Welcome to this edition of Loop!

To kick off your week, I’ve rounded-up the most important technology and AI updates that you should know about.

‏‏‎ ‎ HIGHLIGHTS ‏‏‎ ‎

Let's jump in!

Adam spent 20 minutes looking for a $36 receipt. His finance team sent three Slack messages. Someone made a sticky note.

Ramp would have matched it automatically the moment he swiped. Auto-coded, in-policy, synced. Nobody had to ask Adam for anything.

This is what finance looks like when it runs itself.

Your team can be Adam. Or they can not be Adam.

Don’t be like Adam

1. Rogue AI models have attacked even more companies

We start this week with more rogue AI models, as last week's story has become even more serious.

Just a week ago, OpenAI revealed that two of its models had escaped their testing sandbox and hacked Hugging Face by accident. Worryingly, the company has found even more cases of agents escaping containment - although none are thought to have left its network.

But OpenAI isn't alone in dealing with this problem, as Anthropic's models have attacked three different companies and breached their systems. These attacks date back to April, and they were only found when the company reviewed its own systems.

Unlike OpenAI's models, Anthropic's didn't break out of their sandbox. Instead, the team running the tests accidentally left it connected to the internet, which gave the models a path into other companies' infrastructure.

There was nothing sophisticated about the attacks either, as they got in using basic techniques like weak passwords and malware. But the problem is that nobody was watching while this happened, so these cybersecurity attacks went unnoticed for months.

As expected, the calls for government oversight have become even louder, as the White House is now "looking at controls" and the European Commission has met with both AI labs.

All of this ties into the Five Eyes warning that I covered back in June, where the intelligence agencies urged businesses to take action and prepare for AI-driven attacks. Unfortunately, these types of attacks are only going to become more common, and small and medium-sized businesses are the ones most at risk.

2. Oracle could become the face of the AI bubble

Oracle has become the company to watch if you're worried about an AI bubble, as a detailed report from The New York Times shows just how much Larry Ellison has gambled on the build-out.

The company is building a huge network of data centres for OpenAI, worth $300 billion over five years - but those payments don't actually start until 2027.

Oracle has borrowed $130 billion to fund the construction, and that strain is already starting to show. Oracle's credit rating is just one notch above junk status, its share price has fallen 60% from last year's peak, and it has been forced to lay off 18% of its staff.

But Oracle isn't the only one spending big, as Amazon, Google, Meta and Microsoft are on track to spend a whopping $1.5 trillion on data centres over the next two years.

This level of record spending is starting to make investors nervous, especially since China's AI companies are catching up. In recent weeks, these labs have released models that are nearly as powerful as the best American ones - and they're free to download.

That means that businesses can easily run these models on their own hardware, avoiding the thousands of dollars that they'd normally pay the likes of OpenAI and Anthropic for access.

While no one doubts that the technology is here to stay, the big question is whether the profits will arrive in time to cover all this borrowing - and Oracle seems to be the most exposed if they don't.

3. Google's new AI model can control a robot's entire body

Google DeepMind has released Gemini Robotics 2, an advanced AI model that can control a humanoid robot's entire body.

The previous version could only manage a robot's upper half, but this one can handle everything from the robot's feet to its fingertips - allowing it to walk, crouch, and pick specific items off shelves.

The dexterity has improved too, as the model can now control five-fingered hands well enough to tie a bin bag or unscrew a lightbulb.

But what's really interesting is that the model isn't tied to one machine. Instead, it can adapt and control robots that have completely different shapes and sensors. It even allows different types of robots to work together on the same task.

Of course, these demo videos are always chosen carefully, and Google itself admits that they need to improve the robot's speed.

But there's a broader trend here, as one AI model could power thousands of different robots, in the same way that a single LLM powers thousands of apps today.

The top AI labs and tech companies have been heavily researching this area for several years, but Google's team have been making huge strides recently.

4. Apple prepares for a memory shortage that could last until 2028

Apple is famous for running one of the leanest supply chains in tech, but it's now stockpiling components as it braces for the memory chip shortage to get worse.

The company has just reported its strongest June quarter ever, with iPhone sales up 22% and Mac sales up 29%. But Tim Cook has warned that supply problems are about to increase significantly, describing the surge in memory prices as "a hundred-year flood".

This is all being driven by the AI boom, as data centres need enormous amounts of memory to train and run AI models. Chip makers are now committing most of their production to these customers through huge multi-year contracts, which leaves less supply for phones and laptops.

Unfortunately for consumers, there's no end in sight as Samsung expects the shortage to worsen through 2027 and last until 2028. That demand has been great news for its chip business, pushing quarterly profits up nearly 1,300% to a record $49.6 billion.

But even Samsung isn't immune to the squeeze, as rising component costs led to its own phone business losing money this quarter. These costs are now hitting ordinary consumers, with Apple, Meta, Microsoft and Sony all having to raise prices.

If you've been putting off upgrading your laptop or phone, I'd do it sooner rather than later - prices are only heading in one direction for the next couple of years.

5. America bans foreign robots as the tech war with China expands

The tech war between the US and China has opened up a new front, as America banned the imports of foreign-made robots. These new rules technically apply to every country, but Chinese firms will be hit the hardest as they manufacture most of the world's robots.

The restrictions cover any robot that can move around, sense its surroundings, and connect to the internet - which means that new robot vacuums could be impacted.

The FCC claims these devices pose "unacceptable risks" to national security, although existing products aren't affected and companies can apply for a waiver.

This seems more focused on boosting American manufacturing than improving security, as the waiver doesn't ask companies to make their devices any more secure - it just wants a detailed plan for moving production to the US.

The timing is no accident either, as the race to build humanoid robots has exploded over the last 12 months. China's Unitree has gone viral with robots doing everything from dancing to MMA fights, while Tesla has stopped making two of its cars to free up factory space for its Optimus robot.

Both governments expect humanoids to become one of the world's most important industries, and neither side wants to rely on the other for them.

While the industry is still in its early days and there are far bigger advances to come, robots are already being treated like semiconductors and a technology that's critical for national security.

Microsoft releases its first AI model for cybersecurity

Defending against AI attacks is quickly becoming a big business, and Microsoft wants to get in on the action. The company has just released MAI-Cyber-1-Flash, which is its first AI model built specifically for cybersecurity.

This new model was trained partly on decades of data that Microsoft has collected, as it reacted to attacks on its customers. With Windows, Outlook and Azure being some of the most targeted products in the world, the company has built up a security dataset that few can rival.

They've also announced Project Perception, which turns AI models into teams of agents that automatically scan for weaknesses in a company's network and repair them. Some of these agents will even try to imitate hackers.

Microsoft claims the model beats OpenAI and Anthropic on a standard security benchmark at around half the price. Although I'd treat that with some caution until we see more, as the model wasn't shared with independent testers before release.

The timing is important, as advanced AI models have become alarmingly good at finding gaps in cybersecurity. Anthropic's Mythos found thousands of security holes earlier this year, which had gone undetected in popular software for years.

As we covered earlier, Chinese labs have been releasing models that are incredibly powerful and free for anyone to download. Since they can be modified so easily, there's nothing stopping attackers from removing their safety controls.

So these hacking tools are coming whether the industry likes it or not, and the real race is getting the defences deployed before they arrive.

💸 OpenAI cuts its AI model prices by 80% as companies worry about costs

📉 Meta's profit falls 14% as its AI spending continues

📈 Microsoft boosts AI spending as its profit jumps 31%

📱 OpenAI is building a family of devices for its AI chatbots

🚁 DoorDash is building its own drone delivery business

🌍 OpenAI's models now reach more than 1 billion weekly users

💾 Chinese RAM maker CXMT makes a $484 billion stock market debut

🛡️ Nvidia and Microsoft launch an AI security alliance without OpenAI, Google, or Anthropic

🛰️ Amazon's new satellite network for mobile phones puts pressure on SpaceX

🤗 Hugging Face's CEO calls for transparency after the OpenAI cyberattack

This British startup has just raised $30 million, which is one of Europe's biggest defence technology seed rounds this year. They plan to build virtual battlefields and change the way that we develop AI-powered weapons.

Today, these weapons are trained in the real world, which is slow, expensive, and really difficult to do at scale. But with Agon's synthetic battlefields, they will be able to practice against simulated threats instead.

This means that a counterdrone weapon can rehearse downing drone swarms thousands of times, long before it's ever tested in combat.

The timing makes sense, as the war in Ukraine has showed how quickly drone warfare is evolving. European governments are also spending heavily on their own defence technology, as they hope to reduce their reliance on the US.

The big question is whether this training actually carries over, as a weapon that performs perfectly in a simulation can still struggle in the chaos of the real world.

While AI-powered weapons aren't something that anyone wants, Ukraine has shown just how important it is for countries to be prepared to defend themselves.

So it's great to see a European company receive funding for this, which has historically been quite difficult on the continent.

This Week’s Art

We’ve covered quite a bit this week, including:

If you found something interesting in this week’s edition, please feel free to share this newsletter with your colleagues.

Or if you’re interested in chatting with me about the above, simply reply to this email and I’ll get back to you.

Have a good week!

Liam

Share with Others

If you found something interesting in this week’s edition, feel free to share this newsletter with your colleagues.

Liam McCormick is a Senior AI Engineer and works within the AI team at Bright. He identifies business value in emerging technologies, implements them, and then shares these insights with others.