✈️ Why the US is testing AI fighter jets against human pilots

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 ‏‏‎ ‎

• Why Microsoft are preparing to launch AI Agents

• Boston Dynamics’ new humanoid robot

• How Google uses AI to help predict natural disasters

• … and much more

Let's jump in!

1. Boston Dynamics’ humanoid robot goes electric

We start with Boston Dynamics, who have completely redesigned their Atlas robot. The announcement video was both fascinating and weird, as it moves like something you’d see in the “I, Robot” movie.

The new robot looks a lot more agile than its predecessor, with it no longer using hydraulics. Hyundai will be one of the first companies to pilot the new version, but Boston Dynamics are facing much stronger competition than before.

Several startups have been given hundreds of millions to develop humanoid robots, such as 1X, Agility, Figure, and Tesla. Some are even working with top AI companies, like OpenAI, to develop custom models for their robots.

While Boston Dynamics have led the field for many years, the other companies aren’t far behind them.

2. Meta releases Llama 3, says it beats other open-source models

Llama 3 is available in different sizes, with 8 and 70 billion parameter options to choose from. Meta claims that it outperforms similar models, such as Mistral 7B and Google's Gemma 7B, which are pretty popular in the open-source community.

However, we should take notice of the dizzying amount of data that Meta used to create it. Llama 3 was trained on over 15 trillion tokens, which is 7 times bigger than Llama 2’s training set.

It’s difficult to comprehend just how much data these companies are collecting for their new models. By using much more data, they’re able to increase performance and make the model more useful than before.

As I mentioned recently, this trend is accelerating at an exponential rate - as companies race to collect everything they possibly can. They’re going to reach the limit at some stage, but it looks like simply giving these models more data will lead to better results.

Meta is working on a 400 billion parameter version of Llama 3, which will be even more capable than what we have today - especially when it comes to multi-modal abilities.

3. Microsoft and Google are preparing to launch AI Agents

This is a major technology shift that’s on the horizon and will change how companies work. Technically, AI Agents are already here. Microsoft have their open-source framework called AutoGen and Google released a tool last week.

But there’s much more on the table here. Businesses want an easy-to-use interface for creating multiple Agents and have them talk to each other. They can then customise the Agents to make them specific for their use case.

This is what the top companies are currently working on, with the possibility that Microsoft might even announce some new tools at next month’s Build conference.

Multi-agent workflows have the most potential for businesses, as they allow you to generate content and then ask the other bots to check it for mistakes. This can happen in an iterative loop, with the content automatically improved each time.

4. UK Government is drafting new laws for AI models

The Department for Science, Innovation and Technology is developing legislation to mitigate potential harms caused by AI, but we don’t expect a bill to be announced just yet.

Government officials have also been looking at changing copyright law, with companies & individuals able to opt-out and stop their data being used to train AI models.

Previously, the UK Prime Minister has argued against regulating the sector. But as the EU and US are bringing in their new laws, the UK will need to follow suit and provide clarity for their own tech companies.

The UK Government has been criticised for their light touch approach, as many companies aren’t sure where the boundaries are.

Regardless, the UK won’t have much room to manoeuvre outside of what the EU and US have set, as their market size is considerably smaller.

5. Google uses AI to help predict natural disasters

Their moonshot factory, known as X, has released details about new AI tools for the US National Guard. These will help the National Guard to respond more quickly to natural disasters, which have become a lot more common over the last few decades.

Google’s prototype is able to analyse satellite imagery of a disaster in seconds, identifying damage to critical infrastructure and then producing labelled maps for the National Guard to efficiently deploy resources.

Currently, the organisation’s damage assessment process is manual and very time-consuming, as analysts have to sort through thousands of images. But thanks to this project, this can be done much faster than before.

Microsoft’s VASA-1 can create deepfake videos

Researchers from Microsoft have unveiled VASA-1, which is able to create a video of someone talking or singing - using just one photo.

The technology could potentially power virtual avatars, but it also raises concerns about the misuse of such tools to create misleading or harmful content.

VASA-1 uses machine learning to analyse both a static image and an audio clip. It then generates a realistic video with precise facial expressions, head movements, and lip-syncs to match the audio.

The model was trained on the VoxCeleb2 dataset, which contains over 1 million examples of celebrities talking.

While the videos aren’t 100% perfect, they are really good when you consider only one image is needed. Although, the potential here for misuse is huge and it’s why Microsoft have decided not to make it available for the public.

It follows on from similar research by Alibaba’s team, which I covered last month, and they were able to achieve very similar results.

This technology will only get better over time, but will become more difficult to detect as a result. There are several companies out there who can reliably detect deepfakes, with Reality Defender being one of the stand-out options.

US successfully tests dogfight between an AI and human pilots

DARPA has revealed that an AI-controlled jet has successfully faced a human pilot during an in-air dogfight test, which was carried out last September.

The US Air Force has been experimenting with AI capabilities since December 2022, which forms part of their Air Combat Evolution (ACE) program.

The AI system was able to autonomously fly a modified version of the F-16, with human pilots on board and able to disable it when necessary - although they never needed to do so.

The AI-controlled jet went against another F-16 that was controlled by a human pilot. The US said that both aircraft demonstrated “high-aspect nose-to-nose engagements” and got as close as 2,000 feet at 1,200 miles per hour.

The US military aren’t the only ones who are testing how AI can be used in fighter jets. Shield AI is an American company that’s working on the same thing, along with Helsing in Europe.

It’s worth stressing that these AI systems are completely different to the type we see in ChatGPT, which regularly gets things wrong, and is being used in a more restrictive way to control the aircraft.

Over time, we will start to see AI systems that can control military aircraft - especially those that have no humans on board, like drones.

The war between Russia and Ukraine has led to some innovative strategies with drones, which will only encourage more AI adoption as Ukraine seeks to defend their citizens.

🚗 6 Waymo robotaxis block traffic to San Francisco freeway

🏥 Hugging Face releases a new benchmark for testing GenAI on health tasks

❌ Tesla lays off more than 10% of workforce, including high performers

🔍 Meta adds its AI chatbot to the search bar

🇯🇵 OpenAI opens Tokyo hub, adds GPT-4 model optimised for Japanese

✂ Stability AI lays off 10% of its workforce

NZero

NZero is an interesting startup that’s focused on tracking carbon emissions. Their main focus is to help building owners in NYC to estimate their carbon emissions, even down to the hour.

Their algorithm uses the building's address and any additional information provided by the owner - such as square footage and the types of heating & cooling systems - to generate more accurate estimates than was possible before.

Building owners can also use it to identify the most cost-effective upgrades and retrofits, which have the added benefit of reducing the building’s emissions.

As governments around the world come under increasing pressure to meet emissions targets, they will be looking at how large buildings can reduce their own footprint. You can check out the company below, if you want to learn more.

This Week’s Art

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

• The new humanoid robot from Boston Dynamics

• Meta’s release of Llama 3

• Why Microsoft and Google are preparing to launch AI Agents

• UK Government’s plans to regulate AI models

• How Google uses AI to predict natural disasters

• Microsoft’s latest research for creating deepfake videos

• US military’s dogfight between an AI and human pilots

• And how NZero are helping building owners to reduce emissions

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