You may never have heard of PHP, but you’ve certainly used websites built with it, because it powers seven in ten websites.
AI managed to screw up PHP?
Researchers at OpenAI uncovered two critical security flaws in PHP.
And you think, Hurray! Because finding them was supposed to be the first step in better defending the websites against attacks.
However, what you’ve forgotten is that this isn’t a turn-based game. Evil doesn’t wait for defenders to take theirs.
Cloudflare, a security SaaS business, recorded tens of thousands of exploitation attempts within 45 minutes of the announcement.
By the CEO’s update, millions.
I’m glad AI helped researchers establish the danger. I truly am.
But finding the problem is supposed to make people safer, not the other way round.
Elsewhere this week, Anthropic and DoorDash are following Muse into the personal assistant business.
I like Claude, but would I buy shares in the company? Musk is collecting rent from companies competing with Grok, while the SaaS businesses AI was supposed to kill keep growing.
We’ll also get to Character.AI’s bankruptcy and why AI companies are so obsessed with maths.
Shall we?
Claude Found Its Muse
Muse has given Anthropic something to worry about.
Think about who would try an assistant that connects to their apps and gets on with work while they’re away. Probably the same early adopters are already paying for Claude, experimenting with its tools, and looking for more things to hand over to AI.
Those are customers Anthropic already has. But having them doesn’t mean keeping 100% of their attention.
They don’t even need to cancel Claude. They can keep using it for coding while asking Muse to organize everything else in their life.
In this case, who do you think learns more about users’ routines, gets more context, and becomes the assistant they reach for first? That’s right, the users who interact with tasks that are always receiving new context.
If I were on Anthropic’s product team, that would worry me more than another model beating Claude on a benchmark.
You can improve a model and make it the default in the software industry. Which Claude already is, however, only if Anthropic’s ambitions stop at being the best coding assistant. No, because eventually they want to be the best at doing everything for you.
Fortunately for Anthropic, Meta has already done the hard work for them.
Meta has put a product in front of people and proven it solves a pain point.
Anthropic can study how Muse gets used, when people interact with it, and what keeps bringing people back to the chatbot, free of cost. Of course, they paid by not being the first one to come up with this viral product. But they could easily take the Apple route, just polish something and make it exceptional.
So, meet Claude Duet.
Anthropic’s new proactive assistant checks connected apps, brings things to your attention, and keeps working while you’re away. Simply presenting Duet to the existing users, it carries over everything you interacted with Claude.
You can see what Anthropic is hoping for.
This is no doubt a pie to grab.
But even if they have a successful Duet launch that’s as grand as Muse, they still don’t have enough income to justify everything Anthropic is spending.
So, the next piece of news, or my question to you, if Anthropic is getting ready to sell shares to the public, would you buy them?
An Expensive Time
I like Claude. But liking the product and wanting to own the company are two different things.
Anthropic has reportedly committed nearly $450 billion to cloud computing agreements. Spread that over ten years, just to understand the scale, and it’s roughly $45 billion a year. The payments won’t fall evenly, but that gives us a starting point.
Before buying in, I’d want to know how it expects to earn that money. Especially when customers can switch to cheaper models or better-suited products.
To be fair, customers are spending more. Anthropic told investors its third-quarter revenue reached $16.8 billion, up 45% from the previous quarter. However, they also have an $80 billion net loss in the first half.
No wonder some investors are already discussing a $0.5 trillion reduction in valuation. How much is $500 billion? To put that in context, it’s roughly six months of spending by every public school system in America combined.
OpenAI is competing for that investment money, too.
It’s looking to raise about $30 billion, and getting their commitments first could leave less money available for Anthropic.
OpenAI might have just realized that it needs a bridge round before a reasonable IPO. Especially the higher interest rate has already dragged a few down: Oura postponed its IPO, despite profitability, or Holtec Nuclear (a nuclear tech) postponed its filing, being uncertain about the data-center need.
Not to mention, OpenAI isn’t doing as well as they hoped, earning $20 billion less than forecast (though arguably the numbers were calculated differently).
But perhaps this slowdown is exactly what Amodei and Altman not so secretly hoped for.
Dinner, With Motives
DoorDash would also like a piece of the personal assistant pie!
Its new feature, enabled text-ordering beta, lets you request food and confirm an order all inside a message thread.
The scenario is that you order the same meal every Friday. Though I’m not convinced, I’m not sure how much effort this saves. Reordering in the app was hardly an expedition.
That said, groceries are a more convincing use. You have an idea for dinner, but still need to work out what to buy. An assistant can help assemble the basket.
DoorDash agrees: it says grocery baskets built through its assistant average nearly 50% more value than the same customers’ traditional orders. So a basket that would otherwise have been $100 is, on average, closer to $150.
You can see why DoorDash would be interested.
However, I also wonder how much of this is preparation for assistants such as Muse becoming the place people start shopping.
Imagine asking your assistant to find dinner. Would you want it to check only DoorDash? Or compare the final bill with Uber Eats?
DoorDash’s assistant already looks for cheaper equivalent products within a store. However, it still doesn’t compare to other store options outside DoorDash’s platform.
I can see why DoorDash wants an assistant.
That said, as a consumer, I’m less sure why I would want a separate assistant for every company trying to sell me something.
It’s now proven! Models ≠ moat
Musk seems quite comfortable with a different arrangement.
Grok Bot can use outside models as well as xAI’s own. It chooses a model for the task, and users don’t necessarily see which one handled the request.
So you can pay for Grok Bot and have a competitor’s model doing some of the work.
Fine by me if the work gets done properly.
Yet it does make the race to own the smartest model look rather different.
If access to a capable model is getting cheaper every day, why would anyone spend enormous sums reproducing it? As I explained in the Microsoft earnings, apparently, Musk has just reached the same conclusion.
Musk doesn’t have to beat Claude to make money from an AI model. As I said in the SpaceX analysis, Anthropic pays SpaceX $1.25 billion a month for computing capacity. While Grok competes for customers, SpaceX can also collect money from its competitors.
That gives Musk more ways to benefit from AI. He can sell the service using somebody else’s model and sell computing capacity to companies building models.
Apple has its own version of this opportunity, ie, the Mac mini for an at-home or startup AI setup, w/o the need to develop its own version of ChatGPT.
And if the next model is better, these companies can simply use that one.
That’s a comfortable position compared with paying to build the model everyone expects you to beat next month. No point doing the hard and expensive work yourself if there’s a better and proven route to revenue.
SaaS, Still Here
What about the supposedly imminent death of SaaS?
It’s worth checking what customers have actually been doing.
In its latest reported quarter, Salesforce’s subscription and support revenue grew 12%. Some of that came from its Informatica acquisition, so we shouldn’t credit AI with the whole increase. But customers have plainly not abandoned the business.
ServiceNow’s subscription revenue grew 24.5% in its second quarter. That’s roughly $125 coming in for every $100 a year earlier.
Hang on, what about Adobe? It seems like somewhere we should see the damage. AI can generate images, edit photographs and make videos. Yet its latest results show creative and marketing subscription revenue growing 13%.
The list goes on: Atlassian’s cloud revenue grew 31%. CrowdStrike’s subscriptions grew 27%. Workday’s grew almost 14%. Yesterday, SAP reported 29% growth in its Cloud ERP Suite.
Design tools, collaboration, security, HR, and finance. Customers are still spending across quite a few supposedly endangered categories.
Why? Well, being able to build an app doesn’t mean a company suddenly wants to replace the software running its payroll.
Someone still has to move the records, recreate the approval rules, connect everything else and take responsibility when somebody gets paid incorrectly. The demo where you build an app over lunch usually ends before that part.
I keep returning to this because the prediction skipped over rather a lot of work.
Getting AI to generate a customer dashboard is one thing. Replacing the system holding your company’s customer records, permissions, approvals, and years of slightly embarrassing workarounds is another.
You may hate the software. Your company may still depend on it.
What most people forget is that the incumbents adopt AI too! Suprised!
They already have customers to sell the new AI capability to, as opposed to a startup that has to persuade those customers to move.
RIP, Character.AI
I love how history teaches us lessons that most people ignore but eventually repeat.
Heavy borrowing and overexpansion helped produce bankruptcies, when the tracks remained useful.
Now Character.AI is in Chapter 11 seeking a buyer.
Don’t remember character.ai? The troublesome AI app whose chatbots families say drove their children to self-harm and suicide.
This isn’t the first, and won’t be the last.
Gartner reports that more than 80% of specialist AI cloud providers are forecast to cease operating independently by 2030. So we can expect a few more of these bankruptcies and acquisitions.
About Those Maths Problems
Have you ever wondered why these AI model companies are always going into math competitions?
For those of us who don’t spend our afternoons trying to prove a theorem, what should we make of it?
OpenAI has published hundreds of mathematical research results produced by an unreleased model.
So it turns out Maths gives these companies something useful: a way to check whether an impressive-looking answer actually holds up. You can sound very convincing in an essay and still be talking rubbish.
A mathematical proof has to survive someone checking the steps.
That’s why I’m interested in what OpenAI has published. How much of this is AI getting better at passing a test, and how much is it doing work that mathematicians couldn’t do before?
And of course, I don’t need AI to collect another maths medal. Same as you, I hope this leads to an actually logical AI rather than whatever we currently have.
One small footnote: OpenAI has announced invisible watermarks for eligible ChatGPT and Codex text in the EU. They work through patterns in word choices, rather than hidden characters. Though the watermark doesn’t indicate how much a human contributed, someone could have come up with the whole idea in Japanese, and having it translated into English could still be marked as GPT-generated!