OpenAI has added an Apple Messages plug-in to the ChatGPT desktop app on Mac. Connect it, and ChatGPT can search your conversations, catch you up on what you've missed, draft replies, and with your approval actually send them. It works across iMessage, SMS and regular text messages, on Apple silicon Macs.

The useful part is real. A surprising amount of daily admin is really just texting: "reply to Sarah with a few times I'm free next week", "what did I promise the plumber?", "find the follow-ups I owe from yesterday". That's now a prompt instead of twenty minutes of scrolling.

The part worth pausing on: you're giving an AI standing access to your private messages. To OpenAI's credit, the plug-in runs locally on your Mac, doesn't build an index of your message history, and nothing sends until you approve the message and the recipient.

My honest read: worth experimenting with for catching up and drafting, but keep the final send in your own hands. There's an option called persistent approval that lets ChatGPT skip the confirmation step. OpenAI's own docs warn it removes your last chance to review a message before it goes out as you. Leave it off.

Try this: once connected, start with a read-only job:

xAI launched Grok Bot in beta this month: "always-on AI teammates" you message like colleagues, from your desktop or iPhone. Each Bot works on a computer in the cloud, which means it keeps working after you close your laptop.

Two things make this one different from the wave of agent launches before it:

The pitch is essentially "hire AI like a staff member": xAI says its own teams use Bots for inbox management, CRM updates, outbound research and expense processing. That's small-business admin, word for word.

My honest read: this is the most convincing version of the AI-teammate idea so far, but it's brand new and the fine print matters. It's included with SuperGrok plans (business access is waitlisted), and xAI's own docs note that your Bots share one cloud computer. Their words, not mine: don't treat separate Bots as a security boundary. So don't hand it your banking login on day one.

Try this: pick one recurring prep task (the weekly report, say, or the Monday research round-up), show a Bot how you do it once, and set the routine to deliver to you for approval, not act on its own. Same rule as the story above: the approval step is yours to keep.

Not sure where to start? I shared a quick video on Instagram about a directory of 150+ free, open-source Grok Bots, where every listing is one prompt that builds the bot for you. No signup, and honestly the directory isn't even the useful part: open three listings and steal the structure. Worth two minutes before you build your first one.

Anthropic launched Claude Academy, a free learning platform with courses, tutorials and real-world use cases built partly from how Anthropic trains its own staff to work with AI.

The courses are free and open to anyone, you can track progress and earn badges, and much of it is deliberately not Claude-specific.

What this could actually look like for you:

If "get proper AI training" has been sitting on your someday list, this is the lowest-friction version of it that currently exists. Start at academy.claude.com.

Read Anthropic's announcement →

Reddit's share of the sources ChatGPT cites in its search answers fell 86% in about a fortnight, from a 3.83% average to just 0.52% by mid-August, according to tracking firm Promptwatch (reported by Axios). Nobody knows exactly why yet, and Google's AI results barely moved over the same period.

A lot of "get recommended by AI" advice lately boils down to "get talked about on Reddit". This fortnight is your reminder that AI answer engines are a moving target: a source they love today can quietly drop to near zero next month.

The practical rule: don't bet your AI visibility on any single channel. Spread your footprint (your own site, genuine reviews, industry press, directories) so no one algorithm change can erase you.

(For AI Systems Lab members: this is exactly the diversification logic from our AEO/GEO playbook call.)

Read the Axios report →

Adobe rolled out Firefly's audio tools to everyone: Generate Music creates original tracks tuned to your video's length and mood, Generate Speech turns a script into a voiceover with control over pacing and emotion, and Generate Sound Effects fills in the rest. There's now a free tier with daily limits, so you can try it without paying.

The bit that matters for business use: Adobe says the output is commercially safe. The music is original and licensed, so you're not gambling on a trending audio clip getting your ad or reel muted or taken down.

Try this: next time you make a promo video or reel, generate a 30-second track in Firefly to match it instead of borrowing audio. Two minutes of work, zero takedown risk.

Read Adobe's announcement →

Everything in this newsletter circles the same question: how much should AI do for you before you ask?

On Thursday's Lab call, we answered it properly. The session was all about proactive agents: AI systems that own one recurring job, collect fresh evidence on a schedule, and deliver finished, checkable preparation before you've even opened your laptop.

The key idea: proactive doesn't have to mean autonomous. The agent researches, analyses and drafts, and you keep the approval gate before anything is published, sent or spent. We walked through the seven-stage loop behind a reliable agent, looked inside two live examples, and filled in a one-page spec you can use to build your first one this week.

If you've prepared the same report or research three times in the last month, you already have your first candidate.