OpenAI has made their deep research feature available on the free plan of Chat GPT. This lightweight version of Deep Research is powered by a version of O4 Mini, which is nearly as intelligent as the original Deep Research feature, but significantly cheaper to serve.

The key details are:

• Free Chat GPT users get 5 uses of the lightweight Deep Research per month.
• Plus, Team, and Pro users will have access to the original Deep Research, with the lightweight version available once the original limits are reached.
• The lightweight version will provide shorter responses, but maintain the depth and quality of the original Deep Research.
• This change is aimed at increasing the current rate limits for Deep Research across all plan levels.

So in summary, free Chat GPT users now have access to a powerful deep research tool, while paid users will benefit from expanded usage of the original Deep Research feature.

Rumor has it that OpenAI is planning to release a new open model around June. This model will be available to download at no cost and won't be gated behind an API, allowing users to run it on their local machines.

The new model is reportedly targeting performance superior to open models from Meta and Deepseek. It's said to have a larger context window than OpenAI's current closed models, potentially matching or exceeding the 10 million token context window of Meta's Lama 4 model.

One interesting feature of this open model, if the rumors are accurate, is the ability to call upon other models. If the prompt is too complicated for the open model, it may be able to hand off the query to one of OpenAI's larger, closed-source models via the OpenAI API to provide a more substantial computational lift.

However, it's still unclear whether this open model will support capabilities like web search or image generation. Nonetheless, the release of an open-source model from OpenAI is an exciting development, as it could address some of the common objections to using their closed-source tools, such as the inability to run them locally and concerns about data privacy.

Perplexity, the AI-powered search engine, has rolled out a new feature called the Perplexity Assistant in their iOS app. This feature aims to provide a more comprehensive digital assistant experience, going beyond basic question-answering capabilities.

The Perplexity Assistant allows users to perform a variety of tasks, including:

• Playing media (e.g., podcasts, songs, videos)
• Drafting emails
• Scheduling and managing meetings
• Booking rides and reservations
• Setting reminders

The CEO of Perplexity, Raven, has cautioned that some of these features are still a bit "janky" and the company is working to improve them. However, the introduction of these voice control-enabled capabilities represents a significant step forward in making Perplexity a more versatile and user-friendly AI assistant.

During testing, the Perplexity Assistant demonstrated mixed results, with some features working well (such as opening podcasts and booking reservations) while others encountered issues (like inconsistent calendar integration). Nevertheless, the overall direction of the Perplexity Assistant suggests an effort to create a more comprehensive digital assistant that can handle a broader range of tasks beyond just information retrieval.

As Perplexity continues to refine and expand the capabilities of its Assistant, users can expect to see a more polished and reliable experience that challenges the limitations of traditional virtual assistants like Siri.

Microsoft showcased new updates to its 365 Copilot, including more agentic features. The key highlights include:

• AI-powered search to help find information faster
• New "create experience" with Copilot notebooks
• An agent store to find and access various agents

The demo showcased the researcher and analyst agents, which are optimized for deeper research and data analysis tasks. Users can give these agents data from Excel, Word, and other sources, and the agents will analyze the information.

Microsoft plans to roll out these new Copilot features in the spring, allowing most users to start accessing the enhanced capabilities in late May.

The recall feature, which allows users to retrace their digital steps on their computer, is finally rolling out from Microsoft later this spring. This feature provides a way to quickly and securely find and get back to an app, website, image, or document that you were previously working on.

Some key details about the Microsoft Recall feature:

• It is an opt-in experience, not enabled by default, to address privacy concerns.
• There are controls and filters to specify what gets saved for later recall.
• Recall data is processed locally on the user's device and not sent to the cloud or shared with Microsoft.
• It aims to address the common frustration of trying to pick up where you left off, whether it's a project from last week or one of many open browser tabs.
• Users will be able to use AI-powered search to quickly find and access their previous digital activities.
• The feature includes a "click to do" functionality to summarize, rewrite, or copy/paste text and images from the screen.

Overall, the Microsoft Recall feature seems designed to provide a more seamless way for users to retrace their steps and recover information they've previously accessed on their computers, without privacy concerns around data being sent to the cloud. Its rollout later this spring will be an interesting development to watch in the productivity and digital organization space.

XAI has updated the Grock chatbot to include vision capabilities similar to the Gemini models and OpenAI models. This new feature is available in the Gro mobile app.

To use the vision functionality, users can click the white circle (or dark circle in dark mode) at the bottom right of the chat window, then click the camera icon on the bottom left. This allows Grock to see the world around the user and analyze the visual information.

When testing the new vision feature, the chatbot was able to accurately describe the user's workstation, including the monitor displaying a menu with options like news, movie pick, and NBA highlight, as well as the camera mounted above the monitor and other tech gear on the desk.

The chatbot was also able to analyze a picture on the wall behind the camera, describing it as a scenic coastal landscape with a beach, waves, and greenery under a partly cloudy sky. While it couldn't pinpoint the exact location, the chatbot suggested it could be a tropical location like Hawaii or a similar Pacific island.

Overall, the addition of vision capabilities to the Grock chatbot is a significant improvement, allowing users to interact with the AI in a more natural and intuitive way by incorporating visual information into the conversation.

LTX Studio this week added Google's V2 video generation model into their platform. This makes LTX Studio the least expensive way to generate videos with V2, costing about $0.65 per 8 seconds of video compared to $0.50 per second on Google's own cloud platform.

To try out V2 in LTX Studio, you can click into the motion generator and select V2 from the dropdown. You can then start from an image prompt or generate a video from scratch.

For example, I uploaded an image of sea turtles on a beach in Kauai and had the turtles crawl up the beach in the generated video. I also generated a video of a wolf howling at the moon, which looked very realistic.

LTX Studio is becoming more model agnostic, allowing you to use V2 in addition to their own open-source LTXV model. If you sign up for LTX Studio before May 3rd, you can get up to an extra $300 in video generation credits, so be sure to check out the link in the description.

If you have a pair of Ray-Ban Metas, they rolled out some new features this week including live translations. This means that if someone is speaking to you in a different language, the little headphones on the glasses will actually live translate their speech into your preferred language.

For example, someone could be speaking to you in Spanish, and you'll hear an audio translation in English through the headphones. This is a really cool feature that I got to test out at MetaConnect last year, and it worked impressively well.

Another great thing is that you can now download language packs in advance, so you can use the live translation feature even if you don't have an internet connection. This solves one of the issues with using these glasses, where I was previously unable to get answers to my questions about the Grand Canyon because I didn't have a good internet connection.

Overall, the new live translation feature in the Ray-Ban Metas is a really impressive capability that makes communication across language barriers much easier. It's great to see this kind of advanced AI-powered functionality making its way into consumer devices.

According to the report, YouTube is testing a new "AI overview" feature that will use AI to highlight relevant video clips in search results. Instead of just displaying text summaries, the AI will pull short video clips from relevant videos that are most helpful for the user's search query.

The goal of this feature is to allow users to quickly find the information they're looking for without necessarily needing to click into a full video. The AI-selected clips are intended to encourage users to engage with the full videos, but there are concerns that it could also lead to fewer people actually watching the complete videos.

This feature is currently being tested with a small number of YouTube Premium users in English. It remains to be seen how creators will respond to having their content potentially summarized and clipped by the AI system. Some may see it as a way to surface their content more easily, while others may be concerned about losing views and engagement on their full videos.

Overall, this represents another step in YouTube's efforts to leverage AI to improve the search and discovery experience on the platform. As with any new AI-powered feature, there will likely be an ongoing balance to strike between user convenience and creator interests.

Anthropic has emphasized the importance of paying attention to a wide range of potential harms from AI systems, beyond just the "giant doomsday scenarios" that often dominate the discussion. They have outlined a multi-faceted approach to understanding and addressing these harms, including:

• Physical impacts: Considering the real-world effects AI systems could have, such as on the environment or human health.
• Psychological impacts: Examining how AI could affect human mental wellbeing, cognition, and decision-making.
• Economic impacts: Analyzing the potential disruptions AI may cause to jobs, industries, and economic systems.
• Societal impacts: Evaluating how AI could exacerbate existing inequalities or create new forms of social harm.
• Individual autonomy impacts: Ensuring AI respects and preserves human agency and self-determination.

Anthropic claims to have made adjustments to their Claude 3.7 model to reduce the likelihood of it producing harmful outputs, while still maintaining important safeguards. They have also published case studies on how their models have already been misused, in an effort to raise awareness and encourage responsibility among AI developers and users.

The company's CEO, Dario Amodei, has emphasized the urgent need for better interpretability and transparency in these large language models, to gain a deeper understanding of how they operate and the risks they pose. Anthropic seems to be taking a more cautious and comprehensive approach to AI development, compared to the rapid pace of some competitors.

OpenAI has shipped their image generation model in the API. This means developers can now use the same technology behind the image generations seen in ChatGPT to create images programmatically. This will likely lead to a surge of new AI-powered image creation tools and applications.

Additionally, Gro has released a new Gro 3 Mini API. According to benchmarks, this smaller Gro model outperforms several other popular models like Gemini 2.5, GPT-4 Mini, and even Anthropic's Claude 3.7 on a variety of tasks. The pricing for the Gro 3 Mini API is also quite a bit less than the other models, making it an attractive option for developers.

These new API releases from both OpenAI and Gro provide developers with powerful new tools to integrate advanced AI capabilities into their applications. The image generation from OpenAI and the high-performing yet cost-effective Gro 3 Mini model open up new possibilities for AI-powered software and services.

Adobe released a new version of Firefly, their web app, and added the ability to choose other models. The new models available include Firefly Image 4, Firefly Image 4 Ultra, and GPT Image. When using the prompt "wolf howling at the moon", the Firefly Image 4 model produced better results than the Firefly Image 4 Ultra model.

The company Crea AI also rolled out the ability to edit images in chat using the ChatGPT image model. Users can now perform various image editing tasks, such as "giblifying" and "frogifying" images, directly within the Crea AI platform.

Crea AI also introduced a new feature called "Stage", which allows users to create 3D environments with AI from images or text. By prompting "cowboy movie scene", the tool generated a 3D scene with various assets that could be further manipulated and customized.

This week, the company Tencent released a new model called Hunan 3D 2.5. This model is a 10 billion parameter model, which is up from 1 billion parameters previously. It has high-quality textures and an animation boost.

The demo video showcases impressive results, with the model generating visuals that look very realistic. While the company likely cherry-picks the best examples, the overall quality appears to be quite high.

It's only a matter of time before tools like Crea integrate this new Hunan 3D 2.5 model, allowing users to leverage its capabilities. The advancements in 3D content generation continue to impress, and this latest release from Tencent is another step forward in the field of AI-powered 3D creation.

Character AI, a platform that allows users to create and chat with fictional characters, has recently rolled out a new feature that generates visuals for the characters. This new "avatar effects" feature creates animations for the characters, making the conversation feel more like interacting with a real character rather than just a text-based chatbot.

The company is currently rolling out this new feature and it is not yet available for everyone. Users can apply for early access to try out the new avatar video generation capabilities.

This new feature from Character AI has significant implications for e-commerce and marketing, as companies can now create AI-generated avatars or "spokespeople" to hold and showcase their products. This could lead to more engaging and personalized product presentations and brand experiences.

Overall, the addition of avatar video generation to Character AI's platform represents an exciting step forward in the world of AI-powered virtual interactions and digital brand experiences.

Argil, a company specializing in AI-powered solutions, has recently rolled out a new feature that allows users to have AI-generated avatars hold and showcase actual products. This innovative approach is expected to have a significant impact on the e-commerce industry, providing businesses with a unique way to present their products to potential customers.

The new feature enables the creation of AI-generated avatars that can hold and interact with physical products, creating a more engaging and immersive shopping experience. These avatars can be customized to match the brand's identity and can be used in various scenarios, such as product demonstrations, tutorials, or even as virtual spokespersons.

One of the key advantages of this technology is the ability to create consistent and scalable product presentations. Businesses can now showcase their products in a more dynamic and personalized way, without the need for extensive photoshoots or video production. This can lead to increased customer engagement, improved product understanding, and ultimately, higher conversion rates.

Furthermore, the integration of AI avatars into e-commerce platforms can also provide valuable data insights. By tracking customer interactions and engagement with the virtual product showcases, businesses can gain a deeper understanding of their customers' preferences and behaviors, allowing them to make more informed decisions about their product offerings and marketing strategies.

As the e-commerce landscape continues to evolve, the introduction of Argil's AI avatar technology represents a significant step forward in the way businesses can present and promote their products online. This innovative solution has the potential to revolutionize the way customers interact with and experience products, ultimately driving greater sales and customer satisfaction.

Tavis has dropped a new lip-syncing model, which is reportedly the best lip-syncing model available currently. While the AI voice with the lip sync still feels a bit uncanny, the model does seem to match up the lips quite well.

In a demo, we can see the model syncing the lips of a Donald Trump video, even though the audio is muted. The lips appear to move more naturally, though there is still a slightly supernatural look to the movement.

The creator notes that this is the best lip-syncing they have seen so far, and that the technology is continuously improving. As the AI-generated lip-syncing gets more realistic, it will have significant implications for video production and editing.

Descript is testing out new AI agentic features, claiming they are building the "cursor for AI video editing". In a demo, they show the AI responding to prompts to edit a video.

The AI can:

• Write a script draft based on a prompt
• Edit a video by making cuts, adding chapter titles, and inserting stock overlays
• Mask jump cuts with subtle zooms and add relevant screen recordings

This allows users to simply chat with the AI bot, and it will make the edits based on the instructions. Descript is working to create an AI-powered video editing assistant that can handle various tasks like a human editor would.

While not publicly available yet, this technology represents a significant step towards agentic AI video editing. Users will be able to leverage the AI's capabilities to streamline the video production process, without needing extensive video editing skills.

The Oscars have announced that they are open to the use of generative artificial intelligence and other digital tools in the making of films. Their statement clarifies that the tools "neither help nor harm the chances of achieving a nomination." The Academy and each branch will judge the achievement, taking into account the degree to which a human was at the heart of the creative authorship when choosing which movie to award.

The Academy's stance suggests that they are not opposed to the use of AI in filmmaking, as long as the human creative process remains central. While a fully AI-generated movie is unlikely to win awards anytime soon, the use of AI tools to assist in various aspects of filmmaking is acceptable. This decision reflects the evolving role of technology in the creative arts and the Academy's willingness to adapt to these changes.

In a court case against Google, OpenAI has stated that they would be interested in purchasing Google's Chrome browser if it was made available for sale. This comes as the courts are currently trying to break up Google due to monopoly concerns.

Both OpenAI and Perplexity have expressed interest in potentially acquiring Chrome, should Google be required to sell the browser. This suggests that OpenAI may be interested in building an AI-first browser, leveraging the existing Chrome infrastructure.

The motivation behind this potential move is likely to create a browser that is more aligned with OpenAI's vision and values, potentially offering deeper integration with their AI models and services. By owning the browser, OpenAI could have more control over the user experience and how their technology is surfaced to end-users.

This development highlights the growing ambitions of OpenAI and their desire to expand beyond just language models and into other areas of the technology landscape. It will be interesting to see how this situation unfolds and whether OpenAI is successful in their pursuit of the Chrome browser.

In the 60 Minutes interview, DeepMind CEO Demis Hassabis shared his thoughts on whether AI systems can develop self-awareness:

"Is self-awareness a goal of yours? Not explicitly, but it may happen implicitly. These systems might acquire some feeling of self-awareness. That is possible. I think it's important for these systems to understand 'you', 'self', and 'other', and that's probably the beginning of something like self-awareness."

Hassabis believes that while explicit self-awareness may not be a direct goal, AI systems could potentially develop a sense of self-awareness as they learn to understand the concepts of "you", "self", and "other". He sees this as an important step in the potential emergence of self-awareness in advanced AI systems.

The idea of AI systems becoming self-aware is a fascinating and complex topic that raises many philosophical and ethical questions. Hassabis' comments suggest that DeepMind is considering the implications of AI systems potentially developing some form of self-awareness, even if it is not a primary objective in their research.