You’re receiving this email because you subscribed to the newsletter. If you’d like to unsubscribe, you can manage your preferences here with one click (or use the button at the bottom of the email).

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. Global data centre spending has surpassed oil investment for first time

We start this week with the surprising news that global spending on data centres will surpass oil investment.

According to the International Energy Agency, global spending will hit $580 billion this year and surpass oil investment by $40 billion.

While it’s an important shift, it also highlights that much of the modern economy now relies on digital infrastructure. But the figures become even more eye-catching when you look ahead.

It’s expected that AI data centres alone will quintuple their electricity consumption by 2030. The US will account for half this growth, with Europe and China making up the remainder.

That’s leading to problems for ordinary people, as most new facilities are being built around major cities. Dublin has paused all new planning applications until 2028, while data centres are spiking electricity costs in Maryland and Virginia.

On the plus side, the IEA projects that renewables will power the majority of new data centres by 2035.

Over the next decade, renewables should supply around 400 terawatt-hours to data centres, compared to 220 from natural gas, and small modular nuclear reactors could contribute another 190.

2. Coding assistant Cursor raises $2.3 billion, just months after its previous round

Cursor has raised a whopping $2.3 billion from investors, just five months after its last funding round. The company is now valued at $29.3 billion, which is alarmingly high for a company that has no moat.

The Series D round almost triples the company's previous $9.9 billion valuation from June, when it raised $900 million. This new funding will be used to improve Composer, which is an AI model that Cursor released last month.

If you’re not aware of Cursor, they create coding tools for software developers - which allow AI agents to take action and write the code for them. Their products are popular with developers, but they’re facing stiff competition from the larger companies.

Today, Cursor relies on external AI models from Google, OpenAI, and Anthropic to power its product. That’s a dangerous position for the company to be in longer term, as model costs have remained the same for the last year - but demand has spiked as developers use them for more tasks.

This was exactly the problem that forced Windsurf, another AI coding company, to quickly put itself up for sale. It’s incredibly hard for startups to absorb these costs, which are only growing over time.

If you want to read more about Windsurf and the problems with its business model, you can read my write-up here.

Ultimately, Cursor wants its proprietary model to entice new developers and reduce their dependence on OpenAI and Anthropic. But that’s going to be a big test for Cursor in the coming year, as those two companies are refining their own AI tools for developers.

3. OpenAI says the new GPT-5.1 is warmer and has more personality

The startup has released a minor upgrade for GPT-5, following its underwhelming debut in August. It includes two options: GPT-5.1 Instant and GPT-5.1 Thinking.

According to OpenAI, the instant version should be "warmer, more intelligent, and better at following your instructions” - while the extended-thinking model should perform faster.

Depending on your question and how complex it is, ChatGPT can automatically choose the best version for your task.

In my own experience, it seems to be marginally better. But the model still seems to be behind those offered by Google and Anthropic.

I regularly use Anthropic’s models to write code, summarise documents, and get general advice - and Claude Sonnet 4.5 is great at all those tasks. It also responds in a very natural way, whereas OpenAI’s GPT models often sound quite robotic.

If you’ve only tried ChatGPT so far, I’d highly recommend you try Google’s Gemini model and Anthropic’s Claude instead. They allow you to use them for free, with some limits, so you can easily compare the results between the three models and pick your favourite.

4. Yann LeCun plans to leave Meta and create his own startup

For years, Meta’s huge AI team has been led by Yann LeCun - an incredibly influential researcher and is often referred to as the “godfather of AI”.

Recently, his work has spanned across machine learning, computer vision, robotics and computational neuroscience.

But the Turing Award winner is planning to leave within months and will launch his own startup, according to the Financial Times. This new venture will focus on world models, which are AI systems that can understand their environment and simulate cause-and-effect.

Meta’s AI division has faced issues in the last few months, following its $14.3 billion investment in Scale AI and rapid hiring of engineers from the top AI labs.

LeCun's work at Meta's Fundamental AI Research (FAIR) division - which is focused on long-term research that might not pay off for five to ten years - has been somewhat sidelined as Mark Zuckerberg pursues more immediate priorities.

Given so much change within the organisation, and LeCun’s vocal scepticism of the AI hype we see today, it’s unsurprising that he’s keen to move on and start his own research company.

5. Startup admits its “AI notetaker” was actually a human

The co-founder of Fireflies, an AI startup that’s now valued at $1 billion, has said that employees would manually join customer meetings and pretend to be an AI notetaker.

The pair would dial into the call, take notes by hand, and then send them over within 10 minutes - with customers charged $100 per month for what they believed was an automated tool.

After they logged over 100 meetings, the founders finally secured enough funding to keep their startup afloat and eventually automated the process.

I’m sure the co-founder is happy with lots of PR for his startup, but it’s hardly a ringing endorsement for the company and how trustworthy it is.

Lots of companies are worried that their data could be used to train new models, which puts them off adopting these tools.

Fireflies might be another example of Silicon Valley's "fake it 'til you make it" culture, but it's a reminder that you can't always trust what tech companies tell you.

World Labs releases its first world model

World Labs has just launched Marble - its first world model that can transform text, images, videos, or panoramas into 3D environments. You can also edit these virtual worlds and have more control over the creative process.

The company came out of stealth just over a year ago with $230 million in funding, and this release puts them well ahead of rivals like Decart and Odyssey, who've only released free demos.

What sets Marble apart is that it can create downloadable 3D worlds, rather than generating them on-the-fly. This leads to more consistent results and allows you to export the results as a Gaussian splat or video.

You can generate four worlds for free, or pay extra to generate dozens more and have the commercial rights you need.

I have to say, the results are very high-quality (note: I had to heavily compress the gif above). You can walk around and explore the virtual world you built, as if it was a video game.

The graphics look very impressive, but they do become fuzzier when you zoom in. Regardless, this is incredible to see and the World Labs team has done a great job.

I’ve been emphasising world models and the huge progress we’re seeing here, for some time. This could have lots of different use cases, especially in the gaming and robotics industries.

Since it’s incredibly expensive and time consuming to create training data, robotics labs could use world models to simulate this instead.

If you can create thousands of virtual worlds, which look exactly like a suburban home, this will be invaluable for companies and could allow them to speed-up the training process.

It could be a few more years before those companies start to use it, as the technology needs to improve further and become more reliable, but the direction of travel is pretty clear.

Google’s agent can play games

Continuing with the theme of virtual worlds, DeepMind has made progress with their AI agent that can understand and play games.

Unlike its predecessor, which could only follow basic instructions and had a 31% success rate for complex tasks, SIMA 2 has been able to double that performance. Google was able to achieve this huge jump in performance, once they integrated it with their Gemini 2.5 model.

But the real breakthrough here is how SIMA 2 actually thinks through problems. When asked to walk to "the house that's the colour of a ripe tomato," it’s able to determine that tomatoes are red, then navigates to the red house.

Or you can use emojis and tell the agent what to do. For example, if you send it “🪓🌲”, the agent will chop down a tree.

The agent also shows some ability to self-improve. The latest version can generate tasks on its own and uses reward models, which allow it to learn through trial and error.

I’m a bit skeptical about how well this performs in reality, as companies tend to overstate this ability, but this is interesting nonetheless. Similar to World Labs, DeepMind also sees this as a stepping stone towards more capable robots in the real world.

But it’s worth stepping back here and looking at what Google’s trying to achieve. Google is developing several technologies that - when combined - will allow them to create more reliable robots.

Genie is being used to create photorealistic, virtual worlds. SIMA allows advanced AI models to navigate those virtual worlds, complete tasks, and learn from their mistakes.

These learnings can then be applied to their growing list of AI models, such as Gemini Robotics 1.5, which is already being used to control robots in the real world and perform basic tasks.

🚀 Blue Origin becomes the second company to land a rocket booster on Earth

👨‍💼 Tim Cook could step down as Apple CEO next year

👜 Apple made a $230 crossbody sock for your iPhone

💬 OpenAI says it has fixed ChatGPT's em dash problem

🛍️ Google's AI shopping tool can call stores for you

🎙️ ElevenLabs releases an AI marketplace for brands to use famous voices

🚗 Waymo's robotaxis can now use the freeways in LA, San Francisco, and Phoenix

📓 Google's NotebookLM adds a "Deep Research" tool

💰 Anthropic says it will spend $50 billion on data centres

🤖 Wonderful raises $100 million to develop AI agents for customer service

⚠️ Hackers used Anthropic's AI model Claude to attack websites

✈️ Neros raises $121 million to build drones for the US military

This Boston startup has secured $150 million in Series B funding to commercialise a new sensor for autonomous vehicles.

Currently, self-driving cars use a mixture of cameras, lidar, and radar to monitor the real world. But weather conditions can cause these sensors to fail and lead to safety issues.

This is because cameras often struggle with the sun’s glare, lidar doesn’t work well in fog, and radar lacks the resolution that’s needed.

But Teradar has been able to create a sensor that works in all these conditions. To do this, their technology uses the terahertz band, which allows it to combine lidar's high-resolution with radar's reliability.

The company is already working with five major US and European automakers to validate the technology, with plans to bring it to production vehicles in 2028.

The defence sector is clearly interested too, judging by the investment from Lockheed Martin, although the company is primarily focused on the auto sector.

If their trials work as expected, this could be incredibly useful for both sectors and become very profitable for Teradar - especially if they focus on defence.

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