🔬 Meta’s new ambition is to build AGI

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 ‏‏‎ ‎

• A new tool from Microsoft that can compress LLM prompts

• Anthropic’s research on training AI models to be malicious

• Magnific’s incredible image enhancer

• … and much more

Let's jump in!

1. Meta is focusing more on AI products and AGI

There are several things going on here. Firstly, Meta is adjusting how its leadership team is structured - with a greater focus on bringing their AI research into their product lineup.

This isn’t surprising. Meta have a fantastic research team and they regularly release new AI tools around computer vision, generative AI, and robotics.

But it’s not always clear how many actually get integrated into their social media platforms, which is something they hope to focus on going forward.

Secondly, Meta is in a huge battle with the other tech companies to keep their AI talent - and to attract new researchers. To keep their talent, companies like OpenAI and Google are paying as much as $1 million in compensation.

However, it’s not the only factor for some of these researchers - the ability to work on hugely ambitious projects can be even more of a draw.

By publicly stating that their goal is to build Artificial General Intelligence (AGI), Mark Zuckerberg hopes that it will give them a boost in attracting that AI talent - while also making it clear that the “metaverse” isn’t their only goal.

2. Anthropic says that AI models can be secretly trained to be malicious

As time goes on, we’ll start to see models that have been created for more specialised use cases. This is only natural, since the large tech companies can’t account for every industry and job type.

Anthropic’s research team have explored what would happen if you trained a large language model, that’s similar to OpenAI’s GPT, and included both “good” and “bad” behaviours.

If someone developed a model that only had “bad” behaviours built in, like creating malicious code, then it would be easy for us to spot and wouldn’t be commonly used.

But if it has a mix between the two, and only becomes malicious when a specific phrase is used, then the model would be much less likely to be identified as harmful - and could become widespread within businesses.

Anthropic found that they could fine-tune a model and add specific triggers, but that it was almost impossible to prevent the malicious behaviour afterwards. For example, when the user indicates that the current year is 2024, their model was triggered and started writing code that had security vulnerabilities.

The key takeaway from their research is that this behaviour is difficult to spot and existing prompt techniques could not prevent the malicious responses. It shows that we need more robust methods to test new models for AI safety, while also being careful about what AI models we use.

3. DeepMind’s latest AI can solve geometry problems at a gold standard

The company says that their new AI system, called AlphaGeometry, can solve geometry problems at the level of an International Mathematical Olympiad gold medalist.

Mathematical reasoning has been a focus of DeepMind for some time now, with other research projects looking at how they can solve math challenges and identify faster sorting algorithms.

AlphaGeometry was able to solve 25 Olympiad geometry problems within the standard time limit. For context, the average human gold medalist was able to solve 25.9 problems.

To do this, they used a combination of two systems - a neural language model to quickly predict what constructs might be useful, along with a symbolic deduction engine which will slowly analyse each and select the best suggestions.

What’s interesting with this project is that DeepMind used synthetic data to train AlphaGeometry - with over 100 million examples used in the training process.

A lack of training data is a common bottleneck for developing these types of huge, but very specific AI tools. Generating your own synthetic data is an exciting way to try and get around that issue.

4. OpenAI wants to crowdsource governance ideas into its models

A new team will be tasked with implementing the public’s input on how future AI models should be developed. It will be called the Collective Alignment team and it’s part of OpenAI’s strategy to show that they do not completely dictate how these models should behave.

It’s important given that the majority of the world’s democracies will be facing elections in 2024, with there likely to be growing pressure on politicians to regulate new AI development.

Some steps towards this have already been taken in 2023, following President Biden’s executive order on AI safety and the EU reaching agreement on how to implement the AI Act.

Scrutiny from regulators is likely to continue this year. OpenAI will be keen to demonstrate that they aren’t the only people who are writing the rules for their models.

5. BMW will use Figure’s humanoid robot at their US factory

While there weren’t many details on how exactly the robot will be used, it’s likely to focus on tasks like box moving or loading pallets.

Figure have stated that they will be gradually implementing these tasks, one at a time, to test how well it performs. The robot will be able to learn on the job, so it should continue to improve as time goes on.

Amazon have also been exploring how humanoid robots could be used within their warehouses. Just a few months ago, they showed how Agility’s robot was able to work alongside Amazon employees.

Figure’s agreement with BMW follows on from a funding round last May, when they raised over $70 million from investors.

Microsoft’s new tool to compress LLM prompts

There are a few issues with LLMs today, which limit how we can realistically use them. The biggest issue is that they are expensive to use when the prompts become too large.

Another is that they can take too long to return a response for users - which means they can’t be adopted for some use cases. However, this new tool from Microsoft aims to solve some of these issues.

Microsoft’s LLMLingua allows us to compress LLM prompts into smaller versions - which helps to speed up the LLM’s inference and also reduces costs. It’s an impressive feat, given this method is able to make your prompts 20x smaller - and only results in minimal performance losses.

It uses a much smaller language model, such as LLaMA-7B, to identify and remove any non-essential tokens in your prompt. There are two versions available - LLMLingua and LongLLMLingua, which offers better performance for situations where the model needs a longer context.

Google's Virtual Worlds:
Adding 3D Objects via Text

Neural Radiance Fields, commonly known as NeRF, is a technique that allows us to build 3D models from a set of 2D images. It can produce highly realistic renders and Google’s Waymo often use it to improve their self-driving cars - plus it can also be used to add real-world objects into virtual reality (VR).

It’s a cool technology and this research from Google demonstrates how you can easily insert objects into your own 3D environment. You simply select where you want the object to be placed and add a text description.

This research matches well with Google’s ongoing self-driving work at Waymo. Today, their cars will regularly take 2D photos of the world around them, as they continue to drive around the city.

Waymo then uses these photos to create a 3D version of that point in time - which helps to improve their self-driving algorithm, since the virtual world allows them to simulate different weather events such as snow or rain.

Since Google now has the ability to easily add new objects into this virtual world, they can simulate many more scenarios - such as what would happen if a pedestrian was crossing the road, or if traffic cones were blocking the vehicle’s path.

It also comes at a good time, as their rival Cruise has been forced to scale back their self-driving plans and Waymo continues to expand their own program.

📱 Apple has now overtaken Samsung in global smartphone sales, for the first time

💻 Sundar Pichai warns Google staff that more layoffs are coming

🇬🇧 UK Government publish details about their Generative AI experiments, release a new GenAI framework

✂️ CEOs expect job cuts of at least 5% in 2024, due to Generative AI

📈 Reddit plans to launch IPO in March

🕶️ Apple’s Vision Pro will launch with over 150 immersive movies & TV shows

📦 Amazon is testing a new AI chatbot, which can answer questions about products being sold on their store

🎨 You can use AI to remove Samsung’s AI watermark

Magnific

You’re watching a TV show and a police detective is searching through CCTV footage for a criminal. Then they spot something at the edge of the frame, “Wait a second, enhance that image!”.

It’s a common scene in TV shows and movies, but is often very far from reality. While there have been image enhancers for years, they haven’t been very good. That was until recently.

Magnific is a startup that allows you to upscale any image and their tool can add an extraordinary level of detail. It has become a necessary tool for those who regularly use AI image generators.

These generators can produce great images, but they often lack the detail that’s needed to add depth and make it look real. This is where Magnific comes in.

Simply upload your image and the AI tool will do the rest for you. Or, you can even guide the upscaling process with your own prompts. The pricing can seem a little high at $39 per month, but it’s to be expected given the amount of compute that’s needed to add a lot more detail.

If you’ve made an AI image that you really like, but find it still needs that extra bit of magic - their tool might be a good option.

This Week’s Art

We’ve covered a lot in this week’s newsletter:

• Meta’s renewed focus on AI products and AGI

• Anthropic’s research on secretly training AI models to be malicious

• DeepMind’s AI that can solve geometry problems

• OpenAI’s plan to crowdsource governance ideas

• BMW testing humanoid robots at their US factory

• Microsoft’s new tool that can compress LLM prompts

• Google’s NeRF research on 3D objects

• And Magnific’s very impressive image enhancer

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.