Perplexity, an AI-powered search engine, has recently introduced a new feature called "Deep Research." This feature provides an alternative to OpenAI's Deep Research, which is part of their paid plan.

The key highlights of Perplexity's Deep Research feature are:

• It is available on the free plan, allowing users to access this powerful tool without a paid subscription.
• The quality of the results is good, providing a comprehensive summary of the topic, though not as in-depth as OpenAI's Deep Research.
• The output is presented in a bullet-point style, making it more concise and easier to digest compared to the lengthy reports generated by OpenAI.
• Perplexity's Deep Research utilizes a larger number of sources, often over 100, compared to OpenAI's 40-source limit. However, the quality of the information is more important than the quantity of sources.
• For users who cannot afford the high cost of OpenAI's Deep Research, Perplexity's offering provides a valuable and accessible alternative that can still deliver useful insights.

Overall, Perplexity's Deep Research feature is a strong contender in the AI research landscape, offering a free and convenient option for users who need quick and informative summaries on various topics.

Grok Free, the latest release from Anthropic, is a state-of-the-art AI model that stands out for its reasoning capabilities and uncensored approach. Here are the key highlights:

Benchmarks: Grok Free outperforms models like GPT-3, Chinchilla, and even Anthropic's own GPT-3 based models on various benchmarks. It showcases impressive reasoning abilities.

Uncensored Approach: Grok Free takes a more "uncensored" approach compared to other AI assistants. It can be prompted to express opinions, use profanity, and engage in discussions on sensitive topics without restrictions.

Deep Search: Similar to OpenAI's Deep Research, Grok Free offers a "Deep Search" feature that allows users to explore a topic in-depth by leveraging a large number of sources. While not as comprehensive as OpenAI's offering, it provides a solid alternative for those without access to the expensive OpenAI plan.

Speed and Accessibility: Grok Free is faster than many of its competitors and is freely accessible through the Grok website, without the need for a paid plan or API access.

Roadmap: Anthropic is committed to open-sourcing previous Grok models as they release new versions, a commendable approach that promotes transparency and accessibility.

Overall, Grok Free presents an intriguing option for users seeking a more uncensored AI assistant with strong reasoning capabilities and a useful Deep Search feature. While it may not surpass OpenAI's Deep Research in terms of depth and detail, it offers a compelling alternative for those looking to explore the capabilities of large language models.

This week, Perplexity released an uncensored version of their GPT model, removing the censorship around topics related to the Chinese Communist Party.

Specifically, Perplexity outlined how they went about this process. They created a new dataset of 40,000 prompts covering topics known to be censored by the Chinese government. They then used this dataset to fine-tune their GPT model, effectively removing the previous censorship.

This uncensored version of the model is now openly available on Hugging Face. It represents an important step in providing access to information that was previously restricted due to political censorship.

The release of this model highlights Perplexity's commitment to intellectual freedom and their willingness to challenge censorship, even when it comes to sensitive geopolitical topics. It's a notable development in the ongoing efforts to make large language models more open and transparent.

This next section will be super quick. It's just a new leaderboard for various agent-focused large language models (LLMs). The definition of "agents" varies, but here they refer to LLMs that are particularly good at using tools.

If you've been looking into creating something yourself, this might be a potentially valuable resource for you to quickly gauge which models you might want to use when building an AI application that is supposed to do things for you.

This section covers the model specification released by OpenAI, which outlines their vision for the behavior and capabilities of their large language models.

The model specification serves as a "job description" for ChatGPT and other OpenAI models, defining the desired behaviors and principles that will guide their development. Some key points from the specification:

• Intellectual Freedom: The models will assume the user's best intentions and help with tasks that may be in a "gray area" legally or ethically, rather than refusing outright.
• Transparency: The models will aim to be transparent about their capabilities, limitations, and decision-making process.
• Impartiality: The models will strive to be impartial and avoid taking strong political stances, while still acknowledging the complexity of many issues.
• Safeguards: The models will have safeguards to prevent misuse, such as refusing to help with illegal activities.

This specification provides insight into OpenAI's philosophy and the direction they plan to take their language models in the future. It serves as a high-level guide for how these models should behave, even as their underlying capabilities continue to advance.

The latest update to Gemini, a popular AI assistant, has introduced a new feature called "Improved Memory." This feature allows the chatbot to actively reference previous conversations as context for the current chat session.

Unlike the standard memory feature that simply gathers information about the user, Improved Memory delves into the history of past chats. This provides the chatbot with a more comprehensive understanding of the user's context and preferences, allowing for more personalized and coherent responses.

When enabled, the Improved Memory feature can be accessed through the "Activity" tab in the Gemini interface. Here, users can review their past conversations and see how the chatbot is utilizing this information to enhance the current interaction.

It's important to note that this feature is enabled by default, but users who prefer more manual control over the chatbot's context can easily turn it off. This allows sophisticated users to craft their own prompts and personas, ensuring the most precise and tailored responses.

Overall, the Improved Memory feature in Gemini represents a step forward in the development of conversational AI, providing users with a more seamless and personalized experience. As language models continue to evolve, features like this will become increasingly important in delivering intelligent and contextual assistance.

Project Starlight is a video enhancement model that allows you to improve the quality of your video footage. It has free renders available on the free plan, making it accessible to everyone.

To test the capabilities of Project Starlight, I uploaded a random 1920s video footage from YouTube. I selected a section with people in the frame to see how the model would handle the enhancement.

The results were quite impressive. The model was able to render the frames, effectively improving the quality of the video. It was able to enhance the details and clarity, making the footage look more polished and modern.

This release is an exciting development in the world of video AI, as it demonstrates the ability to improve the quality of existing footage. With Project Starlight, you can take your old or low-quality videos and give them a fresh, high-quality look.

Overall, Project Starlight is a powerful tool that can be a valuable asset for anyone working with video content. Its free renders make it accessible to a wide range of users, allowing them to experiment and see the benefits of AI-powered video enhancement.

In the video AI world, there were two big releases this week that I'll touch on briefly:

1. Adobe Firefly Video: This is the first video generator that, as stated by Adobe, uses only copyright-free materials to train. If you want to be super safe and responsible when using AI video models, this would be the one to use. We ran our classic test prompts through it, and the results can be seen on the screen. While it can hold up with some of the other top-tier models, it's not the new king. In fact, some of the Chinese models might be beating it. However, the selling point of this one is its responsible approach, being trained on copyright-free footage.

2. Sky Reels V1: This is a fine-tuned version of the Hunan video model, one of the top contenders for the best video model out there. The fine-tuning was done on Hollywood footage of strong emotions being displayed. This model is supposed to be really good at portraying humans expressing emotions. We also ran test prompts, and while the results for prompts that don't include human emotions are good (since the base model is good), the clip where the person is smiling with the dog looks particularly convincing. This suggests that having a base model with fine-tunes for different scenarios, like expressing emotions, might be the way to go.

I've included links to all of this below if you want to use any of these models. That's all I have for this week!

In this week's AI news, we covered several exciting developments in the world of language models and video enhancement:

1. Perplexity's Deep Research Feature: We explored Perplexity's new deep research feature, which provides a free alternative to OpenAI's powerful Deep Research tool. While not as comprehensive as OpenAI's offering, Perplexity's feature is a solid option for those looking to leverage advanced research capabilities without the high cost.

2. Anthropic's Grok Free Release: The release of Anthropic's Grok Free model has shaken up the AI landscape. With its impressive benchmarks and uncensored approach, Grok Free presents a unique alternative to other language models, catering to users who value a more unhindered conversational experience.

3. Perplexity's Uncensored R1 Model: Perplexity has also released an uncensored version of the DeepSpeech R1 model, removing the Chinese Communist Party censorship that was previously present. This move expands the model's capabilities and accessibility.

4. Gemini's Improved Memory Feature: Gemini's new "Improved Memory" feature, currently in alpha testing within ChatGPT, allows the model to reference previous conversations, providing a more contextual and personalized experience for users.

5. Video Enhancement Tools: We explored two new video enhancement tools, Adobe Firefly Video and Sky Reels V1, which offer different approaches to improving video quality and expressiveness.

These developments showcase the rapid pace of innovation in the AI industry, with language models and video enhancement tools continuously pushing the boundaries of what's possible. As these technologies evolve, it's important for users to stay informed and explore the various options available to find the solutions that best fit their needs.