Hello, Starters!

The current of AI innovations doesn't stop, and as one gets delayed, there's another company taking advantage of the circumstance to come up with their own. But in the end, who will be the one to drive advancements? We'll see!

Here’s what you’ll find today:

• Kyutai introduces its AI assistant Moshi

• Google’s new method for image description

• Is ChatGPT really good at coding?

• ElevenLabs unveils an AI voice isolator

• Apple Intelligence is coming

• And more.

🤖 Kyutai introduces its AI assistant Moshi (2 min)

After announcing its Voice Mode feature, OpenAI decided to delay its launch, allowing French startup Kyutai to have a head start. They've developed Moshi, an AI assistant based on an "Audio Language Model" that can have natural conversations with users in real time.

Moshi has some very interesting qualities to offer, starting with the fact that it was developed in just six months. This model works with audio data in a way that makes it natively a multimodal model, and most importantly, Kyutai plans to release it as open source to further drive more advancements.

🖼️ Google’s new method for image description (5 min)

A team of researchers at Google has come up with a method to enhance how AI image generation systems work, and they've called it ImageInWords. This approach mixes human knowledge with AI to improve the descriptions given to image models, resulting in more accurate outcomes.

Most AI systems are trained on content found on the web, which is often unreliable and inaccurate. As a result, image models trained with this data tend to generate biassed descriptions. With ImageInWords, this changes as an AI recognizes individual objects in an image, followed by a human annotator refining and using it to fine-tune the models.

🤔 Is ChatGPT really good at coding? (3 min)

ChatGPT's coding capabilities are always in the spotlight, sometimes by people praising it and claiming how it could potentially replace programmers, and other times by people saying just how bad it is. But what is the truth? Is ChatGPT as good as a human programmer or is there still a lot of improvement to do? A recent study sheds light on this approach.

The results highlight that ChatGPT is indeed able to produce functional code; however, there are variables such as the difficulty of the task, programming language, and more. Also, researchers noticed that its potential decreases when facing algorithm problems that go beyond its last training date. So basically, any problem found on the web after 2021 will be harder for the model.

🍎If you're looking forward to Apple's upcoming enhancements to Siri and its whole Apple Intelligence revamp, you'll need to be a little patient, as the major changes are expected to come in spring 2025 with iOS 18.4, according to Bloomberg's Mark Gurman. However, ChatGPT's integration is confirmed to arrive this fall.

🎙️ElevenLabs doesn't stop shaking the industry with its AI voice models, and its latest offering is a tool that will change creators' lives. Its new AI voice isolator allows users to remove background noise, improving the voice of the speaker and removing interferences that may damage the final product's sound.

⚡️Quick links

OpenAI breach is a reminder that AI companies are treasure troves for hackers

Whiteboard of Thought: New method allows GPT-4o to reason with images

Tool preventing AI mimicry cracked; artists wonder what’s next

Tokens are a big reason today’s generative AI falls short

Samsung shares hit over 3-year high after better-than-expected guidance on strong AI demand

Thank you for reading!