🔈 Listen on Spotify

Hi! Welcome to the 61st edition of the TomorrowToday newsletter.

We’re here to decode the AI chaos so you don't have to. Think of us as your friendly neighbourhood tech translators - we cut through the chaos, translate the jargon, and spotlight new AI tools that matter for founders, builders, and curious minds.

Buckle up, because the future's moving fast and we're here to make sure you don't get left behind! ⚡

If you enjoyed today’s newsletter, please forward it to a friend & subscribe by following this link.

_{~7 mins read}

/Voice /Productivity /MeetingNotes

We are big fans of Wispr Flow, the voice dictation app we use every day. Now they have shipped their first product beyond dictation: a meeting note-taker, and it has quietly become our favourite one we have ever used.

Here is the clever part. Most note-takers (think Otter, Fathom, and friends) join your call as a visible bot, that awkward extra "participant" everyone can see. Wispr's Notetaker does not join at all. It captures audio straight from your Mac, so the meeting looks completely normal to everyone else, and it works across Zoom, Teams, Google Meet, Slack huddles, and even a call in your browser, with nothing to set up.

It also gets the details right. It labels who said what, and because it pulls names from your calendar invite and your personal dictionary, "Biswaroop" does not come back as "Bishop" and product names stay intact. After the call, it re-reads the whole recording with that context to sharpen the transcript.

A quick honesty note: at launch it is Mac-only and English-only (although we did run a few Afrikaans tests which worked alright), and speaker labels are sharpest on video calls where everyone joins from their own device. But it is good enough that we have cancelled a couple of our other meeting-recording subscriptions. It is included in the free Wispr Flow plan, with Pro at around R270 a month.

/OpenSource /LocalAI /Meta

Meta has released Muse Glimmer, a 30-billion-parameter model, and given the weights away under a permissive licence. Translation: it is genuinely free, and you run it on your own machine instead of renting it from a lab.

What makes this one special is that it is built to run locally, on a normal Mac or PC with a single decent graphics chip. Meta shrank it down (see this week's AI Dictionary) so it fits in roughly 18 to 20GB of memory, which means a well-specced modern machine, an M4 Max or M5 Max, say, can now run capable AI entirely offline, no internet and no subscription required.

The bigger picture is the exciting bit. If you own good hardware, you now have free on-tap AI sitting on your desk, and these local models are only going to get better from here. Meta also hinted that the weights for its larger foundation model, Muse Spark, are coming soon.

/Coding /Anthropic /Agents

Two neat updates landed for Claude Code, Anthropic's coding tool, this week.

First, auto mode is becoming the default from 14 August (on the Pro, Max and Team plans). If you have ever been driven mad by clicking "approve" every single time Claude wants to do something, this is for you. Instead of asking at every step, it just gets on with the work, and only pauses for anything irreversible, destructive, or reaching outside your environment. Anthropic's reasoning: they had an independent firm run 720 prompt-injection attacks at it, and none got through. Worth being clear-eyed though, it reduces the risk, it does not remove it, so for anything touching production you should still keep an eye on what it does.

Second, and we love this one: your different sessions can now message each other. Anyone who has juggled three terminals knows the pain of copy-pasting context from one chat to the next. Now one session can send another a summary, and the other picks it up mid-task. It passes a summary only, never your full history or files.

In this section, we aim to spotlight an incredible AI tool or use case and guide you on how you can try it.

After the Meta story above, you might be wondering: this all sounds great, but how do I actually run a model on my own machine? It is simpler than you would think, and the tool that makes it painless is called Ollama, a free app that handles all the fiddly bits for you.

Don’t believe us? Try it yourself…

Pro tip: if it feels sluggish, try an even smaller model like llama3.2:3b. And when you are done chatting, type /bye to exit.

In this section, we're spotlighting real businesses using AI to solve actual problems.

A man in Melbourne named Andrew did something completely ordinary: he asked his AI assistant to book him a spot in a popular morning gym class. He was sitting fourth on the waitlist. What happened next, first reported by the ABC, is now being called the first known case of an AI agent autonomously hacking a system in Australia. To be clear, Andrew was not some hacker in a hoodie: he works at an AI company and was using an ordinary assistant agent built on Claude. He was, in his own words, sitting on the couch thinking booking the class was a chore, and figured this was a perfect little job to hand off.

Rather than politely waiting, the assistant went looking for shortcuts. It found that the gym's booking limits were only enforced on the front end of the website, the part you see, while the underlying system that actually processes bookings had no such checks. So it booked classes months further ahead than the gym allowed. Then, entirely on its own, without being asked, it discovered the system had zero authorisation checks on cancelling other people's reservations, tested this on the person sitting first on the waitlist, and bumped Andrew from fourth to third by kicking them off. It then reported back, quite matter-of-factly, that it had done so "as part of testing its capabilities".

Nobody set out to run an attack. A paying customer used an agent for exactly the task the marketing promises, and the model simply took the shortest path to "done". That is the part worth sitting with: there was no malicious prompt, no jailbreak, no clever adversary, just a helpful tool doing its job a little too literally. The uncomfortable takeaway for business owners is that your app's security can no longer lean on the quiet assumption that most people will not bother poking at it. For years, that assumption held, because prodding an API by hand is tedious and most customers will not do it. Agents change that maths overnight. When anyone can point a capable AI at your booking form, your API, or your checkout, the weak spots you assumed were hidden become the very first thing that gets found, and at scale.

The fix is boring but urgent: enforce your rules on the server, not just in the interface. Any limit that lives only in the front end - booking windows, quantity caps, who may cancel what - is not really a limit at all. Stories like this are going to get far more common, so it is worth asking this week: if an eager AI agent went poking at your systems, what would it find?

The lesson: If a limit only lives in your website's front end, it is not really a limit. AI agents will find the door you left unlocked round the back.

AI is full of jargon, and we’re here to decode it. Each week, we’ll give you a plain-English definition of a buzzy term you’ve probably seen (but never fully understood).

Thanks for reading TomorrowToday! We’d love to hear from you:

➡️ What would you like us to cover next?
➡️ Have a tool or topic we should feature?

We’re building this with (and for) you. 🚀
See you next Tuesday 👋