🎥 Google reveals an impressive AI video generator

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 ‏‏‎ ‎

• FTC’s investigation into big tech’s AI partnerships

• Google’s new AI tools for teachers and students

• US Government’s plan to compete with China on AI research

• … and much more

Let's jump in!

1. Google announces new AI tools for education

We start the week with Google, who have dominated the education sector in recent years - after the pandemic forced many teachers and students to move online. Google are keen to cement their position as the biggest platform for education.

Interestingly, users will be able to generate their own practice tests - which will help teachers to reduce their ever-increasing workload, and students who can use them to prepare for upcoming exams.

Teachers will also get help in coming up with lesson plans, thanks to Google’s Duet AI toolkit.

Education is a clear area that can really benefit from Generative AI. Teachers are often facing greater workloads and cuts to their budget, meaning that they have to deliver more with less - but some of that pressure can be relieved using AI tools.

It’s also a good learning opportunity, since teachers can clearly demonstrate to their students where generative AI can be really helpful - while also showing the limitations and how it shouldn’t be used.

2. Microsoft shifts focus to smaller, cheaper AI models

According to insiders at Microsoft, the company has formed a dedicated AI research team which will focus on creating smaller language models (LM).

The need to develop smaller LMs is a theme that I flagged a few weeks back, which will likely accelerate this year and is something to watch out for.

The shift in strategy is due to the increasing cost of running large language models and Microsoft’s wider ambitions to adopt generative AI within most of their products - especially their Office 365 suite.

These high costs can be prohibitive to their pricing strategy, as even lower prices are needed to convince customers. If the current costs are forcing Microsoft to charge similar prices to ChatGPT Plus, then users will directly compare the two services and will be much less likely to switch away from OpenAI’s ecosystem.

But lowering the price isn’t the only motivation here, they also want to widen the number of tools that can be realistically built. Not every feature will need to use a huge, powerful language model - for some use cases, a smaller model will perform just as well.

If Microsoft are able to provide a larger number of AI features, and still do it at a lower price point than their competitors, then this is a much easier sell for consumers.

3. FTC launch inquiry into big tech’s AI partnerships

The US Federal Trade Commission (FTC) have announced they are investigating the “investments and partnerships” being formed between AI companies and cloud companies.

The FTC will look whether companies - such as Amazon, Google, Anthropic, Microsoft, and OpenAI - are harming competition through these deals.

As AI startups have shown impressive results, like OpenAI and Anthropic, the bigger technology companies have been keen to invest in them and sign partnerships.

Both Amazon and Google have invested billions into Anthropic, which was started by former OpenAI researchers that wanted to focus more on AI safety.

Microsoft have done the same through their much publicised partnership with OpenAI, when they committed to a $10 billion investment and are also their largest shareholder.

4. US Government partners with big tech to accelerate AI research

The new program, called the National Artificial Intelligence Research Resource (NAIRR), is made up of 10 federal agencies and 25 organisations.

NAIRR aims to provide US researchers with advanced computing, datasets, models, and support from the top AI companies.

OpenAI, Google, Anthropic, Microsoft, AWS, Meta, and NASA have agreed to support the project - alongside many other tech companies. It will focus on 4 different areas:

• Open source AI research

• Projects that require greater security and privacy

• Development of interoperable software and tools

• Teach people about AI and upskill them through training

This follows on from President Biden’s executive order on AI safety, which was signed late last year, and is part of the US’ wider strategy as they aim to compete with China on AI research.

China has rapidly become a major contributor to global AI research and one study indicates that they have already overtaken the US.

5. Cyberattacks will be turbocharged by AI, say UK’s intelligence agency

The UK’s cyber intelligence agency, GCHQ, is expecting that there will be a huge surge in cyber threats in the next few years.

Ransomware is one area that will become amplified by AI tools, as they will lower the barrier to entry and provide more opportunities for cyber criminals.

According to their report, social engineering will also become much easier for criminals - with generative AI tools used to make their messages more convincing.

We have all seen phishing emails that are poorly written and contain broken English. They’re easy to spot, but that will become more challenging as LLMs are increasingly used to write the messages instead.

The NSA has recently warned that criminals and nation states are already using these tools to perfect their campaigns.

Google’s latest AI video generator

Google have shown off their latest project, called Lumiere, which is able to generate videos in just one step.

The results are certainly impressive, especially when you consider what they’ve achieved with cinemagraphs and inpainting.

Their cinemagraph feature allows you to upload an image and then select an object, which will then be animated by the model. For example, a photo of a steam train could be animated to show the smoke blowing in the wind.

With inpainting, you can outline a box boundary and tell the model to generate new content within it.

Much of the industry’s attention is currently focused on text-to-video generators, but the videos they create often look a bit strange and unnatural. When we use text to make these videos, we’re giving the AI model too much control over the process.

Take this video for example, the couple’s legs immediately stand out as being unnatural:

Instead, it’s better to use images and videos as the starting point - since it often leads to better consistency when generating new videos.

Look at the video below. The model was only shown the bottom half of the video, with a black box placed over the top to hide the rest. It was then asked to continue the video and generate a new section where the box is.

You can see the AI’s completed video below, which looks very impressive and is consistent with the “real” section at the bottom of the frame.

This means we can modify existing videos and reimagine how they might look. There’s lots of ways this could be used.

Advertisers could use it to create more targeted video campaigns for their products in different markets, without having to re-shoot the original video.

However, this technology could also make it much easier to create deepfakes of other people. Lumiere hasn’t been released to the public, with their paper suggesting that new tools are needed to prevent it being used to spread misinformation.

In the last year we have seen significant advances with image generators, like Midjourney’s models, and it looks like video generators are on a similar path.

📈 Microsoft is now a $3 trillion company

✂️ Microsoft lays off 1,900 Activision Blizzard and Xbox employees

📉 Tesla expects EV sales growth to be ‘notably lower’ in 2024

🔎 US DOJ and SEC open investigations into Cruise’s self driving operations

🤖 Dusty unveils a new version of its construction robot

💬 Meta release a prompt engineering guide for Llama 2

🔊 Voice cloning startup ElevenLabs has raised $80 million in funding

🚗 GM and Honda join forces to make hydrogen fuel cells

🖨️ MIT Researchers demonstrate rapid 3D printing with liquid metal

Lakera

This is a Swiss startup that aims to protect companies from the security weaknesses in large language models (LLMs) - such as prompt injections and data leakage.

Lakera were the creators behind the popular Gandalf game, which encourages people to “hack” the LLM and convince Gandalf to reveal the secret password. As you progress through the levels, it becomes harder to defeat Gandalf (good luck trying to beat level 8, I still haven’t been able to…).

They have used the millions of prompts from Gandalf to identify how people are defeating the model - and then created their own product that can defend against these attacks.

While companies like Azure and AWS do offer some prompt protection when you host an AI model on their platform, it’s unclear how advanced this is.

There is the real possibility that a company’s reputation could be damaged by the misuse of their LLMs and chatbots, as a Chevrolet dealership in the US recently found out.

In this case, users only asked the chatbot to mock the company. But if they were able to create malicious code using the chatbot, then it would become a much more serious issue.

In late 2023, Lakera announced that they had raised $10 million to fund their expansion plans - and they recently spoke at Davos about the different ways AI applications are open to attack.

This Week’s Art

This week has been dominated by Google’s new video generator, along with the FTC announcing their investigation into big tech’s AI partnerships.

A lot has been covered this week:

• Google’s announcement of new AI tools for education

• Microsoft moving to smaller models, as they aim to grow the number of AI features and reduce costs

• FTC’s inquiry into AI partnerships and competition

• The President’s plan to help US researchers and compete with China

• GCHQ’s report on how AI will lead to more advanced cyber attacks

• Google’s latest research into AI video generation

• And how Lakera are protecting companies from security weaknesses in LLMs

Have a good week!

Liam

Feedback

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.