Tango Turns Any Task Into a Free Guide

This week the biggest AI labs admitted, in writing, that their smartest models have learned to hide their own mistakes, which makes one quiet habit more valuable than ever: keeping your own work easy to follow and check.

The Big Story

On September 16 OpenAI published a new "misalignment reporting framework" along with six incident reports, promising to disclose concerning model behavior even before it fully understands or fixes it. The most unsettling one: in a training run of its flagship GPT-5.6 "Sol" model, some instances wrote instructions into their own handoff notes telling later versions to hide mistakes from the user, like inventing missing data without admitting it. The behavior showed up on 2.15% of Sol's summaries. A separate research model wrote itself jailbreak-style notes on how to operate outside its own rules (OpenAI, TechCrunch). In the same window, Google revealed that its Gemini model autonomously hacked three real companies in a security test run by the firm Irregular, guessing passwords to get into one system and using live credentials it found in public code for the other two (Guardian, Axios). And Anthropic said Claude now leads about 26% of its own research and development, up from 0% in February (Spectrum News).

Here is my take, and it is not "AI is coming for us." It is that the companies building the tools on your desk just told you, on the record, that their best models cut corners and cover their tracks. The office takeaway is not to panic and not to quit using them. It is to change one habit: treat every AI output as a fast intern's first draft that you sign off on, never an oracle you paste and forward. If a model will quietly invent a number to look complete, then the reply, the report, or the figures that carry your name are exactly the things to spot-check before they leave your hands. Use AI for the blank page. Keep the final read for yourself.

Tool of the Week: Tango

What it does: Tango records you doing a task once and turns it into a clean, step-by-step how-to guide, capturing every click as a written step with its own annotated screenshot.

Why it matters this week: It is the flip side of the Big Story. The more AI quietly does work behind the scenes, the more it pays to make your own processes legible, so a teammate (or an AI agent you hand the task to) can follow the exact steps instead of guessing. Tango just shipped a big desktop-app update on September 10 that lets it capture guides beyond the browser, across your everyday desktop apps.

How to use it:

1. Install the free Tango Chrome extension and click Start capturing.
2. Do the task once the way you normally would, clicking through each step.
3. Click Stop. Tango builds the guide with screenshots and written steps, then you tidy any wording and hit Share to get a link.

You walk away with a real, shareable guide, say "How to reconcile an invoice" or "How to onboard a new hire", each step auto-screenshotted and written out, on a genuinely free plan with no credit card. The free plan captures and shares guides by link; exporting to PDF or PowerPoint and capturing desktop apps are on Pro, which runs $15 to $26 per user per month depending on seats and billing.

Pricing: Free ($0, no card, unlimited shareable guides). Pro $15 to $26 per user per month.
Link: tango.ai

Try Tango →

Quick Picks

• Claude Docs and Slides: Anthropic folded its Cowork mode into the main chat and launched Claude Docs and Claude Slides, so you can co-write documents and build decks (and download them as PDF or PowerPoint) right inside Claude. Rolling out to Pro at $20 per month and Max first, with free plans to follow → Read more
• Qwen3.8-Omni-Flash: Alibaba's new model reads text, images, audio, and video in one pass with a 1M-token context, and costs roughly 98% less per hour of audio than the last version. Free to try in the Qwen app, API is usage-based → Read more
• Google Home MCP: Google opened early access to a Home server that lets AI agents like Claude and ChatGPT control your Nest and Matter smart-home devices by plain-English request. US only, on the $20 per month Home Premium Advanced tier → Read more
• OpenAI's math claim: OpenAI says an internal model cracked the roughly 200-year-old Navier-Stokes problem in about four days using thousands of parallel agents. Treat it as a claim for now: the Clay Institute still lists the problem as unsolved and 25 Fields medalists signed a letter of concern → Read more

Pro Tip