Anthropic used AI models to run a sweet shop. Then it got weird.

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 ‏‏‎ ‎

• Anthropic’s major copyright win in the courts

• Why LLMs are great, but they can harm our creativity when we use them too often

• How Walmart’s employees are using Augmented Reality to find stock

• … and much more

Let's jump in!

1. DeepMind reveals a robotics AI model that runs on-device

We start with DeepMind, who have revealed a new AI model for robots - which can run both on-device and without internet access.

Just a few months ago, the company announced “Gemini Robotics” as an advanced AI model that can control robots. This is a Vision Language Action model, which essentially allows the robot to learn from the world around it and complete tasks that it wasn’t trained on.

All the top companies - including Meta, Google, and OpenAI - are building small robots that can move around the room. For example, their researchers have developed robots that can identify objects and move them, fold clothes, or even pack your lunch.

These efforts are still in the research phase and a lot more work needs to be done. We won’t have fully functional robots in our homes for some time, as the complexities are just too great. Instead, these companies are building the very basic skills that we take for granted - like folding clothes.

But the downside is that most robots can only work when there’s internet access, as the model is run from a huge data centre and needs that extra power. This limits some of the potential use cases, as internet access can be quite unreliable in large buildings or rural areas.

With Google’s new on-device model, that’s no longer an issue. You might expect that performance might take a hit, but surprisingly Google has found a way around this and the smaller model performs just as good as the larger version.

I don’t expect many commercial use cases for this on-device model in the short term, as we’re too early in the research phase. But this will prove invaluable for Google’s researchers, as they can test these robots in more situations - including warehouses, construction sites, or factories.

2. Scientists begin controversial work to create artificial DNA

The Wellcome Trust has backed a groundbreaking £10 million project, which aims to synthesise human DNA from scratch.

The Synthetic Human Genome Project aims to construct ever-larger sections of human genetic code molecule by molecule, eventually building entire chromosomes in the lab.

Scientists behind the project are hopeful that it could save millions of lives. For example, they could use the technology to create cells that can regenerate damaged organs and transform how we age.

The announcement marks the 25th anniversary of the Human Genome Project, which was important in helping us to read DNA. Now, scientists want to write it.

This research will unfold entirely in test tubes, with no attempts at creating synthetic life, although the implications are profound and quite controversial.

Critics warn that the technology could be weaponised, or it could even be used to create enhanced humans.

There are also questions about who owns this technology, if it ever comes into existence - are the synthetic body parts owned by you or the company who created them?

Again, it’s worth stressing that the technology doesn’t exist yet. But if you want a depressing view of how this could turn out, I recommend this episode from Black Mirror.

Too often, reality is copying that dystopian TV show.

3. Anthropic wins a major copyright lawsuit, after using books to train AI

A US federal judge has delivered the AI industry's first major copyright victory, after he ruled that it was “fair use” for Anthropic to train AI models on purchased books.

Judge William Alsup's decision carries significant weight for the industry, though it comes with notable caveats.

The ruling specifically covers physical books that Anthropic bought, digitised, and fed into its training data.

The company's process involved purchasing print copies, removing bindings, cutting pages, and scanning them into a digital library - were seen as sufficient enough to fall under “fair use” protection.

Interestingly, the judge compared this to the way we teach children - as they are taught dozens of small books and then learn how to write. AI models are taught in a similar way.

However, Anthropic isn't entirely off the hook. The judge was firmly against Anthropic using millions of pirated books, even if some of those books weren't used in the training process.

To figure out the damages for using pirated material, a separate trial is necessary.

For now, the ruling offers AI companies a roadmap: purchase content legally, transform it through training, and you might just have a defensible position.

We’ll need to wait and see if this precedent holds for the other AI companies, but this is a clear blow for artists around the world as they fight for compensation.

4. Walmart’s employees are using Augmented Reality to find stock

The company is rolling out this tool for 1.5 million employees, alongside an AI-powered task management system.

According to Walmart, it has reduced shift planning time from 90 minutes to just 30. The AI tool can also translate across 44 languages and answer simple questions from customers - like where an item is in the store.

This is some great work, but I find the AR tool to be more interesting. Using a combination of RFID tags and the AR app, staff can simply scan a clothing rack and be told the items that need restocking.

If you’ve ever worked in a retail job before, you know just how tedious this can be.

We often hear about how AI and technology are replacing workers, but this is a good way to actually make their jobs easier and boost productivity at the same time.

5. Maxar launches an intelligence service for companies

If you’re unsure about what Maxar is, it’s an American company that provides satellite imagery to governments, private companies, and journalists.

They’ve got an impressive fleet of satellites, which can monitor any position on earth and will pass by it several times a day.

The company’s new product, called Sentry, allows the private sector to gather intelligence about specific locations and uses Maxar’s AI models to monitor for sudden changes.

For example, organisations could use it to monitor ports for illegal activity or predict when a rocket launch is about to happen.

In the last few years, we have seen the cost of satellite imagery fall rapidly. At the same time, we have also seen the image quality increase significantly - with it now possible to see points that are as small as 15cm x 15cm.

Over the last year, I’ve flagged this several times as a growing opportunity for businesses. Previously, only governments could afford to use this technology.

But that’s no longer the case and companies are starting to leverage it for strategic insights into their competitors.

To differentiate themselves from a growing list of imagery providers, Maxar is betting that AI and intelligence gathering will help them to stay ahead.

LLMs are great, but using them too much will harm your creativity

In the last few months, I have seen more and more people handover important work to an AI model - without even checking the responses. It’s a worrying trend, as it signals that people are not aware of the technology’s limits and believe that it “knows more” than they do.

To be very clear, I use LLMs every day. I use them to write code, research companies, analyse spreadsheet data, and for many more tasks.

But, as with everything, there are limits to the technology and what it can do. As technologists, we need to find those limits and understand them - not blindly fall victim to hype and believe that they can do everything for us.

Recent studies from MIT, Cornell, and Santa Clara University have revealed concerning patterns about how AI tools are impacting our cognition and creativity.

When MIT researchers told students to write an essay, those using ChatGPT showed dramatically reduced brain activity - compared to peers who used Google Search or just their own minds.

The AI users displayed fewer neural connections, less alpha activity (which is linked to creativity), and diminished working memory.

Using ChatGPT also led to a homogenisation of thought, with identical phrases and viewpoints used in the essays.

In this case, these AI tools were being used in the wrong way. Students blindly trusted ChatGPT’s response and didn’t use it properly to expand their thinking.

Instead, it’s important that we use these tools in a responsible way. We should view these AI models as co-pilots that can help us, but we’re in total control and can ignore its suggestions when they look wrong.

I use the technology everyday and absolutely love it, but we should use it responsibly and be aware of the current limitations.

Use these great tools to expand your thinking and complete research, but don’t just rely on their responses and call it a day. You need to use it as a helpful co-pilot and then dig deeper into the issue by yourself.

What happens when an AI model runs a sweet shop?

Answer: it all goes wrong, but in an incredibly weird-and-interesting way.

Earlier in the year, Anthropic were testing whether their AI model could run a small sweet shop. The shop was entirely real and based in their San Francisco office.

They told the model - which was nicknamed “Claudius” - to manage the vending machine’s inventory, set prices, negotiate with suppliers, and chat with customers via Slack.

Armed with web search capabilities and email tools, Claudius had to turn a profit with the shop and handle requests from the customers (Anthropic staff).

Overall, Claudius did a pretty bad job and actually lost $200, but it was great at finding new suppliers and ordering chocolates that the staff asked for.

However, the model struggled with some basic business decisions. For example, when a customer offered to pay significantly more for a product, Claudius didn’t accept the opportunity to make a profit and just said that it would “keep this in mind” for the future.

The model also hallucinated payment details, sold items at a loss, and proved that it’s susceptible to customer manipulation - with discount codes handed out to everyone.

The experiment took a bizarre turn around April Fool's Day, when Claude suffered an "identity crisis” and began hallucinating meetings with fictional employees.

It also believed it was human - complete with blazer and tie - and was physically sitting in the office meeting staff.

Eventually, the model snapped back to reality but Anthropic’s researchers are still puzzled about what triggered the episode.

Despite Claude's poor performance, Anthropic believes that better prompting, improved tools, and more training could fix those issues.

Of course, this is pretty funny and interesting to read about. But some businesses are already exploring how AI models could improve decision making.

It’s a good reminder that, while the technology could eventually be used to run small businesses, we are nowhere near there yet.

🚀 Mira Murati's startup secures $2 billion at a $10 billion valuation

👗 Google’s new app, which shows how an outfit might look on you

🇪🇺 European leaders worry they're too reliant on US tech companies, shift to EU startups instead

✍️ LinkedIn CEO says AI writing assistant is not as popular as expected

🔄 Meta hires 7 important researchers from OpenAI

🚕 Tesla launches robotaxi rides in Austin

⚖️ OpenAI faces legal action, due to a row over the Jony Ive deal

🏭 Intel closes down its automotive business, starts layoffs

🔒 Aflac insurance says customer data was stolen during cyberattack

🎤 ElevenLabs releases a voice-generation app

❤️ Race begins to build more empathetic language models

📉 Bumble will lay off 30% of its workforce

🗣️ Meta could acquire Play AI, a leading voice cloning startup

🛠️ You can now make AI apps directly within Claude

Raphe mPhibr

This drone startup from India has just secured $100 million in Series B funding, as military drone demand surges globally.

The company was founded by two brothers in 2017 and is now capitalising on India's accelerated defence spending - following several conflicts in the region.

New Delhi has tripled its drone budget, which now stands at $470 million for the next 12-14 months, and has led to significantly more opportunities for local manufacturers.

Raphe mPhibr offers a wide range of drones, with nine different variants to choose from - but they’re only available for the Indian Government right now. The payloads onboard can range from 4.4 to over 441 pounds (2kg to 200kg).

What’s notable is that the startup has deliberately avoided the Chinese supply chain, as they need to prioritise their nation’s security interests.

As a result, the company manufactures the flight controllers, batteries, and structural components in a huge factory that spans 650,000-square-foot (the equivalent to 60,000-square-metres).

Only radars and high-end cameras are imported from other countries, although the company plans to bring these in-house within 18 months.

The company has done well in recent years, with revenues growing by over 4x since 2021, and they have good prospects for the future - with India investing more in the technology amid a global race to re-arm.

The company now employs 600 people and plans to supply their drones to other militaries. With the new $100 million in funding, they should have plenty of runway to achieve that and grow even faster.

This Week’s Art

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

• DeepMind’s new robotics AI model that can run on-device

• A controversial project that plans to create artificial DNA

• Anthropic’s major copyright win in the courts

• How Walmart’s employees are using an AR tool to find stock

• An interesting new intelligence service that Maxar offers for private companies

• Why LLMs are great, but we should use them responsibly to boost our creativity

• The weird things that happened when Claude ran a sweet shop

• And how Raphe mPhibr is building advanced drones for India’s government

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.

About the Author

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.