🤖 Google unveils Gemini as their rival to GPT-4

Hello,

Welcome to this edition of Loop! We aim to keep you informed about technology advances, without making you feel overwhelmed.

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

In this edition, we’ll explore:

- How Bing is using GPT-4 to enable Deep Search
- Meta & IBM’s open-source AI Alliance
- The EU’s agreement on the AI Act
- … and much more

Let's jump in!

Top Stories

1. Agreement reached on the upcoming EU AI Act [Link]

The landmark agreement will set out a new framework for AI applications, using a risk-based approach. The EU aims to strike a balance between innovation and ensuring that AI systems are used in an ethical way - with full enforcement expected by 2026.

For example, companies will be prohibited from using a person’s behaviour or characteristic to assign social scores. Using systems that analyse a person’s emotions, in a workplace or education setting, is also banned. However, systems that will be used only by the military or security services will not be affected by the law.

Lawmakers have also reached agreement on what a “high risk system” is and where foundation models will fit within that approach. Any companies who are making high risk systems will need to maintain documentation to demonstrate their compliance - including the datasets used, training methodologies, and how the system is being monitored. Those who don’t comply will face heavy fines, similar to GDPR, which can be as much as €35 million or 7% of global turnover.

2. Meta launches their own AI image generator [Link]

"Imagine with Meta” aims to compete with DALL-E, Midjourney, and Stable Diffusion. It’s powered by Meta’s Emu model and is currently free for users in the US. Any images that are made using the tool will contain an invisible watermark, so that others can identify it as AI content.

For police forces around the world, who often rely on evidence submitted by the public, this is a real concern. It’s possible that AI generated images and videos could be used to mislead investigators. Meta’s announcement follows on from work by Adobe, where they already embed watermarks into images that are created using Firefly.

3. Meta and IBM form an AI Alliance to promote open-source development [Link]

Meta and IBM have created a new AI Alliance, which advocates for an "open-science" approach to AI development - positioning themselves against Google, Microsoft, and OpenAI. This new group, which includes major tech companies and universities, emphasises the importance of scientific exchange and open-source technologies in AI's future.

They currently have 50 members, which include AMD, Intel, UC Berkeley, Yale, Stability AI, and Dell. Their approach is at odds with OpenAI and Google, who instead prefer a more proprietary and closed AI development model.

There’s a growing debate over the open-source development of AI models, with a lot of concerns about safety and how the technology could be misused by bad actors - especially in regards to disinformation campaigns. Following the recent executive order by President Biden and with the European Union making progress with their AI Act, it’s likely the safety debate will only intensify.

4. Bing is getting a new feature called Deep Search [Link]

Deep Search is using GPT-4 to add more detail to the user's query, which can lead to more relevant and detailed answers. We already see this with DALL-E 3, since it will refine a very simple prompt, like “spaceman on the moon, realistic” to a more complete prompt like “A spaceman is walking on the moon. Behind them is the American flag and they look up at the stars”.

There’s lots of ways Deep Search could be used. For example, a question about loyalty programs in Japan would be expanded to cover the different types of programs, their benefits, requirements, and then include a comparison with other payment methods. By expanding on what the user asked, Bing will have a greater understanding of what the user is looking for and can therefore provide better results.

It will take longer for Bing to generate a response, but it’s useful for those who need more comprehensive answers - rather than quick results.

5. Microsoft want to use MatterGen to discover new materials [Link]

Following in DeepMind’s footsteps and their announcement of GNoME last week, Microsoft’s Research team has released details about MatterGen. Interestingly, this is a new diffusion model that can use specific criteria to then generate novel materials.

Similar to DeepMind, the research team has proposed that this new model could help drive progress in a number of areas - such as energy storage and carbon capture.

Announcement

Google finally unveils Gemini as their rival to GPT-4

Google's DeepMind team has recently introduced Gemini, a new family of AI models that has three versions: Ultra, Pro, and Nano - with each of these models tailored for different applications.

Gemini Ultra is their flagship and most powerful model, Gemini Pro is a lighter version that’s similar to OpenAI’s GPT-3.5, and Gemini Nano is designed for mobile devices - with there being two further variants for low and high-memory devices.

In contrast to OpenAI’s models, Google claims that Gemini is “natively multimodal” - meaning it’s been pre-trained and fine-tuned on text, audio, images, and videos - rather than just text on its own. In some benchmarks, Google has suggested that Gemini can outperform OpenAI’s GPT-4, but that remains to be seen in the real-world.

The downside is that Gemini Ultra won’t be released to the public until “early next year”, although Gemini Pro is already being used to power Bard - which is Google’s competitor to ChatGPT. The Pro version will soon be available for enterprise customers via Vertex AI and for developers via Google’s Generative AI Studio. Gemini Nano will feature on the Pixel 8 Pro and offer features like summarising voice recordings, or suggesting replies within their messaging app.

You can read more about Gemini on their website.

Byte-Sized Extras

• Cruise faces fines for withholding video footage of their robotaxi accident [Link]

• Anthropic’s researchers show that phrases like “really really really“ can improve a model’s results - e.g. “We really really really don’t want to discriminate users” [Link]

• Autonomous plane completes a 12 minute flight, without a human onboard [Link]

• Google create a library that can convert Python code to LaTeX expressions [Link]

• Runway, who develop AI video generators, are partnering with Getty Images to build AI tools for enterprise [Link]

• Google’s AI note-taking app, NotebookLM, is now available to users in the US [Link]

Startup Spotlight

Liquid AI

This is a promising start-up that has spun out of MIT and aims to commercialise a new architecture called Liquid Neural Networks (LNN). This is a more flexible architecture that’s smaller in size, meaning it needs significantly less compute power to run.

Since far fewer parameters are needed to handle tasks, such as text generation or coordinating autonomous vehicles, it could be run on low-powered devices - such as a Raspberry Pi.

Earlier this year, the MIT researchers showed how LNNs can enable drones to autonomously navigate around an unfamiliar environment. There’s a lot of potential use cases for autonomous drones, such as search and rescue operations, monitoring wildlife, and delivering packages.

The company believes LNNs are particularly good at analysing fluctuating phenomena - such as electric power grids, medical readouts, financial transactions, or changing weather. They’ve recently raised $37.6 million in funding and have offices in both Boston & Palo Alto.

If you want to read more, you can check out their website.

Analysis

After much fanfare, Google has finally given us a glimpse into what Gemini can do. It’ll be interesting to see how it compares to GPT-4, especially when it comes to multi-modal applications.

Following the difficult launch of their Bard chatbot earlier this year, which immediately hurt the company’s share price, Google was keen to get this one right. It’s clear a lot more work was put into what was possible with their new AI model.

Although, it may be the case that they’ve leaned too much into the future possibilities. The company has come under fire for a video that seemed to showcase Gemini’s multi-modal capabilities - where it could spot that the user was drawing a picture of a duck, or that they were playing a game of Rock Paper Scissors.

But the issue with the video is that while Gemini can do those things, more prompting and hints are needed from the user to get those results. While Google did say that the video was edited and some scenes were shortened, this could have been made a bit clearer.

Regardless, it was a small mistake in what was otherwise a pretty good launch. Much has been said that Google is trailing behind the competition, specifically OpenAI. That was true at the start of 2023, when they didn’t have many commercial applications that had Generative AI, but isn’t true anymore.

In fact, Google is well positioned when we look at the much wider AI sector. DeepMind has been releasing new AI advances on a very regular basis - from accelerating drug discovery, to identifying new materials, to training robots that can do household chores. Google is in a much better position than is currently being portrayed.

This Week’s Art

Prompt: A cinematic high-definition image in a 21:9 aspect ratio, depicting a scene from an action film with 'Winter' coloring. Two men are walking in the Arctic during the day. The setting is vast and open, with a clear blue sky and a snowy landscape stretching far into the horizon. The men are dressed in heavy winter gear, suitable for Arctic conditions. They look determined, suggesting a sense of purpose in their journey. The image captures the raw, harsh beauty of the Arctic environment, emphasizing the challenge of their trek.
Platform: DALL-E 3

End Note

While Gemini stole the show this week, we’ve also looked at the EU’s agreement on their upcoming AI Act, Meta’s AI image generator, the open-source AI Alliance, Bing’s deep search feature, MatterGen (sorry if you’ve got a bit of déjà vu, following DeepMind’s announcement last week), and how Liquid AI are looking to commercialise Liquid Neural Networks.

Have a good week!

Liam

Feedback

Share with Others

If you found something interesting, 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.