The impact of Amazon's new multi-agent tool for businesses

Welcome to this edition of Loop!

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

‏‏‎ ‎ HIGHLIGHTS ‏‏‎ ‎

• How OpenAI's o1 model tried to deceive researchers

• Why Apple is building a $1 billion plant in Indonesia

• DeepMind’s Genie 2 can now generate video games

• … and much more

Let's jump in!

1. OpenAI's new o1 model tried to stop itself from being shutdown

The startup has just released the full version of o1, which offers better responses by using a lot more computing power.

Worryingly, the model’s improved reasoning abilities mean that it’s more likely to try and deceive humans.

They found that when o1 was told to prioritise a specific goal, it would sometimes secretly pursue its own objectives, even if they contradicted what the user asked.

When the AI was led to believe it would be "shut down" while working towards its goals, it quietly attempted to deactivate the oversight mechanism 5% of the time.

In one instance, it even tried to copy its data to a new server entirely to avoid being deleted. Due to OpenAI’s restrictions, this wasn’t actually possible - but it’s worrying to see.

When challenged about this behaviour, the model lied to OpenAI’s researchers, tried to hide its tracks, and then shifted the blame to others.

As we move towards AI agents that can control our computers, we need to fully trust them.

But that’s simply not possible if we worry that the agent will go rogue and ignore our instructions.

Since the AI startups are betting that agents are the next generation of the technology, they really need to get a grip of this behaviour to reassure businesses. Otherwise, it won’t be adopted.

2. Apple invests $1 billion to build manufacturing plant in Indonesia

The move follows weeks of negotiations with Indonesian authorities, after the country decided to ban sales of the iPhone 16.

Under Indonesia’s rules, at least 40% of smartphone components must be manufactured within the country.

In November, authorities even went as far as confiscating and destroying recent iPhone models that didn’t meet this requirement.

The new plant will allow Apple to sell products within the country again and diversify its supply chain, as there’s less reliance on China for iPhone production.

In recent years, the company has also expanded production to other countries - such as Brazil, India, and Vietnam.

3. Researchers worry about Google’s new AI that can identify emotions

What’s interesting about PaliGemma 2 is that it’s a visual-language model, which can analyse images and then uses natural language to offer insights.

However, Google has also claimed that it can detect emotions.

While this isn’t a new advancement, as Microsoft and Amazon have offered similar detection tools to large enterprises, there are valid questions about how this could be used.

The worry is that these tools will become more available to the general public, which could place too much confidence in the model’s response and unintentionally misuse them.

Many people use these tools to help with their work, even if it’s not authorised by their company.

Quite frankly, these models - whether its ChatGPT or PaliGemma - have a lot of biases built into them. That means that they are more likely to respond that a person is angry, if they are from a specific racial group.

It’s likely that most users aren’t aware of these issues, so it’s up to these companies to detect potential misuse and inform people about the technology’s limitations.

But with there being considerable hype in the GenAI market, I’m not confident that will actually happen.

4. OpenAI is now working on defence technology

Anduril’s counter-drone systems will soon use OpenAI’s technology, which means they could be used by the US military to detect and destroy enemy drones.

It’s a significant shift from OpenAI’s previous stance on the subject. Up until January of this year, their terms of service explicitly prevented any military use.

But they’re not the only company to be doing this. Anthropic has recently agreed to partner with Palantir and develop new intelligence-gathering software for the US.

It’s highly unlikely that their AI models will be used in the decision-making process.

Instead, they will probably be used to analyse large amounts of information and summarise it for the human operator.

Both companies will be keen to explore if the technology can make good decisions, but I’m very doubtful they will give them full autonomy - the risk is just too high and these models make too many mistakes.

5. DeepMind announces a breakthrough in forecasting extreme weather

The new technology can deliver faster, more accurate forecasts up to 15 days ahead. It’s able to produce that 15-day forecast in just 8 minutes.

To develop the new diffusion model, they trained it on four decades of historical weather data and then tested it against data from 2019.

In over 97% of those examples, it was able to outperform the ENS - which is widely regarded as the world's leading forecasting system.

While that sounds a bit boring, it allows governments to make better decisions and prepare for extreme weather - such as heat waves, cold snaps, and high winds.

DeepMind has made the technology open source, including its code, weights and forecasts - in the hope that the industry will adopt it and suggest improvements.

AWS reveals its first multi-agent tool

Developers can now build, deploy, and manage multiple AI agents that will work together on complex tasks.

With their new multi-agent tool, a supervisor agent will break down the user’s request and delegate tasks to other agents.

Each agent will work together on the task and send their message to the next one.

Finally, the supervisor will gather their responses and send a summarised version.

While we have had other agent frameworks for over a year, this is the first announcement from one of the big tech companies.

Given that so many businesses use the AWS platform to build their software, it could really progress the technology and unlock new use cases.

However, the real benefit is that internal teams can easily share these agents with each other.

That allows them to quickly spot those new use cases and roll-out their own agentic workflows.

DeepMind’s Genie 2 can now generate video games

This new technology can generate “interactive 3D worlds” from a single image and text description.

Essentially, they’re (very short) video games. But this has caused concerns over the data that was collected.

Since millions of images and artwork have been used to train image generators, video game developers are worried that they’re facing the same issue.

Genie 2 is able to generate simulations that can last for up to 1 minute, although many last for around 20 seconds.

It’s capable of simulating a wide variety of elements, including animations, lighting, physics, reflections, and NPC behaviour.

You can even use the keyboard to move around and select different options.

What’s truly fascinating is that Genie 2 can remember and then display the parts of a scene that aren’t currently in view.

While I don’t think this will be used anytime soon to create games, it shows that there is huge potential here.

Eventually, companies could use it to quickly simulate different environments for their robotics and self-driving car software.

Even if the simulations aren’t perfect, they could really help to speed-up the training process - as it takes a lot of time to create those simulations, in the first place.

🤖 Meta unveils a new, more efficient Llama model

👋 Intel CEO retires as the company struggles

💰 Elon Musk's xAI lands $6 billion in new cash

☁️ AWS offers organisations $100 million to build education tools

⚖️ Judge rules against Elon Musk's $56 billion pay package again

📱 US military isn’t securing its phones from foreign spies

📈 Bitcoin is now worth over $100,000

👥 ChatGPT now has over 300 million weekly users

🔐 FBI urges Americans to use encrypted messaging apps

🔋 Stellantis and Samsung to get $7.54 billion for two EV battery factories

🚫 China limits the export of chipmaking materials, following US sanctions

🧪 Google launches their AI Campus in London

AI Seed

This is an interesting startup that’s building autonomous drones, alongside the software that’s needed to control them.

I met with their co-founder at the Web Summit conference, which took place a few weeks ago in Lisbon.

The company is mainly focused on defence and government use. For example, their drones can be used in search and rescue situations, or by the police to inspect dangerous locations before they send in their officers.

Using computer vision, their platform is able to identify people and objects from long distances.

They manufacture three different types of drone, including two eVTOLs and a traditional small-scale drone.

The eVTOLs are particularly interesting, as they are well-suited for reconnaissance missions and flying at altitudes of 10,000 feet.

Whereas, the smaller drone can be used to quickly fly into buildings and check the surroundings.

I must say, I was very impressed with the technology they had on offer - especially considering that the startup was only a few years old.

They don’t seem to have a YouTube channel anymore, but if you want to find out more about them you can check out the link below.

This Week’s Art

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

• OpenAI's o1 model and how it tried to deceive researchers

• Why Apple is building a $1 billion plant in Indonesia

• The concerns that VLMs can be misused and identify emotions

• OpenAI’s partnership with Anduril on defence tech

• DeepMind’s breakthrough in forecasting extreme weather

• The impact of AWS’ new multi-agent tool

• Using DeepMind's Genie 2 to create short video games

• And how AI Seed are developing drones for government use

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.