What's happening

On Monday, OpenAI quietly filed paperwork to go public. The reported target is a valuation of up to $1 trillion USD, with a market debut possible as early as September. The same day, Sam Altman and Jakub Pachocki published a plan called "Built to benefit everyone." It argues that concentrated power makes societies fragile, and that AI's upside should be spread as widely as possible. Read the full plan →

What it means

Read those two things back to back and something does not sit right.

A company files to raise tens of billions on the public market, at one of the largest valuations in history, while publishing a document about not letting a few institutions hold most of the capability and most of the upside.

Both can be true. You can believe in spreading the benefits and still want a trillion-dollar listing. But once OpenAI answers to public shareholders, "benefit everyone" has to share the table with "return value to investors." When those two pull apart, the share price usually wins.

The plan does drop one real marker worth keeping: Altman expects a meaningful share of OpenAI's own research to be done by AI by March 2028. Take that part seriously.

Why you should care

The AI you are starting to lean on is about to be run by a public company. That is not automatically bad. It does mean its incentives are shifting, and you should read its promises with that in mind. When a business tells you what it stands for right before it asks the market for money, watch what it files, not what it writes.

SpaceX has signed Google up for around 110,000 Nvidia AI chips at $920 million USD a month, from October through to mid-2029. That is about $30 billion USD in total, landing the week before SpaceX is tipped to list at over $1.7 trillion USD. The strange part: SpaceX built that capacity for Elon Musk's own AI company and is now leasing the spare to a rival. When the picks and shovels pay this well, you can see why everyone is sprinting to the market at once. Read more →

Apple's "boring" AI plan is starting to look clever

While rivals commit a combined $900 billion USD to AI this year, Apple is spending about $14 billion USD, and it still posted record iPhone sales. Its bet is simple: skip the arms race, bake a few useful AI features into the phone you already own, and let everyone else burn cash proving the technology. Slow is not always behind. Sometimes it is just patient. Read more →

One to watch: Anthropic is expected to release Mythos to the public

Mythos is Anthropic's model for finding software security flaws, and in restricted testing it has already uncovered thousands of serious ones. A public release is rumoured within days, though Anthropic has not confirmed it. If it lands, expect a real debate about AI that can break software as easily as it patches it. Read more →

What I tested

I spent a week building a personal knowledge system out of plain markdown files. Obsidian holds the notes, Claude Code is the brain that reads and writes across them. Most people keep four different chatbots with four separate memories. This turns that into one shared memory you can ask in plain words.

Caption: Most people switch between ChatGPT, Claude, Gemini and Perplexity, and each one forgets you the moment you leave. The whole idea here is one shared memory.

How I set it up

I backloaded everything I had into it: old Claude sessions, imported context, notes, transcripts. The files are free. I run a Claude Max subscription and pay about $4 AUD a month for Obsidian Sync so it spans both my devices. That is the whole bill. The idea is not mine. Andrej Karpathy, who co-founded OpenAI and ran AI at Tesla, posted a version and called it a hacky collection of scripts. His is read-only, built per topic. Mine vets what goes in, links it cheaply, and grows itself. I call it big brain.

Caption: Karpathy's original on the left: drop in raw notes, the AI compiles them into a wiki you can ask. Mine on the right adds the missing layers, a step that checks what is actually true and a system that grows on its own.

What actually happened

The real cost is one week of heavy Claude usage while you set it up. Push through that week and it settles. What you are left with is not a filing cabinet. It does the retrieval and the thinking for you, across everything you have fed it, and it surfaces connections I would never have found on my own.

Who should try it

If you are getting into AI and do not know where to start, start here. Build the architecture first, and every project after it gets captured in the same place. One honest warning: someone will turn this into a one-click product soon, and when they do, everyone gets a personal intelligence layer and the edge fades. Right now, for a short window, the people who build it themselves are ahead of nearly everyone they meet. That window is open today.

Resources

That's the week in AI. If something sparked a thought, hit reply. I read every one.

Back again soon. Maran