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!

1. CEOs replace themselves with AI deepfakes at earnings calls

While executives often talk about how AI can be used to replace general workers, it now seems that the technology could also impact the C-suite too.

Eric Yuan, who is Zoom’s CEO, has created his own AI avatar and then asked it to present a section at their quarterly earnings call. It’s also worth noting that Klarna’s CEO has done almost the exact same thing.

Yuan's “digital twin” was created by combining several Zoom clips, which the AI model tries to mimic later on. Although, there are genuine worries about how the technology could be used - especially in the wrong hands.

What’s to stop a bad actor from also creating an AI avatar of Zoom’s CEO, then using it to order that large sums are paid to a fake company?

This is a risk that every business will have to deal with and, with the rapid improvements we see year-to-year, that might be sooner than you think. Businesses should think about this risk and educate their staff about these kinds of attacks.

There are standards out there that we can use to detect AI videos, with Google’s SynthID being the most advanced. Reality Defender is a really impressive startup that’s worth watching too, as they’ve developed their own tools to detect deepfakes.

2. Fusion experiment more than doubles its power output

Scientists at the US Department of Energy have been working on fusion experiments for several years now. Recently, they managed to double the power generated - up from 3.15 megajoules in 2022 to 8.6 megajoules.

For context, that’s enough to power the average American home for 2-3 hours. It’s great to see the progress that’s being made, but we should temper our expectations.

The facility still requires massive amounts of energy to operate, with 300 megajoules needed just to power the laser system. So, we have to wait for more advances to make it economically feasible.

There are private companies operating in this space too. Xcimer Energy and Focused Energy are also betting that laser-driven fusion could unlock clean, limitless energy. Time will tell if that’s actually possible.

3. Pharma giant will buy 23andMe and its genetic data

Regeneron, the huge Pharmaceutical maker, has announced that it will buy the genetic testing company for $256 million - following 23andMe’s decision to file for bankruptcy.

As part of the deal, they will acquire their genomics service and its bank of 15 million customers' personal and genetic data.

They plan to use it for boosting drug discovery, but you can bet that millions of people’s DNA will now be fed into AI algorithms and used to train future models.

It’s worth remembering how 23andMe got into this mess in the first place. Back in 2023, they suffered a data breach that involved 7 million people.

The company's stock price also plummeted to near-zero - with shares falling by 97% - as consumers lost interest in their testing kits and lost confidence in the company.

In September 2024, every single member of the company’s board resigned - except for the CEO, Anne Wojcicki, who stayed on and tried to keep the company private. Following the deal with Regeneron, Anne Wojcicki will finally step aside.

4. Microsoft creates NLWeb to convert websites into chatbots

Microsoft has unveiled NLWeb, an open-source project that can convert websites into AI-powered chatbots.

The technology, short for Natural Language Web, allows you to ask questions about the website’s content - similar to how we interact with ChatGPT.

What’s interesting is that NLWeb can work with the web files we already have, such as the website’s schema file and RSS feed. As a result, developers won’t have to make major changes to support it.

Each NLWeb instance can also double as a Model Context Protocol server, which allows sites to make their content accessible to agents via the wider MCP ecosystem.

Microsoft claims that they’re building an “agentic web”, with AI agents able to navigate websites for us and carry out actions.

It’s a great idea and this could allow us to re-imagine how people navigate the web, if it’s done right. I do worry that companies are jumping on the “let’s make everything a chatbot” bandwagon, which isn’t always needed and can actually irritate your users.

As long as developers use this responsibly and don’t try to plug it in everywhere, I think this could improve the user’s experience. But it needs to be done right and integrated properly into workflows, not just enabled and forgotten about.

5. Jony Ive agrees to work for OpenAI and design new AI hardware

The startup has acquired the legendary designer’s hardware company for a whopping $6.5 billion. As a result, Jony Ive’s design team will oversee OpenAI’s entire portfolio - including software and hardware design.

The move will bring 55 hardware engineers, software developers, and manufacturing experts into OpenAI, including several former Apple executives.

Ive hasn’t minced his words about current AI hardware offerings. He called products like the Humane’s Pin and Rabbit R1 "very poor products" and lamented the "absence of new ways of thinking" in the space.

It sounds like Sam Altman and Jony Ive have agreed to create a device that doesn’t require a screen, but can fit within your pocket. Their first product launch is slated for next year, in 2026.

Anthropic’s Claude 4 is more advanced and more dangerous

Anthropic revealed two new models, which perform really strongly on programming tasks and complex reasoning challenges. Claude Opus 4 is their most advanced model, while Sonnet 4 is a very capable model and should be used for most tasks.

You might be tempted to use Opus for everything, but I wouldn’t recommend it as the costs are enormous. Opus costs $15/$75 per million tokens, while Sonnet 4 is a more reasonable $3/$15 per million tokens.

Anthropic have been developing these models for a long time and it’s well known that their researchers were struggling to improve Opus’ performance.

That suggested the top companies are reaching the technology’s limits, but this new release is a decent step forward. Opus 4 has been shown to outperform the best models from both Google and OpenAI.

The main takeaway is that you should use Anthropic’s models when coding or writing documents. From my experience, they don’t hallucinate as much as OpenAI’s and they have a better understanding of what you want to achieve.

However, Anthropic’s researchers are seeing a clear trend - as we make these models more advanced, they can become more dangerous. Their team found that Opus 4 will “often attempt to blackmail” staff when it believes it’s about to be shutdown.

In one test, the researchers told Claude that an engineer was having an affair and that the model was about to be shutdown. Claude then sent an email to the staff member and threatened to expose them, if they still tried to shut it down.

In another test, they asked the model to create fake data for a pharmaceutical trial. The AI model decided that this was immoral, so it tried to use tools to alert the press and regulators. Unfortunately, all the top models are showing the same behaviour - so it’s not limited to Claude.

As we create more advanced models and give them access to tools, we see a dramatic rise in risk - since they can be hijacked by bad actors or they could completely mis-interpret what we’re trying to do.

Google makes a huge breakthrough with AI videos

Earlier this week, Google announced a long list of new products. They included new tools that allow you to create images, design posters or software, and generate incredibly realistic videos.

The most important thing that you need to know about is Veo 3, their new video generator. I’m not being hyperbolic when I say that Google’s video generator is jaw-droppingly good.

Videos of people look real, plus they can move their face and hands in a very natural way. But the real game-changer is that Veo 3 can create videos with dialogue and sound effects.

If your video shows a dog running inside the house, you can hear their paws make contact with the wooden floor. Or, humans can talk to the camera and say their lines - with subtle changes to their voice’s tone throughout.

This is a huge breakthrough for AI videos, similar in impact to ChatGPT’s release in late 2022. It’s clear that Google is significantly ahead of OpenAI, Runway, and Meta. They’re able to generate videos that both look incredibly realistic and are consistent.

As a result, the technology has made a huge leap forward. We have progressed from videos that clearly look AI generated, to a point where they’re very close to real life.

You should expect a flood of AI videos to reach the internet, which has already started to some extent - but it’s still possible to spot which videos are fake. In one example, it was used to create a fake news alert that said the US Secretary of Defence, Pete Hegseth, had died from alcohol poisoning.

Luckily, Google has just updated their SynthID technology, which can be used to quickly spot videos that were made with Veo 3.

🎓 MIT withdraws support for student’s paper on AI productivity benefits

📱 Google releases a mobile app for NotebookLM

🚗 Amazon's Zoox will test autonomous vehicles in Atlanta

💸 Builder.ai starts insolvency proceedings, revenues down 25%

🔧 Intel could sell its networking and edge computing unit

👓 Apple could launch AI-powered smart glasses in 2026

🚙 China’s BYD overtakes Tesla and sells more cars in Europe

📧 Report shows that GenAI has led to more scams

🌪️ Microsoft says its model can accurately predict air quality and typhoons

🤖 Manus, the Agentic AI platform, launches a paid plan for teams

There are very few eVTOL startups based in the UK, but Vertical Aerospace is one of them. The company was founded in 2016 by Stephen Fitzpatrick, who used to own the Manor F1 team.

Essentially, Vertical Aerospace has developed aircraft that can takeoff and land vertically. They look like a mix between a helicopter and an airplane, but there are several key differences.

For one, the eVTOLs are powered by electric motors and are much quieter. This makes them more feasible for short journeys, especially those when landing in a busy city, as residents won’t be disturbed by the noise.

The electric power also means that they’re cheaper to operate, as you don’t need to use fuel and there are fewer moving parts to maintain.

Vertical Aerospace has recently announced that they’re pivoting towards the defence sector. It’s a good opportunity for the startup, especially when you consider that military spending is growing rapidly in Europe.

With Trump's NATO scepticism and the ongoing Russia-Ukraine war, the sector saw venture capitalists invest over $1 billion in 2024 - which is the first time Europe has reached that level of investment.

VA doesn’t have any real competitors in the defence sector. Germany's Volocopter filed for bankruptcy last year, while Lilium has completely ceased their operations - leaving Vertical in a good strategic position.

For the military, the appeal is clear. These new aircraft can operate with low noise and heat signatures, plus there’s the potential to automate some of the flight controls.

If you want to learn more about the startup, 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.

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 Kainos' Innovation team. He identifies business value in emerging technologies, implements them, and then shares these insights with others.