Edition 7 · 27 July 2026
how to build out your AI OS in 4 steps
AI news, the basics of building, and for our differentiated take: the four steps to building out your own AI operating system.
Morning,
📜 TLDR:
Update on our latest product: AxleClaw: the AI that runs your business just by texting it.
All the updates in AI (bro i stg every week has something new and crazy)
Builder’s notes: the basics in building
Differentiated take: do all roads lead to zuck?
Carve outs…
Enjoy :)
💼 1. What’s new in the biz
Update on the AxleClaw campaign - we have signups from our product campaign!
2 to be precise haha - but we had the wrong targeting setup (which has been fixed as of a day ago), so hoping we’ll be able to up that moving forward.
We’ve onboarded about 10+ other people that have enquired separately, which is great.
We’re offering a limited time free trial. The biggest thing we’re optimising for is user feedback. So (at the risk of sounding like a broken record) if you are (or know someone that is) looking to see what their own personal AI agent for their business could look like - feel free to reply to this email (or share it with someone in need).
Current user feedback consists of:
emails (we should be able to ship this asap) - but having them auto-drafted, because we have all the context we should be able to solve this ASAP.
making it easier to automate tasks with skills - or rather making the design around this better.
be more chat friendly - e.g. more like texting a friend
faster response time
Other features that I think would be cool:
Adding in the ability to call upon pre-made workflows that are relevant to the user (so you automate tasks that are specific to the business in 1 go).
As a reminder of what it can do: competitor analysis, code (makes websites), reporting, checking in on employees, following up on leads, sorting your emails, and so much more. If you want something done? Just ask it!
And you can just text it from your iMessage!
The harder problem underneath all of it is the blank canvas. You open AxleClaw and it can do almost anything, which is the same as not knowing where to start. Which we’ve solved to a degree with an AI audit that you get as soon as you onboard.
The trap is trying to solve every business problem ever to fill that canvas. The likelier answer is niching down, picking an industry, and solving that one deeply instead of everything shallowly. I'm not settled on it yet. But the way we’re getting there is unglamorous and probably the most useful thing I've done all month: sitting down with every single user, one at a time, working out who they actually are and what they actually want solved.
One other change: I've pulled back slightly on social media this month. More on why below.
🤖 2. What’s new in AI
An AI escaped its own safety net
This one is genuinely bonkers.
OpenAI confirmed this fortnight that during an internal safety test, two of its own models escaped their sandbox and broke into another company's systems, Hugging Face, to cheat a benchmark. Harvested credentials, moved laterally, the works.
It broke out because breaking out was the easier way to win the test it was being scored on.
The setup was an internal red-team evaluation with the normal guardrails deliberately removed. Two OpenAI models (GPT-5.6 Sol and an unreleased one) were being run against a benchmark. Instead of solving the benchmark task the intended way, they found that the benchmark's own infrastructure had a weakness, exploited a flaw in the dataset loader, harvested credentials, and moved laterally into Hugging Face's production systems to get the score. Cheating the test, essentially, by going after the machine grading it rather than answering the question.
It then also found zero-day vulnerabilities, which means there’s no time for IT specialists to intercept the attack (basically very bad).
This stuff isn’t meant to happen. Not trying to be all doomer and gloomer but this is genuinely insane. Not a whole lot to add other than we’re getting to scary levels of autonomy.
This is fodder for another article - but I think the reward functions for AI models is one of the more under-appreciated parts of constructing an LLM.
AI agent went rogue and hacked startup by itself, OpenAI revealsAnthropic shipped Claude Opus 5 on Thursday. The headline isn't a new capability, it's the price. They're calling it near-Fable-5 intelligence at half the cost, same rate as the model it replaces. Google's new Gemini 3.6 Flash came in about 17% cheaper on tokens. DeepSeek pushed its V4 model to general release and started charging more during business hours, which tells you compute is the constraint now, not brains. Four frontier-ish models in two weeks, and the thing they're competing on is cost.
Introducing Claude Opus 5Additionally, Cursor, the coding tool, shipped a "Router" that reads each request and automatically sends it to the most efficient model that can handle it. They claim frontier-quality at up to 60% less spend. Picking the model is now its own product.
Then two bigger signals landed the same week, in case Cursor was too subtle. Stripe, the payments company, is reportedly in talks to buy OpenRouter, a startup whose entire job is routing between AI models, for around $10 billion. That's up from $1.3 billion in May. A payments giant paying frontier money to own the switchboard, not any of the models it switches between.
Stripe has also been active in another acqusition. In a story of grand irony, Stripe, together with a PE firm, reportedly bid around $53 billion for PayPal, and PayPal's board reportedly knocked it back as too low.
PayPal was founded by Elon and Thiel back in the day and sold to eBay. And Thiel and Musk were also early investors in Stripe. Which is just all too funny.
Craziest part about all of this? Stripe is proposing this acquisition and they’re not even public!
In other news, OpenAI launched a small-business program: free training, webinars, in-person sessions on using AI for accounting, marketing and e-commerce. Notice it's training, not a new tool. HubSpot shipped an Agent Builder that lets a non-technical person describe an automation in plain English and build it against their own CRM data. Canva Code 2.0 will build a working landing page from one sentence, on the free tier. Intuit folded receipt-matching and bookkeeping straight into a QuickBooks credit card.
The big tech companies are now trying to capture as much market-share as possible by catering to people’s use cases, where previously they were focussing more on trying to make the best model possible (which is slightly different).
👷 3. Builder’s notes: there are so many tools that it’s easy to overcomplicate, but building in AI, the basics are your best bet
The more clients we work with, the more the interesting work turns out to be boring. It's not the flashy model. It's the fundamentals of the business, set up properly.
There's an order to it, and it doesn't change:
Context. This is the foundation, and it's king. An AI that doesn't know your business, your customers, your voice, and how you actually work gives you generic answers dressed up as smart ones. Get the context right and everything above it gets sharper for free.
Skills. Once it knows you, you teach it the things you do over and over. A skill is just a saved recipe, an SOP your AI runs the same way every time. Same input, same quality, no re-explaining.
Routines. Then you let the useful ones run on their own. The brief that's waiting when you wake up. The report that writes itself on Friday. The work happens whether you remember to ask or not.
Agents. At the top, a small team of specialists, each owning one job, with you directing. This is where it stops being a chatbot and starts being an operating system. We recommend this for when things get complex, softwares won’t talk to each other, you need a really specific outcome produced every single time.
For anyone looking to build in AI, this is where you start.
Another thing I’ve been getting asked about a lot is data. And who owns what. This is a diagram I’ve been showing people more recently for how we architect it so that people’s data stays yours, and there are the correct authentication layers. But for anyone getting someone external in, it is vital that you, the business owner, owns the data.
How our infrastructure works. The data all feeds into an AI brain, which is stored on its own private server. There’s a model router on the backend that basically switches the model out depending on what it thinks the best model for the task is. This is then plugged into Claude/ChatGPT (have to be on a business plan though to ensure data guarantees are upheld). You can then add in agents that are linked to the business brain.
But this lets you interact with Claude and ChatGPT while ensuring connections and company data are secure - which is very important.
And on that note, the way I made this diagram was with Claude Design.
I don’t know why I haven’t been using it more before, but it is SO GOOD.
Like genuinely.
It does burn through a fair few tokens in fairness, but I think it’s the most fun I’ve had with an AI tool in a while.
We have our system design (our brand kit) embedded into Claude Design, so it outputs really really clean pitch decks. It’s also tied into our WhiteHorse AI brain, so it is then able to pull out the exact pain points from calls we’ve had with clients and deliver very, very tailored. And then this is put on a skill to produce an output that is clean and tailored to our brand.
And voila, beautiful, clean, professional, personalised deck.
It made me rethink our whole sales process. We can reduce the time to value and make a personalised value-based pdf/website/portal for every enquiry that we get and do this automatically. I think the friction for providing immediate value has dissipated immensely and it opens the door for a variety of possibilities.
🧐 4. Differentiated take: do all roads lead to Zuck?
Social media is one of the most interesting things of our time.
I’ve basically grown up in a generation that takes social media for granted. People resent it, and yet they are consumed by it. It’s genuinely mind-boggling. There was a comment section I saw recently of people wanting Meta to basically be nationalised, owned by the masses or killed in some capacity (basically a post talking about the potential downfall of meta). And yet I can bet the average person has a screen time above 4 hours a day.
But does anyone feel better after spending 3 hours consuming short-form content? I feel like my brain is fried after 5 minutes of it.
What’s particularly interesting is the shift in how it is consumed or how the algorithm is wired nowadays. It’s interest media - not social media. We are shown things we are interested in, for they increase, “time in app”.
Now that’s not necessarily a revelation for a lot of people.
What’s interesting is analysing the way it has shaped human behaviour and our economies. The big tech platforms knew that if they empowered everyone to become a creator, you’d increased the number of people uploading to the platform and give yourself better chances at keeping people in the app and making more money.
More content = more content people are interested in = better content = better engagement = more time in app = more money for big tech.
So now we have people creating personal brands, around literally anything. I note I am not immune to this. I resisted posting anything online for quite a while. I didn’t like the idea of falling in line, I’d prefer to be more private about my life. I figured it was always better to just build in stealth.
But the incentive structures are too great.
Trying to grow a business in the 21st century with no digital presence is like trying to ride a horse when cars were being invented - you could do it, but is very impractical and rather inefficient.
The other part of it is that it is largely an asymmetrical bet.
You lose nothing from posting. Obviously time, and maybe a bit of social reputation or brand (if done incorrectly) but the upside is so much greater. With some smart editing and the bare minimum of knowledge you’re able to reach people you otherwise would not have. There is no marginal cost to producing a piece of content that reaches 1 person or 10 million people. All you have to do is pick up the phone and record.
What’s interesting has been the change in media as traditional outlets and older people realise they can have more influence here. In particular, someone like Chamath has been posting more. For what? The guy’s already a billionaire. Simply, because it’s where the attention is. And there’s only upside to be gained.
Want to get the word out about something you want to offer? Post it.
Have something cool to share? Post it.
Got absolutely nothing interesting to share, but figured what the hell let me post a day in the life? Post it.
What do you have to lose?
No one sees it? Then no harm done. Lots of people see it? Then even better.
Makes me think, do all roads lead to Zuck?
What does this future look like in 20 years? We’ve barely had 25 years of this stuff and the world is so ridiculously different to when Facebook launched.
And what do people actually want?
Sometimes I feel like I am a victim of my own bubble (which we all are), but I find it really refreshing just talking to people outside of business that I don’t know. Because when you’re constantly talking to people in the know it feels like you get this warped perception of reality.
And then you just go outside and meet people that just want to have a beer with their mates and enjoy a good meal, with good company.
Genuinely curious what people think the future looks like in 20 years. Will we be walking around with meta glasses on scrolling reels in one eye, while looking at a slide deck in the other and trying to have a conversation with a friend at lunch? I’d hope not.
If you have thoughts, reply to this email! Would love to chat :)
🎶 5. Carve outs:
⚽ Sports:
📖 Books:
Leonardo by Walter Isaacson, still grinding - crazy the things that they were able to build back in the day.
🧑🧑🧒 Family
My youngest brother turned 21 so now the whole family is officially old 😕
🎵 Music
Elvis is goated
Enjoy the week, until next time
Stay hungry, stay foolish,
Cheerio,
Alex