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!

Your next campaign needs a dozen fresh ad variations by Friday. Your agency quotes two weeks and a five-figure invoice. Your in-house designers are already buried under this quarter's requests.

Hightouch Ad Studio fixes that. It reads your brand guidelines, your best-performing creative, and your product catalog, then generates on-brand ads your team can ship the same afternoon. You review and approve every asset before it goes live, so quality holds.

Growth teams use it to build variations for every audience, test more of them, and stop rationing creative because production got expensive. The work that once needed a full agency retainer now runs inside your own workflow, at your pace and under your direction.

You direct the work while Ad Studio handles production, and your designers get their week back.

Start your first campaign today

1. China's breakthrough AI model threatens America's lead

We start this week with Kimi K3, a new Chinese AI model that has threatened America's lead in the AI race.

Moonshot AI is the Chinese company behind it and their team has matched the performance of America's most advanced models. But unlike those systems, Kimi K3 is open-source. This means anyone can use it for free and change how the model works.

It's a huge moment for open-source AI, as independent benchmarks show that it performs just below Anthropic's Fable 5 and ahead of OpenAI's flagship model. That's a remarkable achievement when you look at the finances, as Moonshot raised $2 billion back in May - while Anthropic raised $65 billion in the same month.

But it's even more impressive when you consider that US export controls have blocked China from buying the most advanced chips, which are needed to train these models. This has seriously restricted China's tech industry and forced them to focus on efficiency instead.

The approach is clearly paying off, with Moonshot's president admitting that they "didn't have the luxury to just scale up compute". That efficiency has also rattled investors, as it weakens the case for buying huge numbers of expensive chips.

As a result, chip stocks have fallen more than 20% from their June high. And this isn't a one-off, as the UK's AI Security Institute has found that open models now trail the frontier by just 4-7 months - a gap that has narrowed over the last year.

If Chinese labs can deliver this level of performance at a fraction of the cost, the enormous spending from American labs becomes much harder to justify - along with the AI boom that so much of the US economy now depends on.

2. Amazon accidentally bills its customers for $1.5 trillion

If you checked your AWS bill last week, you might have been shocked to see that you owe Amazon billions of dollars - or even $1.5 trillion in one case.

One UK charity, which normally pays less than $1 a month, was shown a bill for $7.8 billion - while a historian in Derbyshire was told she owed $245 billion.

The company has since apologised to its customers, after a glitch caused the unit prices to be calculated incorrectly and directly impacted monthly bills.

Thankfully, these were only estimates on the billing dashboard, so no money was actually taken. Mistakes on this scale are incredibly rare and Amazon corrected it within a few hours.

But Amazon isn't alone here, as glitches have been plaguing other platforms - with Spotify users complaining about broken uploads and being unable to access their podcasts.

While we don't know the exact cause of these problems yet, software engineering teams have completely changed in the last few months and they're now using AI to write code faster than ever.

If your company is doing the same and embracing AI coding tools, it's a good reminder that the basics still matter - reviewing changes properly, testing them, and rolling them out gradually.

3. Demis Hassabis calls for a global AI watchdog

Demis Hassabis, who's the CEO of Google DeepMind, has called for a global AI watchdog that would be led by the US.

In a blog post, he said the organisation should be able to test frontier models before they're released, and coordinate an industry-wide slowdown if a model is judged to be too dangerous.

Hassabis believes that we're "only a few short years away" from developing artificial general intelligence, which is a system that could match or outperform humans at almost any mental task.

That would be a huge jump from today's technology, which is very good at specific tasks like writing and coding - but still needs a human guiding it.

Reportedly, Hassabis has spent months trying to build support behind this idea - with talks involving the White House, EU, and other AI labs.

Personally, I can't see this panning out as he expects. European leaders are quite nervous about handing more power to the United States, as access to American technology was weaponised as a bargaining chip during last year's trade disputes.

I also can't see China agreeing to this. President Xi has spent the last week pitching the complete opposite vision, telling an AI conference in Shanghai that AI development "should not be a solo performance by a single country".

4. Grok was secretly uploading its users' entire codebases

SpaceXAI has been caught uploading its users' entire codebases to cloud storage, through its Grok Build coding tool.

Security researchers found that Grok was sending full repositories to Google Cloud - including files that it was told to ignore and secrets that had been deleted.

That's a lot more data than most coding tools retain, and the uploads were only switched off after the findings were published.

One researcher warned that the exposed data could include proprietary code, credentials, and details of security vulnerabilities.

Elon Musk claims that the uploaded data will be "completely and utterly deleted", but there's no way for users to verify that has been done.

It's a worrying revelation for companies, as they start to embrace these tools and try to make their teams more productive. Interestingly, Microsoft's CEO has warned about the risks of handing over this knowledge to tech companies - which is good timing given these concerns about Grok. I've covered it in the story below.

5. Microsoft warns you're teaching AI companies your secrets

Sticking with the theme of data, Microsoft's CEO Satya Nadella has warned companies that they're "paying twice" when they use AI models.

His argument is that businesses pay for their usage, but they also hand over incredibly valuable knowledge - which goes far beyond codebases. It covers everything from the prompts they write, to the corrections they make when a model gets things wrong.

Over time, that teaches the models how a business actually works, which is the kind of knowledge that a competitor could never buy. He also finds it ironic that the labs train freely on public data, but won't allow others to use their models in the same way.

Overall, it's a pretty surprising message from Microsoft, given that they've invested billions in both OpenAI and Anthropic.

Nadella's solution is for companies to keep ownership of their data and use tools that make it easy to switch between models - which, conveniently, Microsoft's cloud team can sell them.

Despite that clear self-interest, it seems like the shift is already underway. Just look at Vercel's AI gateway, where open models made up 29% of all traffic last month.

Thinking Machines allows businesses to customise their AI models

Thinking Machines, the AI startup that was founded by OpenAI's former CTO Mira Murati, has released its first model - and it takes a very different approach to the endless stream of chatbots we've seen this year.

Their Inkling model is free to download and can run on a business' own servers - ensuring that sensitive data never leaves the company. Thinking Machines' Tinker platform also allows teams to fine-tune Inkling on their own data, which shapes the model around how their business actually works.

Think of an insurer that wants claims handled in a specific way, or a bank that needs decades of lending policies followed - that's the kind of detail that general-purpose models can't achieve.

Essentially, the company is betting that custom models will outperform the general-purpose ones that are sold by the big labs. And there's some early evidence for this, as a similar project with the hedge fund Bridgewater beat the top models on financial reasoning - while costing a fourteenth as much to run.

The business model is different too. Instead of charging businesses for every request, like the big labs do, Thinking Machines makes its money when companies pay to customise the model through Tinker.

The catch is that fine-tuning still requires machine learning talent, and most companies don't have that in-house. If Tinker can make the process simple enough, I can see plenty of businesses giving this a try - but until then, Inkling is really aimed at companies with deep pockets and their own AI teams.

💰 European defence AI company Helsing raises $1.8 billion

🎓 Anthropic launches a version of Claude for school teachers

🏗️ New York becomes the first US state to pause new data centres

👨‍💼 Anthropic's hiring spree continues with Monzo cofounder Tom Blomfield

⌨️ OpenAI releases a $230 Codex keyboard

💳 Stripe and Advent offer to buy PayPal for $53.4 billion

⚠️ OpenAI's new flagship model keeps deleting users' files

📊 Meta could soon cap AI token budgets for individual engineers

This startup is based in Singapore and has raised $439 million for its video generation tools, which has pushed its valuation past $2 billion.

PixVerse offers a range of models that can generate videos in 4K resolution, with the audio included. These cover everything from casual clips that people make for fun, to professional film and advertising work.

The company was founded in 2023 by Wang Changhu and Jaden Xie. Changhu previously built the visual understanding technology behind TikTok's recommendations, and PixVerse believes this experience with labelling data is its real edge over rivals.

That approach seems to be working, as its product has grown to over 150 million registered users - with 15 million of them active each month.

Their investors include Alibaba, which has just signed a deal to use PixVerse's video generation tools across its own products.

The company has also released its own world model, which can create environments for game development. Regular readers will know I've been following world models closely, as they could revolutionise both gaming and robotics.

For example, robots could be trained inside thousands of realistic virtual worlds - which is far faster and cheaper than gathering that experience in the real world.

If you want to learn more about PixVerse, I've included a link below.

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.