Why Meta is creating a robot hand that can “feel”

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 ‏‏‎ ‎

• Wiz’s CEO faces a deepfake attack

• OpenAI’s transcription tool could be causing medical diagnoses

• How Wonder Dynamics are using AI to accelerate the VFX process

• … and much more

Let's jump in!

1. Wiz CEO was targeted with a deepfake attack

The cybersecurity company recently had to deal with a deepfake attack, which targeted dozens of employees.

Each person was sent a voice message from a hacker, who was impersonating their CEO and asked for login details.

Fortunately, no one fell victim to the attack - as there was a noticeable difference between the fake audio and Assaf Rappaport’s real voice.

While Wiz was able to trace the origin of the voice, they weren’t able to identify who sent it.

2. LinkedIn launches its first AI agent for recruitment

LinkedIn has unveiled its latest AI tool, called Hiring Assistant, which is designed to help recruiters.

They can use the tool to create job descriptions, identify potential candidates, and then message them.

The AI assistant is only available for a select group of customers, but a wider roll-out is expected in the coming months.

This builds on LinkedIn’s long list of AI-powered tools - including its learning coach that suggests ways to improve your performance at work, assistants that can improve your marketing campaigns, and its powerful recommendation algorithm.

Job recruiters are a huge source of income for LinkedIn. Just last year, it saw revenues grow to over $7 billion. The company believes that AI is key to future growth, with a focus on both businesses and consumers.

If the new Hiring Assistant agent works well, it could help to boost those revenues even further.

3. OpenAI’s transcription tool has major hallucination issues, say researchers

Imagine that you visit your doctor and describe a problem. The doctor is using an AI tool to listen to the conversation and summarise what was said.

But the model goes off-course and claims that you violently attacked someone. Medical staff later review the notes and call the police.

It might sound far-fetched, but it really isn’t. Researchers have found that OpenAI’s Whisper model can occasionally make things up and cause real harm.

The Whisper voice model powers many of OpenAI’s features and is used by businesses around the world.

One of those companies is Nabla. Thousands of doctors use their software to transcribe conversations they have with patients and summarise their notes.

But these hallucinations are very worrying. In some cases, Whisper claimed that patients talked about violent assaults - which wasn’t true at all.

If you’re falsely accused of a crime, or medical staff give you the wrong treatment, it could lead to huge legal implications for Nabla - and the same is true for many other companies that use Whisper for critical use cases.

4. Nvidia makes progress with quantum computer research

Nvidia has partnered with Quantum Machines, whose research has shown that Machine Learning can help us to control qubit behaviour.

Qubits are core to how quantum computers work, however they are highly sensitive and can exist in multiple states at once.

To manipulate these qubits, microwave pulses are used. But this is really demanding and requires a lot of computing power - especially when you try to do this in real-time.

Interestingly, a Machine Learning technique can be used to optimise this process - which is called reinforcement learning.

The team found that even a small change can lead to exponential improvements in error correction, which is essential for unlocking the benefits of quantum computing.

It’s still early days, but this is really encouraging progress. The two companies will continue their research into next year and plan to use even more computing power for their tests.

5. OpenAI launches ChatGPT Search to challenge Google

The new feature can quickly answer questions and places more emphasis on online sources, rather than ChatGPT’s training data.

It’s powered by a fine-tuned version of GPT-4o, which is able to provide the sources used in its response.

The new feature will initially be available to ChatGPT Plus and Team users, with plans to expand it more widely in the coming months.

If you use Chrome, you can set ChatGPT Search as your default search engine. To do this, just use their new browser extension.

This is a direct challenger to both Google and Perplexity. But the rise of AI search is worrying news publishers, as they fear that it will reduce their web traffic and ad revenues.

Meta is creating a robot hand that can “feel”

Meta’s AI research team is working to develop a robotic hand that can actually “feel”. The tech giant is working with GelSight and Wonik Robotics on the project.

Ultimately, they want the robot to be able to identify pressure, texture, and movement - without requiring a large database - similar to how we can feel and describe an object in the dark without knowing what it is.

This is because tech companies are reaching the limits of what data can be collected. They can’t keep scaling-up indefinitely.

GelSight is currently working on a fingertip sensor that can “feel” the difference between materials. This is necessary for future robotics systems, as they need to apply the correct amount of pressure when lifting an object - especially if it’s fragile.

It’s impossible to specify this for every single object. Instead, the robots need to be able to determine this on their own.

While Meta has completed a lot of robotics work in recent years, this is pretty out there. And pretty ambitious.

Previously, their work has focused on training robots and ensuring that they can detect objects. They’re now aiming far beyond that.

D-ID reveals avatars that can have real-time conversations

The company provides AI video avatars for businesses, with many combining these avatars with AI models that can respond to user questions.

There are lots of companies doing this, including HeyGen and Synthesia. But D-ID’s new avatars are incredibly realistic and can have real-time conversations with people.

There are two new options. You can now create an avatar that looks just like you, or another person. To do this, you only need to upload one minute of video. D-ID’s avatar will then be able to replicate your head movements.

Another option can be used for more accurate results, which is called the Premium+ model. It does require several minutes of video, but it creates an AI avatar that has hands and a torso - making it seem more human-like.

D-ID have been pitching this towards enterprise customers. They hope that this will be used for engaging customers and marketing campaigns.

But there’s a real risk that the tool could be used to impersonate people.

If you can create highly-realistic videos for a marketing campaign, then why can’t hackers do the same and use it to deceive people?

With so many startups racing to develop AI avatars and make them look realistic, there’s a risk that these concerns could become true.

🔧 OpenAI will start using AMD chips and could make its own AI hardware

🤖 MIT debuts a new way to teach robots and improve their skills

🚀 SpaceX wants to refuel their Starships in space, from next year

💬 Chinese military created a defense chatbot with Meta’s AI

💻 Apple upgrades the MacBook Pro with M4 chips

🚛 Aurora Innovation delays launch of autonomous trucks to 2025

⚡ GitHub Spark lets you use English to build web apps

🎙️ Meta releases an open version of Google's podcast generator

🚕 Waymo could use Google's Gemini to train its robotaxis

📉 Intel's AI chips are far behind Nvidia and AMD, fails to hit $500 million goal

Wonder Dynamics

Wonder Dynamics is a VFX startup that uses AI to create animated characters.

This benefits film studios and directors, as they can quickly see how a shot might look in post-production.

Currently, they have to wait for the CGI work to be completed - which is time consuming and expensive.

Any re-shoots will dramatically increase costs, so it’s important that they get it right the first time.

As you might expect, the company is based in LA, as they hope that filmmakers embrace their tools and experiment with different shots.

The company emerged in 2021 and has prioritised professional-use, not ordinary users.

They have stressed that the tool isn’t designed to create production-level movies. Instead, it should be used within the wider filmmaking process.

They’ve gained a lot of attention with their tool, which is called Wonder Studio.

So much, that Autodesk has recently acquired it. There aren’t any details on the numbers involved, but you can find out more about Wonder Studio below.

This Week’s Art

We’ve covered quite a bit this week, including:

• The deepfake attack that targeted Wiz’s CEO

• LinkedIn’s first AI agent for recruitment

• The worrying problems with OpenAI’s transcription tool

• Nvidia’s progress with quantum computer research

• OpenAI’s ChatGPT Search and how it’s different to the chatbot

• Why Meta is creating a robot hand that can “feel”

• D-ID’s new avatars that can have real-time conversations

• And how Wonder Dynamics are using AI to accelerate the VFX process

Have a good week!

Liam

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 AI Engineer and works within Kainos' Innovation team. He identifies business value in emerging technologies, implements them, and then shares these insights with others.