Uncover the Voice AI No One Expected: Moshi AI Tested

Amidst the hype around OpenAI's GPT-4 Voice Assistant, a new player has emerged - Moshi AI, an open-source voice assistant developed by a French company, Cute AI Labs. This web-based interface promises low-latency interactions and emotional awareness in its voice.

The key highlights of Moshi AI are:

• Low Latency: Moshi AI aims to provide a real-time, interrupt-driven voice experience, unlike the delayed responses often associated with voice assistants.

• Emotional Awareness: The assistant claims to have the ability to detect and respond to the user's emotional tone, though this feature was not consistently demonstrated during testing.

• Open-Source: Cute AI Labs plans to open-source the code for Moshi AI, allowing developers to integrate it into their own applications.

While the initial testing revealed some limitations in the assistant's capabilities, such as inconsistent voice modulation and emotion detection, the open-source nature of Moshi AI suggests that it has the potential to improve over time as the community contributes to its development.

The availability of Moshi AI, alongside the anticipated release of GPT-4's voice features, signals an exciting time in the world of conversational AI. As these technologies continue to evolve, users can expect to see more innovative and accessible voice assistants emerge, potentially transforming the way we interact with technology.

Just 7 years ago, the state-of-the-art in AI image generation was barely recognizable. Now, the latest video generation model, GenFree, has taken a massive leap forward, producing stunningly realistic and creative visuals.

While GenFree is an impressive achievement, it's important to understand its capabilities and limitations. The model excels at generating high-quality cinematic shots, such as drone footage of a lighthouse, thanks to its extensive training on relevant visual data. However, it struggles with more specific requests, like creating an otter surfing a wave, due to the lack of such niche examples in its training.

One of the key strengths of GenFree is its ability to blend different artistic styles. By prompting the model with a request to create a scene in the style of the Dutch painter Hieronymus Bosch, the results are a captivating blend of medieval fantasy and modern GTA-style cinematography.

That said, the cost of using GenFree can add up quickly. Each 10-second generation costs $1 in credits, and achieving the desired result often requires numerous iterations. This can make the tool prohibitively expensive for casual experimentation.

Despite these limitations, the potential of GenFree is undeniable. As the model continues to improve and the cost of usage decreases, we can expect to see more and more real-world applications, such as the Motorola ad campaign that utilized AI-generated video.

In the meantime, the most rewarding use of GenFree may be in the realm of creative exploration and experimentation. Prompting the model with unusual combinations of elements, like a "dance party full of cats with hats," can lead to delightfully bizarre and imaginative results.

As the field of AI video generation rapidly evolves, it's an exciting time to explore the capabilities and limitations of these cutting-edge tools. With patience and a willingness to experiment, users can unlock the full potential of GenFree and push the boundaries of what's possible in the world of AI-powered visuals.

11 Labs has been shipping new features for its voice assistant platform. The key updates include:

1. Iconic Voices: The 11 Labs reader app now allows users in the US, UK, and Canada to have their text read out by iconic voices like James Dean or Bert Lahr. This adds a fun and unique touch to the text-to-speech experience.

2. Audio Isolation: 11 Labs has released a new AI tool that can isolate voices from noisy audio. This allows users to remove background noise and get crystal clear audio, similar to features released by other companies earlier this year.

3. Mobile App for AI Music Generation: Sooner, a music generation AI, has released a mobile app that allows users to generate AI music on the go. However, this is currently limited to iOS users in the US, with an Android version and global rollout planned once multilingual capabilities are integrated.

These updates from 11 Labs and Sooner demonstrate the continued advancement and expansion of voice assistant and audio manipulation capabilities powered by AI. While the iconic voices and audio isolation features are immediately useful, the mobile music generation app points to the potential for AI-driven creativity on the go.

On a quick side note, Luma AI also released a brand new feature called Luma Keyframes. This feature allows you to transform one thing into another, creating smooth transitions with AI video.

We decided to put this feature to the test, using the Mid-Journey generated images from our Star Wars-themed team video. The idea was to see how well the Luma Keyframes feature would handle these transitions.

Unfortunately, the results were a bit disappointing. Out of the 8 team members we tested, 8 of the transitions were borderline unusable. The feature often resulted in a hard cut in the middle of the scene, rather than the smooth transition we were hoping for.

There were a few exceptions, like Ariad's lightsaber transition, which turned out quite nice. And Larry's transition between the Yoda-like and evil characters also had some interesting moments. But overall, the results were not as seamless as we had anticipated.

It seems that while the Luma Keyframes feature is an interesting concept, in practice it still requires a lot of iteration and fine-tuning to get usable results. The AI-generated content, while visually striking, doesn't always lend itself well to these types of smooth transitions.

So in summary, the Luma Keyframes feature shows promise, but based on our testing, it's not quite ready for prime time yet. The technology is still evolving, and we'll have to see how it improves over time. For now, it may be best to approach it with measured expectations.

Motorola has recently showcased a creative use of AI video technology in their latest ad campaign. By leveraging tools like Control Net and Stable Diffusion, the company was able to generate a series of images that seamlessly integrate the Motorola logo into various fashion-inspired styles.

The workflow likely involved inserting the Motorola logo into the images using Control Net, and then using Stable Diffusion to generate the final visuals. These images were then compiled into a commercial advertisement, complete with music and editing, to create a polished and visually striking final product.

This example demonstrates how companies are starting to explore the practical applications of AI video generation in the real world. While the quality may not yet match the state-of-the-art examples showcased by Gen-Free, Motorola's ad campaign shows that AI-powered video can be a viable and cost-effective solution for certain types of content creation.

As the technology continues to improve, we can expect to see more and more brands and businesses incorporating AI video into their marketing and advertising strategies. This represents an exciting development in the field of AI-powered media production, and it will be interesting to see how the industry continues to evolve in the coming years.

Perplexity, the AI-powered search engine, has introduced a new feature called "Pro Search" that aims to provide more advanced and comprehensive search capabilities. This feature integrates multi-step reasoning and access to external data sources, such as math, programming, and Wolfram Alpha, to deliver more accurate and informative search results.

The key aspects of the Perplexity Pro Search feature include:

1. Multi-Step Reasoning: The search engine can now break down complex queries and perform multiple steps of reasoning to provide more comprehensive and relevant responses. This allows users to ask more nuanced and detailed questions, and receive tailored answers.

2. External Data Integration: Perplexity Pro Search can now access and leverage data from external sources, such as mathematical calculations, programming resources, and the Wolfram Alpha knowledge base. This integration enables the search engine to provide more accurate and informative responses, especially for queries that require specialized knowledge or calculations.

3. Premium Feature: The Perplexity Pro Search functionality is currently available as a premium feature for Perplexity subscribers. This allows users who rely on the platform regularly to access these advanced search capabilities and enhance their overall search experience.

The introduction of Perplexity Pro Search represents a step towards more agentic and intelligent search, where the AI system can actively reason and utilize external resources to deliver more comprehensive and useful information to users. As the demand for advanced search capabilities continues to grow, features like this are likely to become increasingly important in the AI-powered search landscape.

One of the most interesting and fun AI experiments featured in the video is the "Interdimensional Cable ONE" website, shared by Carol in the community. This website is a recreation of the iconic "Interdimensional Cable" concept from the animated show "Rick and Morty".

In the show, Uncle Rick brings back a TV set from another dimension, allowing the characters to view random and bizarre TV stations across the multiverse. The WebSim AI team has rebuilt this concept using web AI, creating a website that delivers a similarly random and unpredictable viewing experience.

When you load the Interdimensional Cable ONE page, you never know what you're going to get. It presents a collection of super random and trippy videos, capturing the essence of the show's "Interdimensional Cable" concept. This experiment is likely to resonate most with dedicated "Rick and Morty" fans, who will appreciate the attention to detail and the faithful recreation of the show's unique aesthetic.

The video highlights that not everything in the world of AI has to be about productivity or serious applications. The fun and experimental side of AI, like the funky video generations or the creation of interdimensional cable-inspired websites, is equally exciting and worth exploring. This project is a testament to the creativity and playfulness that can emerge from the intersection of AI and pop culture.

There's a brand new uncensored multimodal model called Dolphin Vision 72B that has been released. This model is an alteration of the Quen 2 model and represents the largest parameter size we've seen to date.

Running this model will be quite challenging, as it requires a monster PC or renting a lot of GPUs. It's not yet widely popular, but its release is an interesting hint at where we're heading in the future with uncensored, multimodal models.

These types of unrestricted models open up a whole new realm of possibilities, both exciting and concerning. The open source community will likely start building all sorts of applications we can't even predict the existence of right now.

However, the lack of any restrictions on this model also raises serious ethical questions. It's a double-edged sword that will lead to interesting results, but also has the potential for misuse and abuse.

As the capabilities of these uncensored models continue to grow, it will be crucial for the AI community to grapple with the implications and work towards responsible development and deployment.

Figma, the popular design platform, has recently announced a suite of AI-powered features that are poised to revolutionize the way designers work. Two of the standout features are the "Prompt to UI" and "Visual Search" capabilities.

The "Prompt to UI" feature allows designers to simply describe the type of user interface they want, and Figma's AI will generate the entire design for them. This has the potential to dramatically streamline the design process, enabling designers to quickly iterate on ideas and explore new concepts.

However, this feature has also sparked some controversy, as it appears to have closely mimicked Apple's Weather app design. Figma has since disabled this feature, acknowledging the need to ensure their AI-generated designs do not infringe on existing intellectual property.

The "Visual Search" feature, on the other hand, is a more widely applicable innovation. By leveraging advanced multimodal models, designers can now search their entire design library using natural language queries, without relying on specific metadata or tags. This opens up new possibilities for designers to quickly find and incorporate relevant assets into their projects.

As these AI-powered features continue to evolve, it will be crucial for Figma and other design platforms to navigate the delicate balance between innovation and ethical considerations. Ensuring that AI-generated designs respect intellectual property rights and do not perpetuate biases will be key to the successful adoption of these transformative technologies.

Overall, Figma's AI revolution represents an exciting step forward in the design industry, with the potential to streamline workflows, foster creativity, and unlock new levels of productivity. However, the challenges posed by these advancements will require careful attention and thoughtful implementation to ensure the responsible and equitable development of these powerful tools.

The Google crossword game features an interesting integration of AI to enhance the player's experience. The key aspect of this integration is the ability to obtain hints from the AI system.

When the player is stuck on a clue, they can request a hint from the AI. The AI will then respond with a simple "yes" or "no" answer, indicating whether the player is on the right track with their current approach. This limited response is a deliberate design choice, as the team behind the game tried to prompt the AI for more detailed hints, but the system refused to provide anything beyond the binary yes/no feedback.

This implementation showcases a thoughtful approach to leveraging AI in a game context. By restricting the AI's responses to a simple yes/no, the developers ensure that the player remains engaged in the problem-solving process, rather than relying too heavily on the AI to provide the solution. The player is encouraged to continue their own thought process and exploration, with the AI serving as a guiding light rather than a complete solution provider.

This integration of AI hints adds an extra layer of interactivity and challenge to the crossword experience, encouraging players to think critically and utilize the AI's feedback effectively. It demonstrates how AI can be integrated into traditional game mechanics to enhance the overall gameplay experience.

Hugging Face, the leading platform for large language model evaluation, has recently overhauled its leaderboard to address the shortcomings of existing benchmarks. The new leaderboard introduces the MLU Pro, GPT-QA, and MSU benchmarks, which are considered more reliable and advanced assessments of language models.

The key highlights of the updated leaderboard include:

1. Normalized Scores: The scores are now normalized to provide a more meaningful comparison across different benchmarks.

2. Community Voting System: A community voting system has been implemented to address the issue of contamination, where models are evaluated on data they were trained on.

3. Quant 72B Model Instruction: The Quant 72 billion parameter model currently leads the leaderboard, showcasing the continued progress in large language model development.

Hugging Face emphasizes that this updated leaderboard is a crucial tool for navigating the modern AI landscape, as it provides a more reliable and transparent assessment of language model performance. The open leaderboard is available for the community to explore and track the advancements in the field.