🌍 UC Berkeley's Large World Model has outperformed GPT-4 Vision

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 ‏‏‎ ‎

• UC Berkeley’s impressive AI model that can analyse videos

• Apple’s move to stop quantum computers from breaking encryption

• Google’s latest struggles with Gemini image generation

• … and much more

Let's jump in!

1. Microsoft and Intel sign a custom chip deal worth billions

Intel has announced a significant partnership with Microsoft, which is valued at over $15 billion. This means that Intel will supply custom chips that have been designed in-house by Microsoft’s engineers.

The chips will be manufactured using Intel’s 18A process and is part of their wider strategy to regain their position in the market.

The deal follows in the footsteps of Apple and TSMC, who have worked on custom chip designs for over the last decade.

It has been hugely successful for both parties, with Apple cementing their ecosystem advantage over their rivals and TSMC becoming a world leader in semiconductor manufacturing.

Microsoft and Intel are hoping that their new deal will have the same impact.

2. Apple prepares for when quantum computers could break encryption

Quantum computing poses a real threat to how we secure our data, as the current standards could be solved by much superior computers.

This is a real concern for everyone, but especially for government agencies. Laura Thomas, a former CIA officer, recently spoke about this issue at a conference in San Francisco.

Nation states are currently collecting all the data they can, whether it’s encrypted or not, with the aim to decrypt it several years from now. For surveillance agencies, this is worrying as their agents and methods could eventually be identified.

Instead of using public and private keys to encrypt/decrypt data, Apple makes the case that encryption keys will have to change “on an ongoing basis”. Their new PQ3 standard will be applied to all iMessage conversations and older versions.

3. Regulators pause Waymo’s application to expand across California

Waymo had applied to bring their self-driving vehicles to Los Angeles and San Mateo, but the plan has since been put on hold by California regulators for another 4 months.

The company already operates their robotaxi service in San Francisco, but lost one of their vehicles after it was set on fire by locals last week.

While cities in California were amongst the first to trial autonomous vehicles, there has been growing frustration with the companies running them.

Local media often reports on vehicles that block traffic, ambulances, or drive even into wet concrete - and Waymo’s rival Cruise was suspended for lying to regulators late last year.

4. Hundreds of AI figures call for deepfake regulation

More than 500 AI figures, including Yoshua Bengio and Gary Marcus, have called for strict regulation against deepfakes.

Their letter raises concerns that “deepfakes are a growing threat to society, and governments must impose obligations throughout the supply chain to stop the proliferation of deepfakes.”

Of course, we have known about the risks associated with deepfakes for years. What’s particularly concerning is how it could be to create content about children, which is specifically raised in this letter.

While last year’s letter on AI development didn’t lead to any meaningful change, we can only hope that this one is different.

5. Bioptimus raises $35 million to develop AI models for biology

The French startup has attracted quite a lot of attention, following their latest round of funding. They aim to train new models which are specifically focused on biology use cases, such as diagnosis, molecule creation, and precision medicine.

AWS will provide the company with the compute they need to train these models, while Owkin - a larger biotech company - will support them with staff and office space.

Owkin are a key investor in Bioptimus and will get early access to their models, which can be integrated into their own products.

It’ll be interesting to see what Bioptimus can come up with. General-task models, such as OpenAI’s GPT, can’t be relied on for providing an accurate diagnosis.

They also face potential headaches with the upcoming EU AI Act. While it would never stop their development work on these models, they would likely be classed as “high risk” due to the range of medical use cases and face greater scrutiny from officials.

Bioptimus will need to clearly show how they’re minimising the risk of harm to patients, once these models have been made and are actually being used.

UC Berkeley releases their impressive Large World Model

The university’s researchers have open sourced their model and made it freely available for others to use.

They were able to achieve some impressive results, which now raises the possibility that we can use vision models to analyse TV shows, movies, and other long form content.

What really stands out is the model’s recall abilities. After it processed a 1 hour compilation video, which contained over 500 different clips, it could accurately answer questions about what was seen.

For example, when asked “What colour jacket was the girl on the trampoline wearing?”, the model correctly identified it was a blue jacket. Whereas, Gemini and GPT-4 weren’t able to do this.

This opens up a new world of possibilities. We can now use models like these to process more information than ever before.

If you want to analyse what being shown in your competitors’ advertisements, you would normally have to do so with a lot of manual effort. With these models, that might no longer be necessary.

One thing I’m interested in is how streaming companies use these models. As I’ve stated before, they’re under immense pressure to retain subscribers and continue growing.

That’s incredibly difficult in a highly saturated market, so companies like Netflix have started offering gaming services as a way to keep audiences onboard.

With AI video models, they could add new features that allow you to ask questions about a TV show or movie - such as why a character made a decision, or what happened in the last episode.

Google admits they “got it wrong” with their latest image generator

It was a difficult week for Google, who’s image generator came under fire for injecting diversity into images - especially those with a historical context.

For example, when asked to generate an image of the US founding fathers, Gemini would return images that only featured people of colour.

In some instances, users shared how the AI model would refuse to generate some people - but be more willing to show others.

This is because of the way the model was trained in the first place, with Google not including enough examples of people it shouldn’t generate.

It’s also down to the model becoming more cautious over time, which led it to reject images of some ethnic groups.

While it’s pretty amusing for us to see an AI make such obvious mistakes, it does raise questions about how in-depth it was tested by Google.

They have been under considerable pressure following OpenAI’s string of product announcements and their close partnership with Microsoft, but this should have been spotted much earlier.

In many ways, this was completely inevitable. Following the meteoric rise of ChatGPT, tech companies have jumped head first into generative AI and aim to create new revenue streams. Google certainly won’t be the last company to make a mistake like this.

💰 Nvidia becomes a $2 trillion company, after it reached $1 trillion last May

📅 AI Pin shipment date is pushed back to mid-April

🚧 Google DeepMind forms a new team focused on AI safety

💻 Leaked files from Chinese firm show a huge international hacking campaign

💵 Reddit files to go public, has already made $203 million from licensing its data

📉 Lucid Motors will only build 9,000 EVs in 2024, previously predicted it would ship 90,000

🇨🇳 China’s Moonshot AI is raising $1B to develop a LLM focused on long context

🤖 Mistral’s models will soon be available on AWS Bedrock

LetsData

Last year, I met with this startup’s co-founder at a conference in San Francisco and was impressed by what they’ve already achieved.

Their platform is used to protect governments and companies against disinformation campaigns, which have increased rapidly in recent years and will only continue to grow.

For businesses, their tools can be used to monitor the web and news articles about what others are reporting - then alert you when there is a PR crisis.

They’re already working with the UK Foreign Office and several organisations in Ukraine, with a focus on flagging disinformation that’s being produced by Russia.

Just last week, they met with Google’s CEO Sundar Pichai and pitched their company at the Munich Security Conference. If you want to learn more, I’ve included a link to their website.

This Week’s Art

There’s been a lot of covered this week, including:

• Microsoft and Intel’s custom chip deal

• Apple’s concern over quantum computers being able to break encryption

• Waymo’s application being paused by regulators

• AI figures call for deepfake regulation

• Bioptimus raising $35 million to develop AI models for biology

• UC Berkeley’s impressive video model

• Google’s struggles with Gemini image generation

• How LetsData are working to detect disinformation campaigns

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.