🔬 Google DeepMind discovers over 2 million new materials with their latest AI tool

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 Microsoft is looking at using GPT-4 for radiology reports
- Stability AI’s real-time image generator
- Pika’s new video generator
- … and much more

Let's jump in!

Top Stories

1. Google DeepMind creates an AI tool that has discovered 2.2 million new materials [Link]

Almost 400,000 of those materials identified by GNoME are stable and could be created by researchers in the future. 736 new materials have already been created by independent researchers, who’ve looked into the feasibility of the AI model’s results.

The world isn’t short of problems that need to be solved. Semiconductor chips have become smaller and smaller to increase performance & reduce power consumption, but we’re already making chips that are 3 nanometers in size. Further advances will likely become more difficult, but DeepMind suggests these new crystal discoveries could help in this area - along with improved efficiency for electric vehicles (EV), which is important as we continue to see a steady rise in EV sales.

This work follows on from the string of AI announcements from DeepMind in the last few weeks. They’ve detailed AI models that can identify new molecules, make much more accurate weather predictions than what we currently use, and even a model that can generate new music.

2. Microsoft explore how GPT-4 could be used for radiology [Link]

Their researchers have partnered with Nuance, an AI healthcare company that was acquired by Microsoft last year, to study how GPT-4 can process and understand radiology reports. Interestingly, GPT-4 was able to generate summaries of radiology reports that are similar to those written by radiologists.

The model’s reports contained greater consistency in the disease terms that were used, compared to those used by human radiologists - where there is often more variation and this can affect how the report is interpreted by others. However, the possibilities go beyond just how the reports are read by medical staff. GPT-4 could be used to translate the medical terminology into a simpler report that’s easier for patients to understand.

3. ChatGPT shown to leak training data when asked to repeat a phrase [Link]

Researchers from DeepMind, and other organisations, have been able to extract significant amounts of training data from ChatGPT for just $200. They did this by asking ChatGPT to repeat “company company company”. After several hundred instances of the phrase, the model will often regurgitate its training data verbatim - which raises the risk of exposing sensitive information to users.

This is a real flaw with Large Language Models, as they often place too much trust in the user and what they ask for. As companies increasingly look to build custom GPTs, whether they are publicly available or just for internal use, it’s a risk that needs to be guarded against.

4. Stability AI introduce SDXL Turbo [Link]

SDXL Turbo is a new text-to-image generation model from Stability AI. It uses a technique called Adversarial Diffusion Distillation (ADD), which allows them to reduce the image generation process from 50 steps to just 1.

This technique differs from Krea AI's approach, which we covered recently. But similar to Krea’s product, it allows for users to generate images almost instantly. SDXL Turbo isn’t available for commercial use just yet, but it looks like we will soon be able to generate lots of image variations and choose between them in seconds - speeding up the design process for everyone.

5. GM to slash spending at Cruise by hundreds of millions of dollars [Link]

General Motors’ CEO has said that they plan to significantly reduce spending at its self-driving vehicle subsidiary, Cruise, by hundreds of millions of dollars in 2024. It’s expected that there will be significant job losses amongst their 4,000 employees, as a result.

Cruise has seen huge rises in spending over the last few years, with $732 million spent in Q3 of 2023. The company is now scaling back their ambitions for piloting driverless cars across the US, after safety concerns led to their permits being revoked by authorities in California.

Closer Look

What AWS announced at their re:Invent conference

The annual conference is used to reveal their latest cloud tools and, following OpenAI’s disruptions with ChatGPT and GPT-4 in the last year, AWS have been keen to show they’re not far behind in the AI race.

There were several announcements, including the launch of Amazon Neptune Analytics for graph and vector database analysis, AWS Clean Rooms ML for different companies to securely share data and train new models, and Amazon SageMaker HyperPod which will enable more efficient training of LLMs.

AWS are also introducing the Titan Image Generator as a rival to OpenAI’s DALL-E 3, Amazon Q which is a chatbot that’s designed to counter the rise of ChatGPT, and some new Guardrails for Amazon Bedrock to limit how your language model can be used - such as a list of banned topics or phrases.

They also revealed their latest generation of chips, AWS Trainium2 and Graviton4, which are designed to enhance AI model training and inferencing performance. Another major update was to its S3 object storage service with Amazon S3 Express One Zone, which offers significant performance improvements for applications that are data-intensive.

If you want to see the full list of announcements, you can check them out here.

Byte-Sized Extras

• OpenAI’s GPT Store is delayed until 2024 [Link]

• 23andMe says hackers accessed “significant number” of files about their users’ ancestry [Link]

• Elon Musk tells advertisers to “go f*ck yourself”, after large numbers leave the X platform [Link]

• Meta aims to launch Threads in the EU this month [Link]

• Krea AI, a startup making AI tools for image generation, now allow you to upscale & enhance your images [Link]

Startup Spotlight

Pika Labs

The startup is developing AI tools for video generation and editing and has recently raised $55 million in funding. Their first model, Pika 1.0, is available for users who’ve been granted early access. It’s able to generate videos using either text or images that have been uploaded - which can then be produced into different video styles, such as 3D animation, anime, and cinematic mode.

Their biggest competitors are Runway and Stability AI, whose announcements have also been covered in the last few weeks. There has been a huge amount of progress with Generative AI video in the last 6 months alone, with the results being higher quality and slightly more reliable.

Although, Pika’s product seems to address one of the biggest issues with GenAI videos right now - which is producing consistent characters and backgrounds as the camera moves around. If you’ve made a video featuring two objects, for example, you might have found that they will morph into something else when the camera changes position.

This is exciting to see, as more consistent results make it feasible for companies to adopt this technology and experiment with various ideas. Plus, it enables the average person on the internet to create their own short stories - just by typing a few sentences or uploading an image.

If you want to read Pika’s announcement, you can check out their website.

Analysis

It has been a rough few months for Cruise. The self-driving company has faced a series of damaging press stories and some very unhappy residents in San Francisco - with their vehicles often spotted blocking traffic, driving into wet concrete, and has even crashed into a fire truck. But when one of their vehicles drove over a pedestrian, who was initially struck by a human driver, things started to unravel.

US authorities have accused the company of withholding key footage from their investigations and immediately revoked Cruise’s self-driving permits. Their CEO, Kyle Vogt, has since left the company and it’s expected that there will be significant job losses amongst their 4,000 employees. The new leadership’s focus is to rebuild trust with regulators and city residents, but their operations are being dramatically scaled back as a result.

It’s a far cry from what the company’s ambitions were in October, when they announced plans to introduce robotaxis in Japan. When I was at TechCrunch’s Disrupt conference in September, Kyle Vogt told the audience that he expected driverless cars to be given priority in cities - due to their improved safety record when compared to human drivers.

However, Cruise’s reputation has been damaged by months of negative stories and how they responded to the accident with a pedestrian. They will need to become much more transparent with regulators on data sharing after incidents, if they want to rebuild that trust with both authorities and locals.

In the UK, regulators are currently working on plans to introduce self-driving vehicles and are developing a strategy for manufacturers to share data with the authorities. Following the issues faced by Cruise in California, they’ll likely be placing much more emphasis on this within their plans.

This Week’s Art

Prompt: An ultrawide image showcasing a child drawing on the ground with chalk, where their artwork is magically coming to life. The child, deeply engrossed in their activity, is surrounded by their creations which are transforming into a vivid, three-dimensional world. This magical realm includes elements like fantastical creatures, whimsical landscapes, and a variety of vibrant colors, all emerging from the chalk drawings. The scene captures the essence of childhood imagination and creativity, where the boundary between art and reality blurs. The background should be an everyday outdoor setting, enhancing the contrast between the ordinary and the magical.
Platform: DALL-E 3

End Note

It’s been another busy week. We’ve looked at DeepMind’s AI tool that has discovered over 2 million new materials, Stability AI’s SDXL Turbo, Microsoft’s research on using GPT-4 for radiology, ChatGPT leaking its training data, Cruise scaling their operations down, and Pika Labs revealing their new GenAI video generator.

Have a good week!

Liam

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.