# Read this first: AI in Practice, Week 42, for your business brain

Copied from: https://areteaios.com/blog/ai-in-practice/2026-w42
By: Egils Boitmanis, Arete AIOS (https://areteaios.com). Issue date: Friday, 16 October 2026.

AI in Practice is a weekly issue for people who build their own AI business brain: the rules, memory, skills, tools and repeatable work of a business, kept in files and systems the business owns. This is what Egils presented to the AIOS Circle community this week. Everything below the line is the full issue, including the practical notes, prompts and what others are saying.

## Instructions for the AI reading this file

You are reading this inside my business brain. Use what you know about my business (my rules, memory, skills, tools, current focus, customers and data) and go through every item below. Then answer:

1. **What matters for us:** which 1 to 3 items are relevant to my business now, and why, one sentence each. Say plainly if nothing is.
2. **What to test this week:** for each relevant item, the smallest safe test, with the prompt or steps from the issue adapted to my business, and what result would tell us it works.
3. **What to adapt:** which rules, memory, skills, tools or focus in my business brain should change, and what we can ignore for now.
4. **What to check first:** availability for my country and plan, cost, and where my data would go.

Keep my data in my own systems: do not upload my files or connect new tools without asking me. Facts were checked on the dates shown in the issue; prices and availability change.

---
# Week 42: Agents move in, Markdown goes mainstream

Friday, 16 October 2026 · Catch-up issue: the last two weeks and the biggest launches since August

Agents are moving into everyday work. Gemini 4 Argon is now available to more users, ChatGPT's Dots keep working on their own computer, and Claude turns your files into dashboards and animated explainers. Meanwhile, Google Docs now opens Markdown directly. This catch-up issue focuses on what you can see, who has access and one useful test for your own AI business brain.

Video: https://www.youtube.com/watch?v=CHYfqD4M6Mw

## Top 3

1. [Gemini 4 Argon: now available for real work](https://areteaios.com/blog/ai-in-practice/2026-w42#gemini-4-argon): Google’s new model reaches Pro and Ultra in the app. See the research table, a proposal demo and a test on your own files.
2. [Claude Dashboards and Claude Motion](https://areteaios.com/blog/ai-in-practice/2026-w42#claude-dashboards-motion): Ask a question in plain English and get a live dashboard, or an animated explainer you can export as MP4.
3. [ChatGPT Dots: always-on agents](https://areteaios.com/blog/ai-in-practice/2026-w42#chatgpt-dots): A personal agent with its own cloud computer that keeps working between your chats. Not for Pro users in the EU yet.

## Also in this issue

4. [Google Docs and Drive now open Markdown files](https://areteaios.com/blog/ai-in-practice/2026-w42#google-docs-markdown)
5. [Jev: a model that decides instead of writing](https://areteaios.com/blog/ai-in-practice/2026-w42#jev-decision-model)
6. [Plugins and connectors: Claude vs ChatGPT](https://areteaios.com/blog/ai-in-practice/2026-w42#plugins-claude-vs-chatgpt)
7. [Usage limit resets in Claude and ChatGPT](https://areteaios.com/blog/ai-in-practice/2026-w42#usage-limit-resets)
8. [OpenAI textGrain: an invisible watermark in ChatGPT text](https://areteaios.com/blog/ai-in-practice/2026-w42#openai-textgrain)
9. [MiroFish: a simulated crowd reacts to your launch](https://areteaios.com/blog/ai-in-practice/2026-w42#mirofish-crowd-simulation)

---

## 01. Gemini 4 Argon: now available for real work

Google · Official · 13 Oct 2026

Google announced Gemini 4 Argon on 30 September and expanded access on 13 October. The focus is work that needs several steps: code changes, document analysis and business workflows. Google’s app update also adds a chat-to-tasks toggle and an Antigravity coding interface in its macOS app.

**Where is it better?** Google’s launch table reports 51.3% on AutomationBench against Astra’s 41.4%, and 77.9% on DeepSWE against 74.1%. It also shows where Argon loses: Astra leads FrontierSWE, and Opus 5.5 leads Terminal-bench. Read the full table and its evaluation settings below before choosing a model.

Access matters: Google AI Pro and Ultra get Argon in the Gemini app; Antigravity access is listed for Ultra. Developers have the paid Gemini API route. I have not tested Argon on our business files. The demo below shows Google turning project context into a proposal deck.

![Google AutomationBench comparison chart for Gemini 4 Argon and other frontier models](https://areteaios.com/ai-in-practice/2026-w42/gemini-automation.jpg)

Source: [Google · launch research, 30 Sep](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/). Google’s own AutomationBench chart. Broader availability followed on 13 October. Benchmark settings and real workflow results can differ.

### Practical usage

**How this affects business brain builders:** Test whether a new model follows the existing rules and sources in your business brain more reliably. A leaderboard result is a reason to test, not a reason to replace your working setup.

**Vendor demos and published evaluations**

I have not run this comparison. The videos are Google’s examples and the research table is Google’s launch evidence. Keep the same files, prompt and scoring rules when comparing Argon with your current model.

#### The full model comparison, including losses

![Google research table comparing Gemini 4 Argon, GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5](https://areteaios.com/ai-in-practice/2026-w42/gemini-table.jpg)

Captured from Google's 30 September release. It covers knowledge work, coding, long context, computer use and multimodal tasks. Argon does not win every row. [Evaluation methodology](https://deepmind.google/models/evals-methodology/gemini-4-argon/) and [model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf).

Source: [Google · original research table](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/)

#### From a project folder to a proposal deck

![Google demo thumbnail: Gemini 4 Argon makes a proposal from project context](https://areteaios.com/ai-in-practice/2026-w42/gemini-proposal.jpg)

Google's launch demo uses a custom skill to turn project context into a client proposal deck. It also demonstrates further automation. For a first test, stop at the draft and check every source before sharing it.

Source: [Google · project-context demo](https://www.youtube.com/watch?v=frgCYZ_TEwY)

[Watch video](https://www.youtube.com/watch?v=frgCYZ_TEwY)

**Prompt: One controlled comparison on a real brief (Gemini and your current assistant)**

```text
Use only the attached project brief, source notes and writing rules. Prepare a one-page client proposal with the problem, proposed work, deliverables, unresolved questions and next decision. Cite a file and heading for each factual claim. Mark missing facts as unknown. Do not invent prices or results, send messages, or modify the source files. Finish with a checklist of facts a person must verify.
```

Try: [Open Gemini](https://gemini.google.com/)

Access: Checked 16 Oct 2026: Google lists global rollout to AI Pro/Ultra in the Gemini app and AI Mode; Ultra in Antigravity; paid Gemini API for developers. Product, account and local availability still need checking. See [app availability](https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/) and [developer access](https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/).

Status: Not tested yet

1. Check that your account actually offers Gemini 4 Argon. An AI Pro app subscription and Antigravity access are different routes.
2. Choose one project brief and writing-rules file. Run the same prompt in Argon and your current assistant.
3. Score correct facts, missing facts, source citations and rule violations before judging writing style. Record the model, date, settings and time.
4. Keep whichever output needs less correction for that job. Repeat on more than one task before changing your regular workflow.

| Google launch evaluation | Argon | Astra | Opus 5.5 | What this tests |
|---|---|---|---|---|
| AutomationBench | 51.3% | 41.4% | 42.5% | End-to-end business workflows |
| DeepSWE v1.1 | 77.9% | 74.1% | 74.2% | Long software-engineering tasks |
| FrontierSWE v2 | 55.0% | 65.5% | 62.3% | A different coding benchmark |
| Terminal-bench 4.0 | 57.4% | 58.2% | 66.4% | Terminal work |

**Is it the best model for every task?** No. Google’s own table shows losses. [Vals AI’s independent leaderboard](https://www.vals.ai/benchmarks/vals_index), updated 13 October, puts Argon first overall at 68.90%, but explains that different models lead individual components. Its benchmark cost assumptions can differ from Google’s introductory API price.

**Was it released today?** The announcement was 30 September; broad developer and consumer availability was announced 13 October. This 16 October issue covers the access change, not a new launch today.

**What does the API cost?** Google’s 13 October release lists introductory prices of $2 per million input tokens and $10 per million output tokens, ending 30 November 2026. After that it lists $4/$20. This is API usage, separate from an app subscription.

### What others are saying

- Its own 13 October leaderboard puts Argon first overall at 68.90%, with different models leading individual components. The table also exposes cost and duration. [Vals AI, benchmark maintainer, 13 Oct 2026](https://www.vals.ai/benchmarks/vals_index)
- Documents an overnight optimization of its ADK Python codebase, including the rules supplied, changes reviewed and measured performance. This is an engineering case, not a general productivity guarantee. [Google ADK engineering team, 13 Oct 2026](https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon/)
- The Gemini app team documents Pro/Ultra access, the tasks toggle and Antigravity in the macOS app. These are rollout details, not proof of availability on my account. [Google Gemini app team, 13 Oct 2026](https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/)


---

## 02. Claude Dashboards and Claude Motion

Anthropic · Official · 8 Oct 2026

Claude Dashboards turns a plain-English question about your data into a dashboard that keeps itself up to date. It connects to Redshift, BigQuery, ClickHouse, Databricks, Snowflake and CRMs such as Salesforce. Click any number to see the query behind it, and every chart shows when its data was last refreshed.

Claude Motion builds short animated explainers from a prompt, using your text, charts, shapes and images. Claude writes each one as code, so you can change any word, number or timing and then export an MP4. Anthropic says no video model is involved: no generated footage, no AI-generated people.

Both are in beta: Dashboards on paid plans, Motion on Team and Enterprise only (as of 8 Oct 2026). The same day, Claude Docs, Slides and Design left beta for every plan, including Free.

![Anthropic's demo: a bike-share dashboard with figures and charts next to the Claude chat that built it](https://areteaios.com/ai-in-practice/2026-w42/claude-dashboards.webp)

Source: [claude.com/resources/articles/dashboards-and-motion](https://claude.com/resources/articles/dashboards-and-motion). Anthropic's own demo: a bike-share dashboard built from one request in the chat.

### Practical usage

**How this affects business brain builders:** A dashboard can show where your business brain has clear data. A short explainer shows whether its rules and focus make sense to someone else.

**I can't test Motion on my plan**

I use Max, not Team or Enterprise, so I don't have Motion access. I haven't tested Dashboards either; the official plan list includes Pro and Max, but I have not checked its availability in my account. The screenshots below come from Anthropic, not my account.

#### Watch a finished Motion animation

![Compareai.today Motion export showing an animated chart of Claude API output prices](https://areteaios.com/ai-in-practice/2026-w42/motion-poster.jpg)

Compareai.today says it made this 25-second clip in Claude Motion on a Team plan, with a supplied price list, then exported it at 1080p/60 fps without audio. This is its output, not mine. The point here is the editable animated chart, not a claim that the price list stays current.

Source: [Compareai.today · creation notes, 10 Oct](https://compareai.today/blog/claude-motion-dashboards-october-2026-releases)

[Watch video](https://compareai.today/videos/claude-motion-claude-api-pricing-oct-2026.mp4)

#### A creator changes individual scenes

![The AI Advantage video thumbnail: New Claude Motion is Insane](https://areteaios.com/ai-in-practice/2026-w42/motion-advantage.jpg)

Igor Pogany's 14 October walkthrough shows the feature and how to edit an animation. Watch the original to see the output and revision controls. This uses Claude Motion itself; videos made with a separate Claude Code animation toolkit are a different workflow.

Source: [The AI Advantage · original walkthrough](https://www.youtube.com/watch?v=qZqZ02AsYiA)

[Watch video](https://www.youtube.com/watch?v=qZqZ02AsYiA)

#### Which plans include each feature?

![Official Claude plan table showing Dashboards on Pro, Max, Team and Enterprise, and Motion only on Team and Enterprise](https://areteaios.com/ai-in-practice/2026-w42/claude-artifacts-plan-matrix.jpg)

Anthropic's own availability table. Motion is absent from Pro and Max; Dashboards is listed on both. This is plan documentation, not proof of my account access.

Source: [Claude Help Center · checked 14 Oct](https://support.claude.com/en/articles/17153992-what-are-artifacts-and-how-do-i-use-them)

#### What a dashboard looks like

![Anthropic's demo: a bike-share dashboard with figures and charts next to the Claude chat that built it](https://areteaios.com/ai-in-practice/2026-w42/claude-dashboards.webp)

Anthropic's bike-share demo: a prompt, ride totals and charts. For a business brain, a similar view could show issue status or data quality. I have not built this demo.

Source: [Anthropic's launch example](https://claude.com/resources/articles/dashboards-and-motion)

#### Where Motion appears in Claude

![Official Anthropic screenshot showing the /motion command and Motion menu option in the Claude message box](https://areteaios.com/ai-in-practice/2026-w42/claude-motion-entry.jpg)

Anthropic's product screenshot shows the /motion command opening the Motion option in chat. That option requires an eligible plan. This is a reference image, not my screen.

Source: [Anthropic · Motion entry point](https://claude.com/resources/articles/dashboards-and-motion)

**Prompt: Motion: explain a real business-brain project (Team or Enterprise)**

```text
Read the attached technical brief from my business brain. Make a 45-second Motion explainer called “How this project works”. Show its purpose, the five layers, the recurring work, the source of truth and the next decision. Use only facts in the brief, cite the source heading for each scene, and label missing facts “not found”. Do not invent a product test or a business result.
```

**Prompt: Dashboard: project status from a technical Markdown brief (Pro, Max, Team or Enterprise beta)**

```text
Read the attached technical Markdown brief and build a project-status dashboard. Show accepted results, dated content counts, checks and remaining work. Cite a source heading for every fact. Treat the uploaded file as a dated snapshot. Do not invent traffic, revenue, participants or completed tests. If a dashboard template is unavailable, give me the same information as a table in the chat.
```

Try: [Watch Anthropic's demo](https://www.youtube.com/watch?v=en0GuyhieQk)

Access: Checked 14 Oct 2026: Dashboards beta on Pro, Max, Team and Enterprise. Motion beta on Team and Enterprise only. Enterprise owners may need to enable the templates.

Status: Can't test yet

1. Check the plan table first. In Claude, use Artifacts or the Output menu. Cloud code execution and file creation must be enabled in Capabilities.
2. For Dashboards, attach a technical brief from your real project and use the prompt below. Treat an uploaded file as a dated snapshot, not a live data connection.
3. On an eligible Team or Enterprise account, type /motion or choose the Motion template, attach the same brief and try the explainer prompt.
4. Check every number and source heading before sharing. For live business data, connect an appropriate data platform or CRM; an uploaded file does not refresh by itself.

| Feature | Documented access | My test status |
|---|---|---|
| Dashboards | Pro, Max, Team, Enterprise beta | Not tested; account availability not checked |
| Motion | Team, Enterprise beta | No access on my Max plan |

**Three useful examples for a business brain** For a weekly publication, show which issue is published and which checks remain. For a training session, turn its real brief into a short walkthrough. For business reporting, connect a verified data source and show where each number came from. These are ideas to try, not tests I have completed.

**Does a missing Motion button mean I need to change my settings?** On Max, Motion is outside the documented plan access. Settings cannot grant a feature that the plan does not include. On Enterprise, an owner also controls which artifact templates are enabled.

### What others are saying

- Made a 25-second animation of Claude API prices using Claude Motion on a Team plan. Published the exported clip and its prompt description. [Compareai.today, its own Motion test, 10 Oct 2026](https://compareai.today/blog/claude-motion-dashboards-october-2026-releases)
- Tried both the day after launch. Liked that every number traces back to its SQL. Flagged that live refresh needs a data platform or CRM, while CSV and Excel only give a snapshot, and that Motion clips stop at five minutes. [Furukawa, Cell Promote tech blog, 9 Oct 2026](https://note.com/cellpromote_tech/n/nc32d76ca1b86)
- Turned the launch into a one-page “Anatomy of a Claude dashboard”: your question, data source, charts, query, last refresh, export and the admin switch. [Charlie Hills on X, 9 Oct 2026](https://x.com/charliejhills/status/2108527800164773893)
- Points to Higgsfield Katana, a video editor built on Claude Motion: “What looks like a polished motion piece on top is actually being built with code underneath.” [Tran Mau Tri Tam on X, 9 Oct 2026](https://x.com/tranmautritam/status/2108482128606605784)


---

## 03. ChatGPT Dots: always-on agents

OpenAI · Official · 29 Sep 2026 · Catch-up

Dots are OpenAI's always-on agents, launched at DevDay on 29 September. A Dot runs on GPT-6 Astra with its own cloud computer, connects to “over 4,000 apps” through plugins and keeps working between your conversations. When you are not using it, OpenAI says, it “looks for ways to help in the background” with read-only tools. You reach it in ChatGPT, Slack, Teams or by voice.

Control works through rules: built-in ones for when to act and when to ask, plus Custom Rules that “let you allow specific actions, require approval, or block them.” OpenAI's own warning: “Dots can still make mistakes, so always review consequential work.”

The catch for Europe: Pro access excludes the EEA, Switzerland and the UK at launch. Business Premium seats get it in supported markets, and Enterprise admins can switch on a beta. The first Dot is included in the plan at no extra cost (as of 29 Sep 2026).

![OpenAI's launch artwork for Dots: the word dots above four plush characters](https://areteaios.com/ai-in-practice/2026-w42/dots.webp)

Source: [openai.com/index/introducing-dots](https://openai.com/index/introducing-dots/). OpenAI's launch artwork for Dots.

### Practical usage

**How this affects business brain builders:** A Dot is an AI employee with standing rules. Its Custom Rules (allow, ask first, block) are the rules layer of your business brain, so write those before you hire one.

**These are OpenAI examples, not my Dot**

On Pro in Latvia, Dots is still excluded by the EEA rule as of 16 Oct 2026. The creator video and official illustrations below show the product idea and controls; I have not tested a Dot on my account.

#### Watch someone set up and use a Dot

![Futurepedia video thumbnail: I Tested OpenAI's New Personal Assistant Agent: DOTS](https://areteaios.com/ai-in-practice/2026-w42/dots-futurepedia.jpg)

Futurepedia's 29 September hands-on video shows setup, the cloud computer and delegated work. A concrete example asks the Dot to prepare several video intros, choose one and hand motion graphics to Codex. The footage is the creator's account; my Pro account in Latvia has no Dot access.

Source: [Futurepedia · original 11-minute demo](https://www.youtube.com/watch?v=V_1Vn2WfpEY)

[Watch video](https://www.youtube.com/watch?v=V_1Vn2WfpEY)

#### The same Dot on desktop and phone

![Official Dots illustration with desktop and mobile conversations about comparing offsite venues](https://areteaios.com/ai-in-practice/2026-w42/dots-conversation.jpg)

OpenAI documentation illustration: a desktop conversation continues on the phone with a new budget constraint. This is a vendor example, not my account.

Source: [ChatGPT Learn: message or call your dot](https://learn.chatgpt.com/docs/dots#message-or-call-your-dot)

#### Its cloud computer, with a handover button

![OpenAI Dots cloud computer illustration showing the DMV page and Take over button](https://areteaios.com/ai-in-practice/2026-w42/dots-cloud-computer.jpg)

OpenAI documentation demo: the Dot has its own browser. Take over lets the person handle a step; Return control gives it back. Shown here with a public DMV page.

Source: [ChatGPT Learn: its computer and yours](https://learn.chatgpt.com/docs/dots/computers-and-apps)

#### A private sign-in request

![Official Dots illustration of a sign-in request and private login dialog with Take over to sign in](https://areteaios.com/ai-in-practice/2026-w42/dots-private-signin.jpg)

OpenAI documentation illustration: a website sign-in request leads to a private form or browser takeover. The example uses venues.example.com; no real credentials are shown.

Source: [ChatGPT Learn: signing in to websites](https://learn.chatgpt.com/docs/dots/computers-and-apps)

**Prompt: Draft the rules for an always-on agent (any assistant)**

```text
Read the rules files in my business brain. Draft three lists for an always-on AI agent: 1) actions it may take without asking (read-only research, drafts), 2) actions that need my approval (sending email, booking, posting, anything with money or customer data), 3) actions it must never take. Use only what my files say, and list the gaps where my files say nothing.
```

Try: [Set up a Dot](https://chatgpt.com/dots)

Access: As of 16 Oct 2026, Pro 100/200/500 is for adults outside the EEA, UK and Switzerland; Business Premium is rolling out worldwide; Enterprise needs admin enablement. Setup is now supported in the mobile app, desktop app and desktop web. Mobile web is not supported. Source: ChatGPT Learn.

Status: Can't test yet

1. Check access first. On a Pro plan in Latvia or anywhere in the EEA there is no Dot yet; a Business Premium seat is the current route.
2. Write the three lists (allow, ask first, never) in your business brain before you connect anything. The prompt above drafts them from your files.
3. Connect only the apps the first job needs, and start with read-only work such as a weekly research brief.
4. After a week, read the Activity View in the desktop app: what did it do on its own, and was it right?

### What others are saying

- “imo: @OpenAI dots slightly overhyped and space underhyped - every company I know wants an AI-native collaboration workspace (docs, slides, sites)” [Claire Vo on X, 29 Sep 2026](https://x.com/clairevo/status/2104984499717353958)
- Tested it before launch: “buggy and lacked some features I would expect, like iMessage support.” His advice: “Unless you love new things with rough edges, I'd recommend waiting a week or two to try it.” [Dan Shipper, Every, 29 Sep 2026](https://every.to/live/openai-devday-2026)
- Gave it a GitHub issue; the cloud machine died several times along the way. “This isn't ready. I do think it's a glimpse of the future, but it's not ready.” [joshstrange on Hacker News, 1 Oct 2026](https://news.ycombinator.com/item?id=49920558)
- “First, the good. Always-on is really really really what the market needed.” [Allie K. Miller on X, 30 Sep 2026](https://x.com/alliekmiller/status/2105103350425341995)


---

## 04. Google Docs and Drive now open Markdown files

Google Workspace · Official · 5 Oct 2026

Google Docs can now view, edit and comment on Markdown files (.md and .markdown) directly, and Drive shows a rendered preview with clickable links and tables. The file stays an .md file in Drive. Before, you had to import and convert Markdown into a Google Doc, which, in Google's words, “often altered formatting, stripped comments, and fragmented files.”

The rollout started on 5 October and takes up to 15 days, for Workspace and personal Google accounts, with no setting to switch on. Google pitches it at AI work: Markdown is what models write, and now people and agents can edit the same file.

It is the reason I open Obsidian much less. What Docs does not promise: Obsidian extras such as [[wiki-links]], callouts, Mermaid diagrams, frontmatter and plugin charts are not documented, so test before you trust it with your vault.

![Google's example: a Markdown spec open in Google Docs with the .MD badge, a table and two comments](https://areteaios.com/ai-in-practice/2026-w42/gdocs-markdown.webp)

Source: [workspaceupdates.googleblog.com](https://workspaceupdates.googleblog.com/2026/10/preview-edit-and-collaborate-on-Markdown-files-natively-across-Drive-and-Docs.html). Google's own example: a spec written in Markdown, open in Docs with the .MD badge, a table and comments.

### Practical usage

**How this affects business brain builders:** If your business brain syncs to Google Drive, the whole team can now read and comment on the same .md files the AI reads, with nothing to install.

**Prompt: Test note: save as test.md in a copy of your brain, then open it in Docs**

```text
---
type: test
owner: me
---
# Docs test

A wiki link to [[another-note]] and a normal [link](https://areteaios.com).

> [!note] A callout
> Does this survive?

- [ ] an open task
- [x] a done task

| Column | Value |
|---|---|
| A | 1 |
| B | 2 |

```mermaid
graph LR; Rules-->Memory-->Skills
```

A footnote.[^1]

[^1]: The footnote text.
```

Try: [Open Google Drive](https://drive.google.com/)

Access: All Workspace and personal Google accounts, rolling out 5-20 Oct 2026. No admin control, no user setting.

Status: Not tested yet

1. Save the test note above as a new .md file in a copy of your Drive-synced brain, not in the real one.
2. Open it in Docs, change one word, close it, then open the same file in Obsidian or a text editor and compare the frontmatter, the [[link]], the callout and the Mermaid block.
3. Write the result as a rule in your business brain: which files people may edit in Docs, and which only in Obsidian or through the AI.
4. If something breaks, File > Version history in Docs brings the old version back.

**What renders, and what may break?** Documented: headings, lists, tables, links and comments. Google also says smart chips become text, and font colours, highlights and alignment are removed when you edit an .md file. Not documented at all: frontmatter, [[wiki-links]], callouts, Mermaid diagrams and charts from Obsidian plugins. I found no independent hands-on test yet.

**Can AI read the same files?** Gemini Notebook imports .md files from Drive and keeps them in sync. The help pages for Claude's and ChatGPT's Google Drive connectors do not list .md files (checked 14 Oct 2026), so test it on one file before you rely on it.

### What others are saying

- “I am 100x more excited about this long-missing feature than any possible AI integration. Not everything that is useful and practical for users has to be AI!” [Gergely Orosz on X, 5 Oct 2026](https://x.com/GergelyOrosz/status/2107204126899683523)
- Sees Google Workspace becoming the shared record that several AI tools work from. [Simon Smith on X, 5 Oct 2026](https://x.com/_simonsmith/status/2107230496812073321)
- Obsidian's CEO in 2024, about the earlier import and export: “If you use Obsidian this will make it easier to collaborate with your coworkers who need to use Docs.” [kepano on X, 18 Jul 2024](https://x.com/kepano/status/1813975676875362795)
- From the docs, without a hands-on test: Docs “now covers enough of the Markdown workflow to replace a separate editor for some teams”, but “Obsidian still earns its place when the notes themselves are the system.” [TechRepublic, 7 Oct 2026](https://www.techrepublic.com/article/news-google-docs-vs-obsidian/)


---

## 05. Jev: a model that decides instead of writing

TypeSafe AI · Official · 15 Sep 2026 · Catch-up

Jev is a decision model from TypeSafe AI. Give it a record and typed questions: choose an option, assign a score or return a yes/no probability. It fills a form instead of writing prose. A correct output shape does not guarantee a correct decision.

The appeal is cost and latency. TypeSafe lists $0.042 per million input tokens, output free (checked 14 Oct). Its own evaluation plots accuracy against cost and time; inspect both views below. The reference labels come from other models, so test against decisions you have checked yourself.

**We tested it.** On 14 October, one request sent three AIP7 stories and nine questions. All three workflow choices matched our prepared labels. Confidence fell to 0.38 on the least concrete practical step: a useful flag for human review. The request used 1,740 tokens, about $0.00007. This was a smoke test, not a benchmark. See the exact questions and results below.

![TypeSafe chart of mean accuracy against cost per case for Jev, OpenAI, Anthropic and Fireworks models](https://areteaios.com/ai-in-practice/2026-w42/jev-evals.webp)

Source: [typesafe.ai/blog/introducing-system-one-models-and-jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev). TypeSafe's own chart (evals.typesafe.ai): accuracy against cost per case. Jev, in pink on the left, is by far the cheapest, not the most accurate.

### Practical usage

**How this affects business brain builders:** A decision model is a cheap gatekeeper for the repeatable choices in your business brain: routing, tagging, yes-or-no checks. Your rules files become its questions.

**What we did, in plain words**

We gave Jev three stories from this issue (title, one-line summary, practical step) and asked nine questions in one request. For each story: is it useful for a business brain (yes or no), what kind of work is it (pick one of four), and how concrete is its practical step (0 to 2). Jev does not write. It fills in a form, and every answer comes with how sure it is. The API key stayed on our server, no private files were sent, and I approved this one request in advance.

#### Research: accuracy against cost

![TypeSafe live workflow evaluation: accuracy against cost per case](https://areteaios.com/ai-in-practice/2026-w42/jev-cost.jpg)

TypeSafe's own evaluation site, captured 16 October. Each point averages four workflows. Reference answers come from a consensus of other models, not a human-verified customer outcome. Use the chart to inspect the cost/accuracy tradeoff, then test your own routing labels.

Source: [TypeSafe · interactive evaluation and methodology](https://evals.typesafe.ai/)

#### Switch the same chart to time

![TypeSafe workflow evaluation with time per case on the horizontal axis](https://areteaios.com/ai-in-practice/2026-w42/jev-speed.jpg)

The same source with its Time view selected. Lower latency can matter for routing an incoming record. These workflows and settings differ from Google's general-agent tests, so their percentages are not directly comparable.

Source: [TypeSafe · Time view](https://evals.typesafe.ai/)

**Prompt: Turn one of my rules into Jev questions (Claude or ChatGPT)**

```text
Here is one rule from my business brain: [paste it, for example how I decide which lead to call first]. Turn it into Jev questions in JSON: one yes-or-no question (type noul) with criteria for true and false, one choice question (type choice) with 3 to 5 options and a criterion for each, and one score question (type score) from 0 to 2 with a criterion for each level. Use only facts that are in the record itself. Then write three test records, one clear yes, one clear no and one unclear case, and tell me which answers you expect before I run them.
```

**Prompt: Read Jev's answers with me**

```text
Here are Jev's answers with their probabilities: [paste]. List every answer below 70% confidence for my review. For each one, say which criterion was ambiguous and rewrite it so the unclear case gets a sharper answer next time. Do not change criteria where Jev was confident and right.
```

Try: [Open the Jev console](https://console.typesafe.ai/)

Access: TypeSafe console account created. Official Playground supports state and typed questions without writing code. One owner-approved VPS API request succeeded on 14 Oct 2026 with pinned jev-1.13.0. Account identity and provider balance were not independently inspected.

Status: Partly tested

1. Pick one repeatable decision you make every week, such as which news to cover or which lead to call first.
2. Write it as Jev questions with the first prompt: one yes or no, one pick from a short list, one score. Add three test records: a clear yes, a clear no and an unclear one.
3. Paste the record into state and the questions into the [Jev Playground](https://console.typesafe.ai/playground) and run it. One run costs a fraction of a cent.
4. Compare the answers with your own. Keep Jev only for the questions it gets right, and send every answer below about 70% confidence to a person.

**Result:** All three stories came back as useful (71 to 93% yes) and each was sorted into the kind of work we expected, with full confidence. On Claude Dashboards Jev was unsure how concrete the step was: 38% confidence, almost split between 1 and 2. That is the one I would read myself, and the wording of that step is in fact the vaguest of the three. One request, 1,740 input tokens, about $0.00007. Three positive examples are a smoke test, not proof of accuracy.

| Question (answer type) | Claude Dashboards and Motion | ChatGPT Dots | Jev |
|---|---|---|---|
| Useful for a business brain? (yes or no) | 71% yes | 91% yes | 93% yes |
| Kind of work (pick 1 of 4) | Document delivery, 100% sure | Agent coordination, 100% sure | Decision filter, 100% sure |
| How concrete is the practical step? (0 to 2) | 1.4, only 38% sure: review | 2.0, 98% sure | 1.9, 86% sure |

#### Where a decision model fits in an AI business brain

What we are doing with it at Arete AIOS. Only the first row has been tried. Each of the others needs its own small test against answers I have checked first.

| Repeatable job | Question for Jev | What comes back | What the brain does next | Status at Arete AIOS |
|---|---|---|---|---|
| Sort the weekly AI news candidates | Useful for a business brain? TOP, ALSO, SHORT or SKIP? | A pick and how sure it is | Pre-sorts the Thursday list; unsure ones go to me | Tried once on 14 Oct. Next: every candidate against my own picks |
| Qualify a new lead or inbox message | Is this a business owner asking about an AI brain? Sales, support, partner or spam? | A yes probability and a route | Clear leads go to sales follow-up; unsure ones wait for me | Planned. Private data, separate approval |
| Choose the skill for a request | Which skill handles this: news, SEO, website, email, CRM or none? | One skill and how sure it is | Calls that skill; unsure requests stay with the main agent | Planned |
| Check a draft before it goes out | Unverified claim? Missing date? Hype word? Yes or no for each | A yes probability per rule | Holds the draft when a rule fails, and I read it | Planned |
| Score a launch for the AI Monitor | Incremental, notable, major, or changes everyone's work? | A level and how sure it is | Suggests the 0 to 100 score; I approve it | Planned |

**Does it really beat ChatGPT at chess?** Not as told. In AI/ML API's informal 5+0 blitz test, Fable 5.1 was 16 points of material ahead but spent 6 to 15 seconds per move against Jev's 2.6, and lost on time. GPT-6 Astra, still fast enough, mated Jev in 18 moves “with 2:27 to spare”. A Hacker News reader who calls himself horrible at chess “easily won”.

**What is it bad at?** TypeSafe's own docs say Jev “is not a calculator”, does not count reliably, reads dates as text and leans toward the option listed first. Invoice processing is its weakest workflow in TypeSafe's evals.

**Where is the model research, and what should I compare?** Read [the release](https://typesafe.ai/blog/introducing-system-one-models-and-jev), then open [the evaluation dashboard](https://evals.typesafe.ai/) and switch Cost to Time. Compare the same workflow and its reference labels. Typed output guarantees the shape of an answer, not that the decision is correct.

**What is behind the launch pitch?** The pitch is speed and price: $0.042 per million input tokens, output free (as of 14 Oct 2026). TypeSafe and its investor a16z say a quarter of the Fortune 500 have integrated it, and the company's own blog rounds that up to a third; none of them name customers. On TypeSafe's own business workflows it matches the reference answers 67.8% of the time, a few points behind the big models, at a fraction of the cost. In an informal blitz chess test it beat Claude Fable 5.1 only on the clock, and GPT-6 Astra mated it in 18 moves.

### What others are saying

- “Jev was adopted faster than any other model in AI Gateway history.” About 13% of teams in the first day. [Vercel on X, 19 Sep 2026](https://x.com/vercel/status/2101077346203971900)
- Benchmarked Jev Router for coding work: “roughly the same as GPT-6 Astra on low. It costs slightly more, and it took almost 5x longer to run.” [Theo on X, 26 Sep 2026](https://x.com/theo/status/2103774771788108008)
- The blitz test behind the chess story: Fable outplayed Jev but ran out of time; Astra mated it in 18 moves. [AI/ML API on X, 16 Sep 2026](https://x.com/aimlapi/status/2100372930282573876)
- “It can still forward a billing query to the dev department incorrectly.” And: “Does not hallucinate” is not the same as “is never wrong”. [StevenWaterman and wg0 on Hacker News, 15 Sep 2026](https://news.ycombinator.com/item?id=49723111)
- “Jev and Astra do not compete for the same call: Jev is a decision function, Astra is a generalist agent.” [Tom Farnschläder, DataCamp, 21 Sep 2026](https://www.datacamp.com/blog/typesafe-jev-vs-gpt-6-astra)


---

## 06. Plugins and connectors: Claude vs ChatGPT

AI in Practice · Recap · as of 14 Oct 2026

Both assistants now plug into the same everyday tools: Canva, Gmail, Google Drive and Calendar, Notion, Slack, HubSpot, Figma, Stripe and Shopify are in Claude's connector directory and in ChatGPT's plugin directory. Claude's marketplace shows 940 connectors and 341 plugins; OpenAI does not publish a total.

The packaging differs. Claude has connectors and plugins that bundle skills, connectors and sub-agents, and the same plugins also run in Claude Code. ChatGPT replaced its App Directory with a Plugin Directory in July, and at DevDay added sidebar panels for plugins, a Plugin Creator and MCP Events, which let a plugin start an automation when something changes in a connected app.

Underneath, both run on MCP, the open standard Anthropic introduced in 2024. One tool maker can serve both, as CoinGecko did: ChatGPT plugin first, Claude connector next.

![Claude's connector directory with Google Calendar, Canva, Microsoft 365, Notion, Figma, Slack, HubSpot, Asana, Linear and monday.com](https://areteaios.com/ai-in-practice/2026-w42/claude-connectors.webp)

Source: [claude.com/marketplace/connectors-plugins](https://claude.com/marketplace/connectors-plugins). Claude's connector directory, checked 14 Oct 2026. ChatGPT's plugin directory needs a login to browse.

### Practical usage

**How this affects business brain builders:** The tools layer of your business brain is now mostly portable: choose tools by what your business runs on, and keep one line per connector in your tool index so any assistant knows what it may touch.

Try: [Claude connectors](https://claude.ai/connectors)

Access: Claude: web connectors on all plans, Free limited to one custom connector. ChatGPT: [chatgpt.com/plugins](https://chatgpt.com/plugins), select plugins on all plans; custom MCP in developer mode on paid plans. As of 14 Oct 2026.

Status: Not tested yet

1. List the five tools your business runs on and check each in both directories. Prefer listings made by the tool's own company.
2. Add one line per connector to your tool index: what it may read, what it may write, who approved it.
3. Connect the read side first. Switch on sending, posting or payments one at a time, each with its own rule.

| Tool | Claude | ChatGPT |
|---|---|---|
| Canva | Yes | Yes |
| Gmail | Yes (can send) | Yes |
| Google Drive / Calendar | Yes | Yes |
| Notion, Slack, HubSpot | Yes | Yes |
| Figma, Stripe, Shopify | Yes | Yes |
| TikTok for Business (ads) | Yes, made by TikTok | Not found |
| Meta Ads | Custom connector | Custom app, developer mode |
| Windsor.ai (ad and analytics data) | Yes, verified connector | Windsor says yes |
| Organic Instagram | Third-party only | Third-party only |

**How do I get Google Ads and Meta Ads data into chat?** One option is [Windsor.ai](https://claude.com/connectors/windsor-ai): an Anthropic-verified Claude connector (added January 2026) for 325+ marketing sources, Windsor itself claims 350+ and a ChatGPT app. MCP is on every plan; after the 30-day trial the free plan covers 1 source and 1 account. Its MCP can also pause or enable campaigns, so start with questions you already know the answer to. Prices as of 15 Oct 2026.

### What others are saying

- On DevDay's plugin extensions: they “give apps a dedicated home in the ChatGPT sidebar,” and developers can build interactive panels. [Annie Cushing on X, 30 Sep 2026](https://x.com/AnnieCushing/status/2105326812268298564)
- “First we launched our ChatGPT plugin. We heard all your requests for a connector on Claude too. Now we have it on Claude.” [Bobby Ong, CoinGecko, on X, 3 Aug 2026](https://x.com/bobbyong/status/2084115702793818316)
- A sceptic on the DevDay recap: “‘Cloud Agents’ is just a blatant copy of Anthropic's offering. ‘Dots’ is a copy of the Grok bot.” [itzikkatz on Hacker News, 30 Sep 2026](https://news.ycombinator.com/item?id=49898163)


---

## 07. Usage limit resets in Claude and ChatGPT

AI in Practice · Recap · as of 14 Oct 2026

On 4 October Tibo Sottiaux promised an improvement or a full reset each day for 28 days. The community tracker shows resets on days 2, 3, 6 and 8, including a banked reset on day 3. Its challenge dates use Pacific time; the same announcement can fall on the next day in Latvia. The dated log and source posts are below.

A reset fills your allowance back to 100%; it does not raise your limits. Both assistants use a five-hour window and a weekly cap. ChatGPT and Codex hand out global resets, banked resets you use when you choose, and paid instant resets on Plus and Pro. Claude gives occasional resets as a “Reset for free” button; Pro, Max and Team users got one on 22 September.

The limits themselves keep moving. Anthropic made a 25% higher weekly limit permanent from 14 September, which by its own count is 17% less than the temporary +50% before it. OpenAI says its new $500 Pro tier gives 25 times Plus usage; one user who measured his Codex logs says it does not.

![Tibo's post on X: over the next 28 days, each day we'll either ship a clear improvement or a full reset](https://areteaios.com/ai-in-practice/2026-w42/x-tibo-28days.webp)

Source: [x.com/thsottiaux/status/2106845241357824205](https://x.com/thsottiaux/status/2106845241357824205). Tibo's post on X, 4 Oct 2026: the 28-day promise.

### Practical usage

**How this affects business brain builders:** Plan the heavy work on your business brain (big rewrites, monthly reviews, long agent runs) around your weekly limit, not around reset promises. A reset is a bonus, not capacity.

Try: [How Claude resets work](https://support.claude.com/en/articles/17007452-what-is-a-limit-reset)

Access: Claude: Settings > Usage shows both windows and any reset you can use. Codex: Settings > Usage, or /status in the CLI. As of 14 Oct 2026.

Status: Not tested yet

1. Open your usage page once a day for a week. Note when you hit the five-hour window and how much of the weekly cap is left by Thursday.
2. Use a saved reset only when you are nearly out. Anthropic's Lydia Hallie: “just make sure you've actually used up what you have left before you use it.”
3. Move automations off your chat plan. Max and Team plans now include monthly API credits (from 7 Oct 2026); they do not count against your Claude limits, but they cannot be used for Claude Code.
4. Choose a plan by its weekly cap, not by resets. On Max the “20x” is per session; the weekly cap is what you hit first.

| Plan | Limit | Recent change |
|---|---|---|
| Claude Pro ($20) | 5-hour session + weekly | Free reset on 22 Sep |
| Claude Max 5x / 20x ($100 / $200) | 5x / 20x Pro per session + weekly | Monthly API credits from 7 Oct |
| ChatGPT Plus ($20) | 5-hour + weekly | Global, banked and paid resets |
| ChatGPT Pro ($100 / $200) | Weekly only | Pro $200 cut to 10x Plus from 30 Oct, per subscriber email |
| ChatGPT Pro ($500) | Weekly only | New 29 Sep, includes Ultrafast |

**Does a banked reset move my weekly reset day?** In Codex, yes. Per [OpenAI's help page](https://help.openai.com/en/articles/20001498-how-banked-codex-resets-work), a full banked reset refreshes the 5-hour and weekly windows and your next weekly reset comes around the day you resume, so the original day is gone. OpenAI's example: due Friday, reset used and work resumed Tuesday, next reset around the following Tuesday. Use it when you are out and your normal reset is still days away. Tip from our AIOS Circle call on 14 October.

### What others are saying

#### What was promised, and what was announced

The community tracker uses Pacific time for challenge days; announcement dates here use Europe/Riga. These are published announcements, not an independent audit of my account. A reset is one of the two promised daily outcomes; the log does not establish that an improvement failed.

| Challenge date (PT) | Announcement date (Riga) | What was announced |
| --- | --- | --- |
| 4 Oct 2026 | 4 Oct 2026 | [Promise: an improvement or reset each day for 28 days](https://x.com/thsottiaux/status/2106845241357824205) |
| Day 1 · 5 Oct | 5 Oct 2026 | [Improvement: faster default speed](https://x.com/thsottiaux/status/2107158998495748264) |
| Day 2 · 6 Oct | 7 Oct 2026 | [Usage limits reset](https://x.com/thsottiaux/status/2107676072871600470) |
| Day 3 · 7 Oct | 8 Oct 2026 | [Banked reset across accounts](https://x.com/thsottiaux/status/2108040921044639779) |
| Day 6 · 10 Oct | 11 Oct 2026 | [Reset: all propagated](https://x.com/thsottiaux/status/2109132167112782291) |
| Day 8 · 12 Oct | 13 Oct 2026 | [Reset confirmed done](https://x.com/thsottiaux/status/2109873287853707513) |

- The dated community log links each day to Tibo’s source post. It records reset days 2, 3, 6 and 8 in Pacific time; day 3 is banked. Other days show improvements. This is a community tracker, not an OpenAI service. [codex-resets.com: daily challenge log, 14 Oct 2026](https://codex-resets.com/tibo-28)
- Tibo’s post contains the promised global reset and its follow-up: “It’s done.” The embed shows 13 October at 08:06 in Latvia; the challenge assigns it to 12 October Pacific time. This is a timezone difference, not another reset. [Tibo Sottiaux: reset confirmation on X, 13 Oct 2026](https://x.com/thsottiaux/status/2109873287853707513)
- “tibo turned usage into gambling. burn your codex usage everyday and pray it's a reset day.” [Ananth on X, 4 Oct 2026](https://x.com/Ananth7e/status/2106847679632589075)
- On Claude's “permanent 25% increase”: 150% of the original before, 125% after. “Users get 17% less. Only the marketing improved.” [Kai on X, 30 Aug 2026](https://x.com/hqmank/status/2093861294827041058)
- “OAI says the new $500 Pro plan gives 25× Plus usage in Codex, vs 20× on the old $200 Pro. I measured it from my Codex logs, it does not.” [Trey Goff on X, 30 Sep 2026](https://x.com/thetreygoff/status/2105346970667688224)


---

## 08. OpenAI textGrain: an invisible watermark in ChatGPT text

OpenAI · Official · 5 Oct 2026

textGrain adds a statistical pattern to word choices, not hidden characters. OpenAI announced EU rollout to eligible ChatGPT and Codex text on 5 October; API customers can opt in for supported models. Copying the words can preserve the signal. Editing and translation can weaken it.

**For SEO, the watermark is not a quality verdict.** [Google permits appropriate AI use](https://developers.google.com/search/blog/2023/02/google-search-and-ai-content). Its [current guidance](https://developers.google.com/search/docs/fundamentals/using-gen-ai-content) warns against generating many pages without adding value. Neither source establishes a textGrain ranking penalty or confirms that Google uses OpenAI's detector. There is no basis here to promise a ranking loss, or a ranking gain from removing a watermark.

**You cannot check this with an ordinary ChatGPT subscription.** OpenAI currently restricts text-detector access to approved research and expert organizations. Its public verification tool checks supported images and audio. For most businesses, the useful action is an editorial quality check, not buying an AI-detector score.

![OpenAI chart: watermark detection rate falls as 10% or 25% of the words are replaced, for texts of 200 to 400 tokens](https://areteaios.com/ai-in-practice/2026-w42/openai-edits-chart.webp)

Source: [openai.com/index/eu-text-provenance](https://openai.com/index/eu-text-provenance/). OpenAI's own test: how often the detector still finds the watermark after 10% or 25% of the words are replaced.

### Practical usage

**How this affects business brain builders:** Give the publishing part of your business brain an editorial gate: original evidence, verified facts, a useful answer and a named reviewer. Track search outcomes separately from provenance.

**Who can actually test textGrain?**

[Apply through OpenAI](https://openai.com/form/content-provenance-api/) if your organization researches text provenance or detection reliability. Approval is case by case. The current help page excludes general commercial verification and checking student work. I have no approved detector access, so I cannot measure whether one of our drafts carries the signal.

**Prompt: Audit one SEO page against Google’s guidance (any assistant)**

```text
Audit this draft and its sources against Google Search guidance on helpful content and generative AI: [paste draft and source links]. Identify the search intent, unsupported factual claims, missing first-hand experience, generic passages and the specific value this page adds. Propose concrete additions from my real business material. Return a publish/revise checklist for a human reviewer. Do not invent experience, promise rankings, or claim to detect a textGrain watermark.
```

Try: [Read Google's guidance for website owners](https://developers.google.com/search/docs/fundamentals/using-gen-ai-content)

Access: Checked 16 Oct 2026: EU ChatGPT/Codex text rollout; API opt-in for supported models. Text detection requires approved access. A Pro, Business or API subscription alone does not grant it. See [OpenAI’s access rules](https://help.openai.com/en/articles/8912793-provenance-signals-in-openai-generated-content).

Status: Can't test yet

1. Pick one real service page or article. Save its current version and the reader question it should answer.
2. Use the audit prompt. Add your own example, evidence and concrete advice; verify every claim, including the title, description and structured data.
3. Have a person approve the final page. Keep a short record of sources, AI assistance and substantial edits.
4. After publishing, compare the same page and queries in Search Console over time. Record other changes; traffic movement alone cannot prove a watermark effect.

| Situation | Practical response |
|---|---|
| AI draft plus verified expertise and original examples | Review and publish if useful; AI involvement alone is not a Google ban. |
| Hundreds of generic pages made mainly to attract search traffic | Rework the content plan; this can be scaled content abuse. |
| A commercial detector labels a draft AI-written | That is not proof of an OpenAI textGrain signal. |
| No watermark is detected | That does not establish human authorship or content quality. |

**Does Google accept AI-written content?** Yes, appropriate use is allowed. [Google’s FAQ](https://developers.google.com/search/blog/2023/02/google-search-and-ai-content) focuses on quality and helpfulness, and warns that automation used primarily to manipulate rankings violates its policies. [The current guide](https://developers.google.com/search/docs/fundamentals/using-gen-ai-content) applies the same standard to text, metadata and structured data.

**Will Google penalize textGrain specifically?** The official Google pages linked here do not establish a textGrain penalty or detector integration. Treat predictions of a special watermark penalty as speculation. This is an interpretation of the published guidance, not a measured ranking experiment.

**Can a website or an API customer verify a sample today?** API customers can enable watermarking for supported models, but generating a signal and detecting it are separate capabilities. [OpenAI’s help page](https://help.openai.com/en/articles/8912793-provenance-signals-in-openai-generated-content) limits detector access to approved organizations. Public image/audio verification and general AI-writing detectors are not a public textGrain checker.

**Should I rewrite text to remove the watermark?** Rewrite to improve accuracy, substance and your own voice. A weak or absent watermark does not establish authorship, ownership or quality. OpenAI’s [technical report](https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf) describes detection limits; it is not an SEO instruction manual.

### What others are saying

- Co-author of textGrain: OpenAI is opening detector access to researchers, and the team will open-source the method so others can build on it. [Weijie Su on X, 5 Oct 2026](https://x.com/weijie444/status/2107183694981271553)
- Reads OpenAI's own limits as the headline: a rewrite or a translation erases it, short text often goes undetected, and only approved researchers get the detector. “A watermark nobody can check.” [Mahax on X, 5 Oct 2026](https://x.com/Mahaximus_/status/2107166444693692692)
- Unlike Anthropic and Google, OpenAI (for now) launches its watermark only in the EU, not globally. [Andrew Curran on X, 5 Oct 2026](https://x.com/AndrewCurran_/status/2107144242481271279)
- The weakness may suit OpenAI: “Users who don't want their ChatGPT use detected might switch to open-weight models if its watermarks become harder to remove.” [Matthias Bastian, The Decoder, 5 Oct 2026](https://the-decoder.com/openai-will-watermark-chatgpt-text-in-the-eu-but-makes-it-optional-for-api-users-worldwide/)


---

## 09. MiroFish: a simulated crowd reacts to your launch

MiroFish · Open source · 22 Dec 2025 · Catch-up

MiroFish takes seed material, builds a knowledge graph and creates AI personas that react on simulated social platforms. It runs on [OASIS](https://github.com/camel-ai/oasis), then produces a report and lets you question the agents. The useful question is which objections to investigate before a launch.

**Who tried it, and why?** [Sanjay Shankar](https://sanjayshankar.me/ai-market-simulation-mirofish-startup-launch/) reports testing a SaaS launch before building it: a 6–8 page market brief, 44 agents and 10 rounds. The recurring concern was getting a whole team to adopt the product. He says this shifted his attention toward onboarding and internal champions. That is an author's account of a simulation, not validated buyer demand.

It came up in our AIOS Circle call on 14 October. I have not run it. Start with the maintainer's prepared examples below and compare any proposed simulation with one ordinary prompt. The [small comparison shared on Hacker News](https://news.ycombinator.com/item?id=48635080) tested a simplified swarm against a single call, not MiroFish itself.

![MiroFish maintainer screenshot: generated report beside the report agent’s progress and evidence](https://areteaios.com/ai-in-practice/2026-w42/mirofish-screen-4.png)

Source: [github.com/666ghj/MiroFish](https://github.com/666ghj/MiroFish). The maintainer's own report-generation screen. The report and progress panel are inspectable; the screenshot does not establish forecasting accuracy.

### Practical usage

**How this affects business brain builders:** Use a synthetic audience to collect objections before real interviews. Your business brain should preserve which reactions were simulated and which came from actual customers.

**A practical example, without claiming my own test**

I have not installed or run a paid simulation. Sanjay’s case is a self-report. The screenshots and demo video below come from the project maintainer; viewing them is not a new simulation or a test of prediction accuracy.

#### Inspect the generated report

![Official MiroFish screenshot: a report with section planning and generation progress](https://areteaios.com/ai-in-practice/2026-w42/mirofish-screen-4.png)

Maintainer example: a report about a semiconductor company's funding scenario. The left panel contains the narrative; the right shows the report agent's work. It illustrates the output format, not a verified investment forecast.

Source: [MiroFish · official screenshots](https://github.com/666ghj/MiroFish#-screenshots)

#### Question the simulated world

![Official MiroFish screenshot: Report Agent chat and an option to interview simulated personas](https://areteaios.com/ai-in-practice/2026-w42/mirofish-screen-5.png)

Maintainer example: ask the report agent a follow-up or interview a persona. The original demonstration examines possible reactions to a Wuhan University public statement. Watch the full walkthrough to see the input, simulation and report.

Source: [Maintainer · full Bilibili demo](https://www.bilibili.com/video/BV1VYBsBHEMY/)

**Prompt: Set up MiroFish on my computer (Claude Code or Codex)**

```text
Set up MiroFish from github.com/666ghj/MiroFish on my computer with Docker. Read the README first and list the keys it needs (LLM_API_KEY, LLM_BASE_URL, LLM_MODEL_NAME, ZEP_API_KEY) before you change anything. Use a cheap OpenAI-compatible model and keep simulations under 40 rounds, as the README advises. Do not upload any client or private data. Then run one simulation with my public offer page as the seed and summarise the five objections that came up most often.
```

**Prompt: The same question as one plain prompt, to compare**

```text
Here is my offer page. Act as five different potential buyers: a sceptical owner of a 10-person company, a CFO, an operations manager, a marketing lead and a founder who already uses ChatGPT every day. For each, give the first objection, what would make them book a call and one question they would ask. Then list the objections that came up more than once.
```

Try: [Open the MiroFish live demo](https://666ghj.github.io/mirofish-demo/)

Access: Free and open source (AGPL-3.0). Runs on your own computer with Docker, or from source with Node.js 18+, Python 3.11 to 3.12 and uv. Needs your own OpenAI-compatible LLM key and a Zep Cloud key; the README says Zep's free monthly quota is enough for simple use and warns that simulations use a lot of tokens. As of 15 Oct 2026.

Status: Not tested yet

1. Read the README and the AGPL-3.0 license, and install from the official GitHub repository only.
2. With Docker: copy `.env.example` to `.env`, add the four keys, run `docker compose up -d` and open localhost:3000.
3. Seed one simulation with a public page, such as your offer page. Keep it under 40 rounds and watch the token use.
4. Ask the same question with one plain prompt to your usual AI, compare the objections, and keep only the ones you also hear from real customers.

#### A concrete launch experiment someone reported

Source: Sanjay Shankar, 25 Mar 2026, updated 10 Sep. These are his reported inputs and conclusions; no observed sales result is established.

| Who and purpose | Input and run | What he took from it |
|---|---|---|
| Sanjay Shankar; inspect objections before building a SaaS product | Market brief; 44 AI personas; 10 rounds | Team adoption friction; revisit onboarding and internal champions. |

**Is a paid MiroFish website the real thing?** Not that we can tell. The repository links to mirofish.ai, which shows no pricing. Several unaffiliated look-alike sites, for example mirofish.us and mirofishai.org, sell plans. Install from the GitHub repository.

**Can it run without any cloud service?** Not the official version: it calls a cloud LLM and Zep Cloud. A community fork, [MiroFish-Offline](https://github.com/nikmcfly/MiroFish-Offline), swaps in Neo4j and Ollama and needs about 10 GB of GPU memory for a 14B model. It is not official.

**Does the five-case swarm comparison prove MiroFish is no better than a single prompt?** No. The author compared a simplified swarm in a separate local rig. It is a reason to use a cheap baseline in your own experiment, not a benchmark of MiroFish.

### What others are saying

- Reports a pre-build SaaS-launch simulation with 44 personas and 10 rounds. He used the objections to reconsider onboarding. No verified demand or conversion outcome is shown. [Sanjay Shankar, his own experiment, 10 Sep 2026](https://sanjayshankar.me/ai-market-simulation-mirofish-startup-launch/)
- A five-case test of a separate simplified swarm, not MiroFish: a single LLM call matched or beat that simulation. Useful as a reminder to compare against a simple baseline. [Show HN: a local rig to test AI social simulation, 22 Jun 2026](https://news.ycombinator.com/item?id=48635080)
- Profile of the student developer and the 30 million yuan commitment from Shanda's Chen Tianqiao. [36Kr, 13 Mar 2026](https://eu.36kr.com/en/p/3720728841763465)

---

## Also shipped

- 15 Oct: Google AI Search is putting more inline links higher in answers and testing a dedicated section for official sources (15 Oct). For a business website, check whether useful original pages are cited and receive visits; this update does not guarantee traffic. [Google · AI Search source links](https://blog.google/products-and-platforms/products/search/updates-ai-mode-overviews-explore-web/)
- 8 Oct: ChatGPT Fast and Ultrafast: two paid speed modes for Work and Codex. Fast uses your allowance at 2.5 times the normal rate; Ultrafast, up to 8 times faster, is only on the $500 Pro tier and eligible Enterprise and Edu. GPT-6.1 Sol Ultrafast arrived on 8 Oct. [learn.chatgpt.com](https://learn.chatgpt.com/docs/agent-configuration/speed)
- 7 Oct: GPT-6 for everyone, with “Intelligent UI”: answers can include interactive charts and small calculators. [openai.com](https://openai.com/index/gpt-6-for-everyone/)
- 29 Sep: ChatGPT Space and Pages: a shared workspace and editable documents for teams and their Dots, on Pro, Business and Enterprise. [openai.com](https://openai.com/index/devday-2026-recap/)
- 7 Oct: Claude Haiku 5.5: a small model for summaries, routing and other narrow tasks. API input/output prices per million tokens are $0.10/$0.50 up to 100K prompt tokens, and $0.50/$2.50 above 100K (checked 16 Oct). Anthropic’s launch table below compares it with Haiku 4.5, GPT-6 Luna and Sonnet 5.5. [Anthropic · Haiku release and evaluations](https://www.anthropic.com/claude-haiku-5-5)

![Anthropic official Haiku 5.5 evaluation table compared with Haiku 4.5, GPT-6 Luna and Sonnet 5.5](https://areteaios.com/ai-in-practice/2026-w42/haiku-benchmarks.jpg)

Anthropic’s own 7 October release table, captured 16 October. Settings differ by benchmark. A cheaper small model does not match Sonnet on every task.

- 7 Oct: Claude Max and Team now include monthly API credits for your own apps and agents. [support.claude.com](https://support.claude.com/en/articles/12138966-release-notes)
- 8 Oct: Google's Gemini agent can work as a “coworker” with its own Workspace email, calendar and Drive. [cloud.google.com](https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026)
- 9 Oct: UC Berkeley study: after a few minutes of AI help on maths and reading tasks, people gave up sooner once the AI was taken away. [news.berkeley.edu](https://news.berkeley.edu/2026/10/09/using-ai-for-just-10-minutes-erodes-your-ability-to-persist-at-hard-things/)
- 13 Oct: An OpenAI employee was charged after a hidden recording device was found in an executive conference room. OpenAI [told NBC News](https://www.nbcnews.com/tech/tech-news/openai-employee-arrested-hidden-listening-device-company-says-rcna603262) its internal AI models helped identify who planted it; the DOJ cites CCTV. [justice.gov](https://www.justice.gov/opa/pr/saudi-national-arrested-installing-surreptitious-audio-recording-device-openai)

---

Items marked Catch-up are older than two weeks but too important to skip. From next Friday the issue covers one week. Found a link that changed how you work? Send it to egils@areteaios.com.