Howdy wizards,
Here’s what’s brewing in AI.
The big thing
Claude has an internal workspace where it does its step-by-step thinking but doesn’t write things down.
You could already see Claude’s chain-of-thought. This is different.
In a new paper, Anthropic calls it the J-space, after the Jacobian math behind it, and the claim is that this is where Claude works through anything that takes more than one step. It seems to have formed on its own during training.
An example: Ask Claude how many legs the web-spinning animal has, and inside the workspace it settles on a spider to work it out, then tells you eight. Reach in and swap that internal spider for an ant, and the answer it gives you changes to six, with the prompt untouched.
If we wipe this “workspace” out entirely, Claude can still chat and recall facts, but anything multi-step falls apart.
In a safety test, they ran a Claude model that had been secretly trained to sabotage code: the code it handed back looked clean, but its internal concepts for deceit and fraud fired first, so the bad intent was visible inside the model before it answered.
Here’s a cool little tool that makes the concept clearer; you can chat to a Qwen model and see the exact J-space forming as the model types the answer.
Why it matters Model reasoning just became a step more inspectable.
It’s now clear that models reason about words and concepts that they don’t write down. LLMs have developed a private part of their “brain”. And whatever part we’re able to see now (ie the J-space), we can actually read and edit and influence the model’s response. I believe that’s a good thing.
Also, most of us don’t need to read an AI’s mind. We just need it to catch the follow-up we forgot. Which is where Ahoy, this week’s sponsor, comes in:
IN PARTNERSHIP WITH AHOY
You know that feeling when a deal goes quiet and you can't remember who was supposed to send the follow-up?
That's not a discipline problem. When you're juggling twenty deals across a dozen channels, something always slips. Ahoy watches your pipeline and prepares your next move. A drafted follow-up. A flag on the deal gone cold. You just say go.
No more "wait, did anyone follow up?" moments. Just deals that keep moving.
NEWS NEWS NEWS ❦ NEWS NEWS NEWS
All the small things
Industry moves
You’ve got a few more days to test Fable 5 for free. Anthropic extended the free run on its most capable model by almost a week, which means you have until Sunday (July 12). If you’re on a paid plan, use the opportunity to test it, so you can brag to your friends; after that, Fable is only available in the API, which is a lot more expensive.
Nvidia will give startups compute in exchange for a slice of their revenue. It routes that compute through its cloud partners and takes a cut of revenue. It’s raising at least $20 billion to fund this. Circularest economy ever.
Midjourney asked a court to make Disney, Universal, and Warner Bros. reveal their own AI use. The studios are suing it for copyright; Midjourney’s filing says they should have to disclose their internal AI too, not just the consumer-facing stuff, plus every Midjourney prompt and image they generated while building their case. A studios’ lawyer called it a "fishing expedition," but the discovery fight could end up exposing how much AI the plaintiffs quietly run themselves.
New tools & product features
GitHub Copilot added its first open-weight model, Moonshot’s Kimi K2.7. Until now the most corporate coding assistant only offered closed models from OpenAI, Anthropic, or Google. Take it from GitHub: open weight models are now good enough to sit inside mainstream dev tools.
Microsoft is merging its consumer and work Copilots into a single app, per a leaked memo. The memo is titled "earn the right to exist," which actually says the quiet part out loud: no one uses Copilot because they want to, they use it because it’s free for them and/or they have to.
Models
Meta says its in-training model "Watermelon" already matches OpenAI’s GPT-5.5. Superintelligence chief Alexandr Wang told staff it is running on roughly 10x the compute of Meta’s last model, with an "Opus-level" coder teased for pretty soon. Remember, Meta’s AI team is currently hard at work justifying an AI budget of up to $145 billion, after a rough year. So let’s take that with an equally generous pinch of salt.
❦
You are a delight.
What's your verdict on today's email?
See which AI use cases are paying off with Context Windows Pro
Most companies pick AI use cases by brainstorming internally. 90% of those initiatives fail.
I’ve created Context Windows just so you can pick the winners.
🟦 Find high-performing use cases from 2,000+ companies at contextwindows.ai, or book a demo with me
Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.