🛢️ DeepMind proves that synthetic data is the new oil

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 ‏‏‎ ‎

• OpenAI’s first step towards AI agent conversations

• Microsoft’s use of Laser to remove AI inaccuracies

• How Apple’s GenAI strategy differs from their competitors

• … and much more

Let's jump in!

1. Microsoft’s Laser project can remove LLM inaccuracies

We start the week with Microsoft’s new project, which has a surprisingly cool name - not something you usually find with AI research.

Layer-Selective Rank Reduction (LASER) is a technique that allows us to improve the accuracy of Large Language Models (LLMs), by removing parts of it that don’t work.

When a weight within the model isn’t performing well, LASER will carefully replace it with a smaller version.

Interestingly, using the smaller weights doesn’t lead to a drop in performance. Instead, the performance can increase by as much as 30%.

Microsoft were able to improve GPT-J’s accuracy at identifying genders - with a huge jump from 70.9% to 97.5% accuracy.

It’s possible that this technique could help us to reduce hallucinations (when a model makes something up) and error rates in LLMs, which are the biggest issues with the technology today.

2. ChatGPT now lets you add other GPTs into your conversation

It’s the first step by OpenAI towards AI agents, which Sam Altman has previously hinted about. With this new feature, paid ChatGPT users can use several GPTs within the one chat.

All you have to do is type “@” and then select the GPT you want to add. The custom GPT bot will be able to understand the previous messages and contribute to the conversation.

This is quite a simple feature, but it gives a glimpse into a future where we expand beyond just one GPT. We are looking at the possibility of multiple GPTs talking to each other.

One GPT is fine, but can be frustrating. We have all faced a situation where you have to continuously ask ChatGPT to change its responses - until you finally get the result you wanted.

By having multiple GPTs, you can ask them to review each other’s work and suggest improvements. Microsoft have made quite a bit of progress in this regard, with AutoGen and TaskWeaver able to coordinate multiple AI agents.

I’ve tried out AutoGen for some time now and the results are very impressive. While OpenAI haven’t officially supported AI agents just yet, it’s clear that is where they are heading.

3. Apple will reveal their GenAI products later this year

In an earnings call, Apple’s CEO has said that they will showcase some of their generative AI research “later this year”. While there wasn’t a specific date, Apple's annual WWDC event in June would be a good opportunity to do so.

In contrast to the other big tech companies, Apple has focused on secretly working on their projects before they’re announced to the public. This silence and aura of secrecy is characteristic of Apple and their working culture.

They’ve also taken a slightly different approach when it comes to how they release their AI tools. Companies like OpenAI and Microsoft have released lots of standalone tools as paid “add ons”, whereas Apple has sought to integrate them into their existing products.

4. FCC plans to ban robocalls that use AI-generated voices

It follows on from a campaign that aimed to suppress votes for President Joe Biden in New Hampshire, with voters urged to stay at home. An AI voice was used to impersonate the President and misleadingly said that if they voted it would help the Republicans.

The robocall has been condemned by both state and federal officials, saying that it aimed to undermine the democratic process. If the FCC succeeds, they will have the legal powers they need to “go after bad actors” behind these types of campaigns. It’s quite a surprise that they don’t have this ability already.

5. Scammers claimed they had hacked Europcar, turns out they used AI to generate the fake data

One user made a post on a well-known hacking forum, claiming that they had access to data about 48 million Europcar customers and wanted to sell it. As per usual, they provided a sample of the data for others to see.

However, it has since been revealed that this data was fabricated - likely with the help of an AI tool, similar to ChatGPT. Europcar found that there were several inconsistencies with the sample data, such as incorrect record numbers, mismatched personal details, and email addresses that don’t exist in their customer database.

Of course, Europcar aren’t an independent party in this scenario. But Troy Hunt, a renowned security expert, has also reached the same conclusion - with fake home addresses seen in the data.

It’s an interesting twist and we could see this happen a lot more often - which wouldn’t be a bad thing. If criminals want to avoid buying junk data, it seems like they will need to do a lot more homework.

Apple’s Vision Pro is a clear step forward

This is the company’s first product launch since the Apple Watch back in 2015. It’s a headset that offers a mix between virtual and augmented reality, which Apple have coined as “spatial computing”.

Initial reviews have been overwhelmingly positive, with many praising the superb user experience and technical advances made by Apple. Some have been critical of the current limitations, such as the need for an external battery pack.

Ultimately, the Vision Pro is a first generation product and we shouldn’t be overly concerned about what the limitations are today.

The first iPhone didn’t have an App Store, 3G, and the battery life was poor - but it quickly overthrew Blackberry and changed how we interact with technology.

The first Apple Watch was too thick, quite expensive for a digital watch and had a battery that barely lasted a day - but in just a few years, it has completely dominated the market.

Instead, we should view the Vision Pro as Apple’s first step in re-defining how we interact with computers. Over time, the cost will decrease as they figure out new ways to improve the design and adjust the number of features.

What Apple aims to do is to create a solid platform and have developers create brand new app experiences, before it makes the technology more accessible for others. This is something that Meta’s Oculus has been trying to do for 10 years, but has struggled to make progress.

In this area Apple has the upper hand, since developers who already have iPad apps can easily port them over to the Vision Pro. To really drive the new VisionOS platform, Apple desperately needs a few killer apps. And to do that, they need to get this technology in the hands of developers early on - even if the cost prohibits mainstream adoption in the short term.

We haven’t seen a new product launch like this for quite a few years, so it’ll be interesting to see what new experiences are made for the Vision Pro.

Add photorealistic objects to any video

Given the constant headlines about autonomous vehicles driving into cement or blocking traffic, it’s clear that there are far too many scenarios for these manufacturers to account for.

Self-driving vehicles use a vast array of sensors and video cameras to capture the world around them. These are then used to re-create virtual worlds, which they use to further improve the system.

The downside is that you need to capture real-world data about all these scenarios, which is tedious and costly. This is where synthetic data comes in. Since companies can’t always capture the data they need, they can try to make it instead.

This project, called Anything in Any Video, allows you to seamlessly insert any object into a video and make it look photorealistic. There’s a lot of potential for how this could be used - such as adding objects into VR environments, simple video editing, or to create new videos for training autonomous vehicles.

For example, you could create test videos to determine how the system would react when a child’s bicycle was visible on the road - without having to actually create that in the real-world.

DeepMind have recently used synthetic data to train their new AlphaGeometry model, which can solve geometry problems at a gold standard.

Since it’s difficult for big tech companies to capture the data they need for their new AI tools, synthetic data will become much more important going forward - especially as shareholders want to see quicker financial returns for AI projects.

🧑‍💻 Meta releases a bigger version of their Code Llama model

🇪🇺 EU’s AI Act passes the last hurdle

🇺🇸 US disrupts a hacking operation backed by China, warns of the threat to American infrastructure

💰 Elon Musk’s $56 billion pay deal is unfair, a judge has ruled

📉 23andme’s valuation falls from high of $6 billion to nearly $0 - a 98% collapse since 2021

🤖 Amazon terminates their acquisition of iRobot

⚖️ ChatGPT is violating European privacy laws, says Italy’s DPA

Kore.ai

This is a startup that’s developing conversational and GenAI products for enterprise. They’ve recently raised $150 million in funding to boost their product offering further.

The company has a no-code platform, which allows businesses to use AI within customer and employee interactions. It’s a crowded market right now, but Kore.ai seems to be focusing on no-code and making it simple to integrate different AI models.

As you can imagine, they make use of Large Language Models for some of these services - such as their search tool, which allows you to upload a document and ask questions about it.

This is a pretty common use case for businesses and saw a huge amount of growth in the latter half of 2023, as more companies started to experiment with generative AI.

Kore.ai are already working with some pretty big names - such as Yahoo, Airbus, and AT&T - and this latest funding round has raised the total amount to around $223 million, which will be useful in scaling their headcount even more.

This Week’s Art

This week was dominated by Apple’s launch of the Vision Pro, which aims to set the foundation for lighter and much cheaper devices.

“Spatial computing” might be a clever marketing strategy from Apple to differentiate their headset with existing AR/VR devices, but it outlines how ambitious their plans are.

We’ve covered a lot this week, such as:

• Microsoft using Laser to eliminate inaccuracies in LLMs

• OpenAI’s first steps towards AI agents

• Apple’s GenAI plans and how they differ from their rivals

• FCC’s plan to ban malicious robocalls

• Scammers using AI to generate the fake data

• How Apple’s Vision Pro could change the way we interact with technology

• Adding photorealistic objects to any video and the rise of synthetic data

• And Kore.ai’s no-code platform for AI

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.