📩 just @iMessage in chatgpt

.Yo yo folks,

📜 TLDR:

• Saw the Stripe and Canva founders chat

• A lab, stealth-dropped a model (and it’s better than Fable)

• You can now @Messages in your ChatGPT and it can read your iMessages.

• And how to start a startup with Paul Graham

💼 What’s new in the business

Not much to report that’s changing in the business.

But in some cool news. Went to the Stripe event this week. Saw my mates speak a Stripe (the boys at fastlane), which was pretty sick.

Also was in the room when John Collison (founder of Stripe) and Cliff Obrecht (founder of Canva) were having a yarn.

Was really impressed with the chat.

One quote that stuck with me (by Cliff), “By building the best product in the market, you're guaranteed to be a big company.” 

Sometimes it really is that simple.

Fun fact: Canva have 200 million free users and 30 million paying ones - their free tier was in essence their growth mechanism and way to acquire clients

Then went and worked out of Lyra’s head office (shoutout to the boys there, absolutely killing it).

🤖 What's new in AI

Somebody is giving away a frontier-class model and won't say who they are.

A model called Ox Alpha appeared on OpenRouter on Thursday. No lab name, no announcement, no paper.

A million tokens of context, text, images and video in, and free. OpenCode says it has capacity for a hundred trillion tokens a day, which is a number only the largest providers claim across their whole business.

Weirdly, this is the fifth of these stealth drops in about six months. A lab tests a model in the wild without its name on it, gets real usage data, then either claims it or quietly kills it.

All signs point to Zhipu, the Chinese lab behind the GLM models. 

But no one has come and claimed it.

What’s even crazier is that the benchmark numbers going around have it beating Claude and GPT-5.6 on coding (which are literally the frontier models).

There are some caveats with the claim, so take it with a grain of salt.

A really cool test I found that came off Reddit rather than a leaderboard.

Someone had a handful of models draw the same thing and put the results side by side. At a glance it doesn't look like much. A few of the others are arguably prettier.

But look at the physics.

Ox Alpha's version has a chain, and both feet are actually on the wheel. The others look more sophisticated as drawings and get the mechanics wrong.

That's usually a good tell. Getting the picture right is a rendering problem. Getting the physics right means something in there is modelling how the object works, not just what it looks like.

Anthropic breaks even

Anthropic reported an $11.5 billion quarter, up more than 14x year on year, with its first positive adjusted operating profit. Reportedly past OpenAI's quarterly revenue for the first time. 

Note revenue is contracted for 2 years into the future.

Nvidia is backing up to $105 billion of financing for a single OpenAI data centre in Ohio. More than most countries spend on infrastructure.

And Lovable raised $400 million at a $13.3 billion valuation, double what it was worth eight months ago, on reported ARR near $600 million. Insane.

I think most people would have thought that they would be wiped out, but by solving one single problem really well they’ve transformed their company.

ChatGPT can now read and answer your iMessages.

OpenAI shipped an Apple Messages plugin for the Mac app on Thursday. It reads and searches your threads and drafts or sends replies, across iMessage, SMS and RCS.

By default it won't send anything until you approve the message and the recipient. You can switch that off per conversation, which I'd think hard about.

It's on every plan, including free.

Two things to note. Apple silicon Macs only, so nothing on an Intel machine, and it wants Full Disk Access.

Note, Full Disk Access is the widest permission macOS hands out. To their credit the plugin runs locally and keeps no index of your messages, which is a better design than the alternative.

(You're still opening a very wide door though).

I know for a lot of founders out there this will be a game changer. I tested it out below and it’s so easy to set up (a little fkn terrifying too). 

Two big dogs in the world of AI going at it

David Sacks, who advises the White House on AI, accused Dario Amodei of regulatory capture. His line is that Anthropic wants a "DMV for AI", written so that only a company like Anthropic could comfortably comply.

Twitter tweet

He framed the split cleanly: "Dario Amodei believes frontier AI is too powerful to distribute; we believe it is too powerful to centralise."

Similar to what was written by Zuckerberg last week.

The Anthropic reply is that Silicon Valley treats every proposed rule as capture by default, which is convenient for anyone who'd rather have none.

Both can be true. 

A rule can be genuinely good for safety and genuinely good for whoever proposed it.

As I’ve documented before, I think all roads lead to open-source and it’s best if this power is distributed as much as possible.

Something happened in mathematics over the past month, that literally hasn’t happened ever

A number theorist at Anthropic, Levent Alpöge, killed a conjecture that had stood since 1939. The Jacobian conjecture, open 87 years in three dimensions and up. He worked with an internal Anthropic model, and the counterexample fits in about 216 characters.

An independent researcher, Dmitry Rybin, says GPT-5.6 Pro helped him break the Dinitz-Garg-Goemans conjecture, roughly 30 years old. The field hadn't formally confirmed it as of publishing, so hold that one loosely.

OpenAI says an unreleased model it calls Astra solved ten open problems in maths and theoretical computer science. Including a new bound on high-dimensional sphere packing that nobody had moved since 1978. Reported inference cost for all ten: 

About $2,000.

Hahahaha like what the actual fuck. 

$2k to solve some of the hardest theories in mathematics.

What people still picture when they hear "AI" is a text box you type into and it types back. What's actually happening is dozens of processes grinding at something overnight while you sleep. 

Proof that if you are still copy and pasting from ChatGPT, you’re barely scratching the surface.

Not everyone in the field is happy though. Terence Tao told the International Congress of Mathematicians in July that the discipline faces a "crisis in the foundations of mathematical values and practices."

Which is a careful way of asking what a proof is for, if a machine can produce one nobody understands.

Notably, nobody in that camp is saying the results are fake…

👷 Builder's notes

The email app. The triage system is running as a working prototype.

It reads the inbox, sorts and labels against a taxonomy, flags what needs me, and drafts replies. It never sends. 

The interesting part wasn't the drafting. Models have been fine at that for two years. It was getting it to run without me in the loop starting it, which is a different problem entirely and mostly not an AI one.

Now everything gets categorised and drafted before I even open up the laptop for the day.

Paul Graham's collected essays, on a Kindle (shoutout AI with Remy for the inspo)

There are a bit over 200 essays on paulgraham.com and no good way to read them away from a screen. So I scraped the lot, sorted them into six themed volumes, and built them into proper EPUB files.

Startups, essays on writing and thinking, the ones about work and life, and so on.

Took literally about 5 minutes.

If you want to do the same simply ask Claude Code:

Download [WRITER]'s essays from [URL] and turn them into valid EPUB files for my Kindle. Look at a few pages first to work out how they're built, strip the site furniture so the text reflows, check no chapter comes out empty, and tell me how to send them across.

If you want the Paul Graham books reply to this and I’ll send them over.

🤔 Differentiated thoughts: how to start a startup

Been reading a lot of Paul Graham recently (after I have been badgered by too many people I know that have told me I’d love his essays… they were right).

3 main ways to start a startup after consulting his readings.

1/ Take an assessment of the world around you and just go deep into where you could disrupt an industry / provide value. E.g. Legora saw law as being disrupted by AI despite not being lawyers and went to work.

2/ Work on what you want and go deep. Be curious about a particular subject and get as close to the frontier as possible. From there you can discover fractal buds of information that can proliferate into new businesses.

3/ Have a larger goal, put all your eggs into one basket and ultimately will this reality into existence, despite not having any prior experience in this field.

E.g. you may want to solve health care, education or energy despite not having any contrarian insight/industry insight as of present.

Elon for example had never been a rocket engineer but said fk it and willed that into existence.

Generally speaking the most common way is by exploring the things that ensnare the passions.

In my study of history it is much rarer to go and out-execute people with real domain knowledge purely on the premise that you are interested in achieving an outcome of say, wanting to make the education system better - despite never having been a teacher.

Generally if that was the goal, you would be passionate about it in the first-order sense and actually be deep into that topic to begin with.

For myself, I’ve recently been thinking about grander aspirations. Things that I think may be cool for the second-order desire or outcome they would achieve.

But I may not be in love with them from a first order perspective.

When you’re constantly involved with startups you tend to hear things like “mission” and “vision”. The higher the mission, the easier it is to attract better people, the easier it is to execute and so forth.

But not every startup has to solve world hunger (and most don’t).

Paul Graham talks of going deep into areas you’re passionate about. The hard thing about searching for startup ideas is that you can’t really go searching for them all that much. At least that can’t be the only reason you do it.

It’s like dating or sales. If you really really want to close the deal or date a cute girl - that desperation can bleed into your actions and constrain the capacity for that outcome to actualise.

What you want to do is to maximise the opportunity for these ideas to come to you.

By focussing on your passions then, you can enjoy the process of discovery and your day to day without being worried about the end outcome - of finding an awesome startup idea.

All the great companies for the most part came as a result of simply following the next thing that made sense.

Larry Ellison for example only ever wanted to work from his sail boat. He’s now the 6th richest person in the world.

Mark Zuckerberg built 9 other websites, applications, etc before building Facebook.

Brian Chesky and co (Airbnb founders) were mucking around for 2.5 years before they actually got traction. They were simply enamoured with the experience they had when hosting people at their bed and breakfast.

There are many more examples.

Now a part of this may be self-rationalising but I think generally it is true.

PG also notes that the amount of luck that goes into creating a great company is astonishing.

And that all you can do is optimise for interestingness.

So on that note, optimise for dad lore.

Carpe diem.

Alex

🎶 Carve outs

Steve Lacy, "Oh Yeah." On loop this week. and the Odyssey playlist 

Hilarious video: AI is getting so good at video.

Twitter tweet

Paul Graham, "SR-71". Short one, built around a photo of the Lockheed Skunk Works in 1962, see below.