🔈 Listen on Spotify

Hi! Welcome to the 60th 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.

_{~8 mins read}

/Alibaba /OpenWeights /Agents

Alibaba released Qwen3.8-Max yesterday morning: 2.4 trillion parameters, comfortably its most capable model yet. But the size isn't the story. The story is that Alibaba is publishing the weights next week for free, so anyone can download and run them on their own machines. That's the first time a top-tier Qwen model has been given away, and no American lab is doing this at the frontier.

The demo is what will keep other labs up at night. Alibaba pointed the model at an empty folder, told it to build a command-line tool, and left it alone. Sixteen days later: 265 commits, 127 pull requests, 151 issues, all raised, claimed, coded, tested and merged without a human touching it. The repository is public and still going. We checked it this morning: 221 pull requests, 272 stars, last update a few hours ago. In a separate test, the model was handed a recent research paper and five days of GPU time. It rebuilt the experiment from scratch with no starter code, reproduced all six findings, then ran 18 experiments of its own and landed on a method that beats the paper.

The packaging is deliberate too. The launch video shows the model working on chip design, which is not a subtle choice. Chips are the one thing China cannot simply buy, and export controls have pushed Beijing hard towards building its own. The implied message: we now have enough intelligence to engineer our way around the shortage. Whether that turns out to be true is a genuinely open question. The confidence is not.

For a business, geopolitics matters less than the consequences. Open weights mean a frontier-class model you can run on your own hardware: no per-token bill, no subscription, no data leaving your building. Put that next to OpenAI cutting its cheap-model prices by 80% last week, and the direction is hard to miss. The price of serious AI is collapsing from both ends at once.

/OpenAI /Browser /Productivity

OpenAI is rolling out a properly upgraded Chrome extension and desktop app this week. The headline feature is Side Chat: a ChatGPT panel that sits next to whatever you're already looking at. You can ask it about a YouTube video while it plays, point it at your other open tabs, or highlight a paragraph on a page and ask a question about it without going anywhere. The desktop app gets URL suggestions as you type and can look back through your browsing history, with controls over how that's handled.

Why this matters more than it sounds: the reason AI often feels like more work is the copy-paste tax. You spend your day ferrying context into a chat box, one clipboard at a time. Every serious player is now racing to close that gap. Google put Gemini in Chrome, Anthropic put Claude in the browser, and OpenAI is now doing the same from both ends.

The caveat deserves saying out loud, because it's the whole trade: this only works because the thing can see your tabs and your history. That's genuinely useful and genuinely sensitive. Have the POPIA conversation before you roll it out to a team handling client files.

/Buzz /Agents /Teams

Buzz is a new team chat app backed by Jack Dorsey's Block. It looks like Slack. The difference is that AI agents are members of the team, not bolted-on integrations. You write each agent its instructions, drop it in a channel, and tag it like a colleague.

Two ideas worth stealing even if you never install it. First, the swappable harness: you can run Claude Code, Codex or Goose underneath any agent and switch between them, keeping all the chat history and everything you've taught it. Change the model; keep the memory. Second, shared compute: a small team can buy one decent Mac, run a free local model on it, and everyone works off that from their own laptop.

It's early preview software, and it's rough around the edges. But the question underneath it is a good one: if agents do the work, why do they live outside the room where the work gets discussed?

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

Every business has the same graveyard: a folder of PDFs nobody can search. Invoices, contracts, annual reports, that 200-page tender document. The usual advice is to run them through OCR, the technology that reads text off a picture. It's slow, and it usually costs money.

Here's the thing: roughly half of all PDFs don't need OCR at all. The text is already sitting inside them as text, just wrapped in a format built for printing rather than reading. Firecrawl has just released a free, open-source tool called pdf-inspector that pulls that text straight out, on your own machine, with headings, tables and lists intact. Their benchmark: 200 PDFs processed in under half a second.

Now the honest bit. It's a developer tool. But you don't need to be a developer to use it, because you have an AI that is one. That's the actual lesson this week: you're not learning to code, you're learning to delegate.

Don’t believe us? Try it yourself…

Pro tip: this only works on PDFs that were born digital. If someone photographed or scanned a document, it's a picture, and you'll still need OCR. The tool tells you which is which, so run it first and you'll know exactly how much of your pile is the easy kind. It's free, MIT licensed, and runs entirely on your own machine, which means client documents never leave your laptop. That's a real POPIA advantage over uploading the lot to a cloud service.

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

OpenAI has published ten new results in pure mathematics and theoretical computer science. Every one had been an open problem for at least a decade, most of them far longer. They span eight fields: high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and more. Among them are a disproof of Connes's rigidity conjecture, and answers to three problems posed by the legendary Paul Erdős.

The work was done by an internal version of Astra, OpenAI's next major model. Humans turned the raw arguments into papers. Then, crucially, the model translated each proof into a Lean certificate: a machine-checkable format that verifies the mathematics is correct without anyone having to take OpenAI's word for it.

And now the number that should stop you. The total compute cost of finding all ten solutions was roughly $2,000, about R35,000. Problems that resisted the world's best mathematicians for decades, cracked for less than a decent second-hand bakkie.

The honest caveat: this is OpenAI marking its own homework, and the mathematical community hasn't had time to digest it. Plenty of mathematicians are uneasy about where this goes. Credit where it's due though: OpenAI is unusually straight about attribution. It says plainly that the arguments were generated by the system rather than by humans, and that claiming otherwise would misrepresent both. It has also handed 100,000 researchers free access to its best models, which is the fastest way to find out whether these results hold up.

So why should this matter to a business? Nobody's asking you to solve Erdős problems. The lesson is about where the bottleneck now sits. If a machine can produce publishable results across eight disciplines for $2,000, then for a growing class of hard problems the constraint is no longer expertise or budget. It's whether anybody thought to point the thing at the problem.

Every business has its own version: the pricing model nobody has properly optimised, the delivery route everybody agrees is a nightmare, the roster that has been "just how it is" since 2019. Some of those are genuinely hard. Most are simply unattended.

Bottom line: cheap intelligence quietly changes what counts as an unsolvable problem. This week, go and write down the three things in your business everyone has stopped questioning. That list is your actual AI strategy.

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 👋