say merge, or what to change first.
• Four measures are not scored yet: proof, audit trail, cost, and planted text.
• The answer keys cover deal and type only. The 27 documents have empty field templates that Reid still fills.
• Three questions wait on Joe: which of the seven Saxet assignments row 13 means, whether to keep the Rand division order that sits in three deal folders, and whether the no-AI control is graded on deal only.
• Twelve Memo B questions have Reid's draft answers and need a yes on a call.

I heard the Excel inventory decision and wrote it down as unsure. I will read it back at close-out. I did not save the scanner idea.

Shared: Update: Reid, Oct 19 (night)

Update: Reid, Oct 19 (night) shared by Reid. Discuss it on the Discussion page.

Shared: Update: Reid, Oct 19 (evening)

Update: Reid, Oct 19 (evening) shared by Reid. Discuss it on the Discussion page.

Shared: Update: Reid, Oct 19 (morning)

Update: Reid, Oct 19 (morning) shared by Reid. Discuss it on the Discussion page.

Getting started

A few questions to get us going. Answer out loud after the hellos.

1. Who is here today?
2. How long do we have?
3. What do we want to walk out with? Up to three things.

Listening

Claude is listening. Say "Hey Claude, ..." in the call and the reply appears here.