ChatGPT has recently introduced a new feature called "memories", which allows the AI to reference your past conversations to provide more personalized responses. This is a significant update that users need to be aware of.

The memory feature is a double-edged sword. For most users, it can be a fantastic beginner feature, as the extra context will make the responses more personalized and engaging. However, for intermediate to advanced users who craft specific prompts for reliable tasks, the memories can introduce unwanted context that can warp the outputs in unexpected ways.

The key is that the memories now not only include the information you've explicitly shared, but also the entire chat history. This means ChatGPT can pull context from conversations you had months ago, which may not be relevant to your current task.

For simple prompts, the extra context from memories can be great. But for more complex, work-related tasks, the memories can introduce irrelevant information that can negatively impact the output. The best workflow for advanced users is to manually manage all context using the "Customize ChatGPT" feature, selectively pulling in only the relevant memories.

Users with less than 50 hours in ChatGPT can likely leave the memories feature on without issue. But those who use ChatGPT extensively for various use cases need to be aware of this new capability and how it can affect their prompts. Carefully managing the context is key to getting reliable, on-target outputs from ChatGPT.

Okay next up we have the release of MidJourney V7 that came out end of last week. So it didn't make it into the cut off for last week's video so we will feature it here. This was a very anticipated release. Everybody was looking forward to Midjourney making the next big move in the AI image generation space. They were probably the first tool to really go viral and where people were stunned by the quality of the outputs. And now they came out with their newest version V7.

Now I don't think there's any other way to put it than to say that this was a disappointment to most people. The hype was very real and maybe the expectations were a little overblown because most people expected them to push the image generation space further than it is right now. But that happening around 2 weeks after Chat GPT released their image generation feature that went gigafiral - probably the most significant mainstream AI moment this year. Midjourney adding a few features that I'll talk about here briefly just didn't cut it for most people.

And I think it does make a lot of sense to me because at the core of the Midjourney value proposition always has been this idea of this utility of it generating stunning visuals - the most pleasing and aesthetic visuals that you can get from any image generator. And it still does that. It's incredible. Did it get better at that? I don't know. I personally don't really think so. They added amazing new features like draft mode where you can turn on your mic or type and it just instantly generates what you're talking in the form of lower quality images. That works super well. But the quality of the base model didn't get much better if better at all. It still has many problems like thumbs still don't work properly as you can see in this image.

But the biggest thing is that this core value prop of it being aesthetically pleasing kind of plateaued since V5. In my opinion, it is already excellent. It's already a 10 out of 10 in terms of aesthetics. Now how are you going to top that? You can't. You can add new features but that's not the reason that 98% of the people use Midjourney.

Now to get a more concrete feeling and opinion of this yourself, we ran a bunch of test images. And if for example I review these cinematic stills, they're still stunning and they're still the most aesthetic image generation model out there. I think that stands, but it doesn't change the game for them. And most people will probably be fine with the Chat GPT output or one of the others. You can see some comparisons to some of the other best image generators right now. And if you want to kind of just work with an image, the text interface of Chat and the editing capabilities are just superior.

So yeah, there you go. It's still stunning. It's still amazing, but it's not really the upgrade that people were hoping for. I just think for most users, Chat image generation does almost everything they can think of, but it's a solid update for all Midjourney users, particularly the draft mode is something that is a lot of fun to work with. It almost generates at the speed of your own fault as you get to talk to it. And I personally think that's great. I just think that most people will prefer a tool that is almost as good but free and more intuitive over V7. So there you go. That's my take. You can make up your own mind in terms of the quality. And let's move on to the next one.

Next up there's a new agentic tool called GenSpark Super Agent. While there are many of these tools coming out, our brief testing suggests this one works better than both Operator and Manos.

What's great about GenSpark Super Agent is that there are free credits available for you to try it right now. This free trial won't last forever, but you can log in with a Google account and give it a shot to see how it works for you. This is not the case with the competition - Manos is now rolling out to the wider public but the paid plans start at $39 per month, and ChachiV Operator is still behind a $200 paywall for the pro plan.

From the few use cases we've tried, GenSpark Super Agent appears to work better than both Manos and Operator. For example, if you wanted to try a deep research task like finding the best camera store in Japan, you can see it work through the step-by-step process, using web searches, map searches, and various tools, before summarizing the results.

GenSpark Super Agent also has all the different video and image generators built-in, along with various presets where you can use the agents to perform specific tasks for you. This could be interesting for things like booking a restaurant, though we haven't tested that functionality yet.

Overall, GenSpark Super Agent seems to be one of the most capable agentic consumer products available right now, and the fact that you can try it for free makes it worth checking out.

Next up we have an application that really pushes what users can do with voice AI agents to the next level. The site called Vapi (which I guess stands for voice API) allows you to combine multiple services. If you've been following the show, you know we've got so many new voice services recently.

What this practically means is you can use the brain of a model like Sonnet 3.7 and the smartness of it, but the voices from GPT-4 Transcribe, which are really good and sound more authentic than a lot of the competition.

Let's just open a dashboard here that you get if you just log in. I know that most people won't be using this, but I think it's interesting to point out that voice is really moving forward super fast.

Fun fact from the research I did on a lecture I held recently on voice assistants: did you know that voice transcription accuracy improved by 5x since 2012? So if you ever talk to one of these artificial support agents at a bank or something like that, and you had a terrible experience where they asked you a question, you responded "yes", and then it kept asking if you actually want that and went in a different direction, that was because these voice transcription services were just not that good and not that accurate. But with the rise of generative AI, that changed, and this platform allows you to combine multiple of them.

Let's have a look. You can see right here this makes it feasible to integrate it into your workflows. On the provider side, let's say we just pick GPT-4 for the transcription service. We're going to use something completely different - I can go to OpenAI Transcribe, or I can use Deepgram. If you're a company in the EU, you might want to go with Azure to make sure that the data stays in the European continent and it complies with GDPR. A lot of options here, including ones like 11Lab Scribe which we talked about.

And then the voice configuration can come from yet another service. As you can see, this makes it really modular. Maybe I want to go with the human voices here - those are a bit more emotional than some of the others. And then you could add all these different tools. These are basically the actions that the agent can take, like the agent could send texts or add data to Google Sheets, whatever it might be. And then many more detailed settings.

But once you kind of go for that, you can just hit "Talk to Assistant", and we're going to use a combination of all these services here with the microphone on the computer.

Plays audio of voice assistant

As you can see, Riley does not comply with that because apparently he's set up here to only help with scheduling in this big system prompt. So there you go, I wanted to point this one out because it's not about just better models coming out. It's really about combining multiple releases that we look at week by week and getting the best of each one to build voice assistants and voice agents that just go above and beyond anything we've seen before. Yeah, you can probably expect a combination of services like this and more apps in the future.

Notebook LM, the popular AI-powered research tool, has recently introduced a brand-new feature that significantly enhances its capabilities. The latest update includes a feature called "Discover Sources," which allows users to effortlessly find relevant sources for their research projects.

Previously, users had to manually curate and add their own sources to Notebook LM. However, with the new Discover Sources feature, the tool can now automatically scour the internet and identify the most relevant sources based on the user's research topic. This streamlines the research process, saving users valuable time and effort.

The Discover Sources feature works by leveraging Notebook LM's advanced natural language processing and information retrieval algorithms. When a user initiates a research task, the tool analyzes the topic and queries the web to find the most pertinent sources. These sources are then seamlessly integrated into the user's Notebook LM workspace, ready for further analysis and processing.

In addition to the Discover Sources feature, Notebook LM has also introduced a new "I'm Feeling Curious" option. This feature randomly selects a topic and generates a set of sources for the user to explore, encouraging serendipitous discovery and broadening the user's knowledge horizons.

These new features further solidify Notebook LM's position as a powerful research tool, making it easier than ever for users to access a wealth of information and insights. Whether you're conducting in-depth research or simply exploring new topics, Notebook LM's latest updates are sure to enhance your productivity and knowledge-gathering capabilities.

This app called Covi presents an interesting interface that could represent the future of AI-powered tools, particularly in areas like crafting marketing materials or learning in tandem with AI.

The key aspects of the Covi interface are:

1. Modular Approach: When given a task, Covi breaks it down into different modules and components, building custom applications and visualizations to assist in the process.

2. Integrated Tools: Rather than just providing text-based responses, Covi generates custom tools and apps that can be used directly within the interface. This includes things like a stargazing guide app, flashcards, and interactive quizzes.

3. Dynamic Adaptation: The interface adapts based on the specific task, creating a tailored experience rather than a one-size-fits-all chatbot approach.

The reviewer believes this type of interface could be particularly useful for learning new skills, where the AI can create a personalized curriculum, compile relevant information, and generate interactive tools to enhance the learning experience.

Additionally, the reviewer suggests that as AI assistants like ChatGPT evolve, we may see them incorporate similar modular, application-generating interfaces, moving beyond the traditional chatbot format to provide more specialized and integrated experiences for users.

Overall, the Covi app showcases an intriguing direction for the future of AI-powered tools, blending language understanding, task decomposition, and custom application generation to create more dynamic and tailored user experiences.

For the next story, I want to briefly point out the monthly challenge that we're running with the AI Advantage for the month of April. The concept is very simple - we're just asking the entire AI Advantage community, including you the viewer of this YouTube channel, to remember the moment that AI clicked for you.

If you think about it for a little bit, for most people there is a moment that you can identify. Whether it's taking a photo of your fridge and asking for a recipe, or it's solving hours of work in a simple intuitive prompt that you just wrote, and from then on out you just start using the product more and more.

Now we talked about this a lot in the community, and we'd love to hear also from you. That's why we created this challenge. The way we kind of frame this challenge is to present that "aha" moment you had in the form of an image that you created with GPT-4.0, and you can submit it here to win memberships to our community and a cash prize too.

We absolutely love doing these challenges because not just the people who participate, but also all the viewers, learn so much from the entire community. So check that out in the free area of our community. The link is below. All the details are below there, and I can't wait to hear about your personal AI aha moment. I'll be sharing mine too, by the way.

The main theme of the Google Cloud Next conference this week was agents. Google has consolidated a lot of progress in the generative AI space and is making it available in a secure way to anyone using Google Cloud.

One new model they introduced was Lyria, a music generation model. This is the first time a major tech company like Google has released a music generation model, as previous efforts in this area have come from smaller independent startups.

Overall, the focus of the conference was on providing access to powerful generative AI models and capabilities through Google Cloud. While most of the announcements are not available for use today, a 10-minute recap video summarizing the key highlights has been shared in the weekly newsletter and is linked in the description below.

Gemini's much-anticipated Project Astra, a visual assistant, has been met with mixed reviews upon its initial rollout. While the preview video showcased an impressive office assistant that could interact with objects and answer questions, the actual implementation seems to fall short of expectations.

According to reports, the visual feature of Project Astra is not true video, but rather a series of regularly taken screenshots that are then processed by a vision module. This, coupled with the voice assistant, has led some to compare it to a Google Lens-like experience with a voice component attached.

Many users who have had the chance to try out Project Astra have expressed disappointment, finding the implementation to be weaker than expected. The comparison to the Chat GPT desktop app, which provides access to the user's camera and desktop, has also been made, with the consensus being that the "fire hose" approach of providing full context is often not ideal for most use cases.

Users typically prefer a more selective approach, where they can choose what information to provide the AI assistant. The step-by-step rollout of Project Astra has also been criticized, as it seems to lack the excitement and game-changing potential that the initial preview video had suggested.

While Gemini is likely to continue iterating on Project Astra, the current consensus among those who have tried it is that it does not live up to the hype and excitement generated by the initial announcement. The comparison to existing features, such as the camera sharing in Google AI Studio, further dampens the enthusiasm for this new visual assistant offering.

Hickfield AI is a new video generator that has been making waves in the AI community. It specializes in two key areas:

1. Generating videos of humans with anatomically correct features: Hickfield AI has demonstrated impressive results in creating videos of humans that look realistic and anatomically accurate, even in complex scenes like a person walking on a beach.

2. Providing custom camera movements with a variety of presets: The tool also excels at generating videos with creative camera movements, including a 360-degree orbit around a subject's face, which is typically a challenging task.

When comparing Hickfield AI's performance to other top video generation models, it stands out for its ability to handle human subjects and complex camera movements. The results are quite impressive, with the generated videos exhibiting a high level of realism and attention to detail.

One key advantage of Hickfield AI is its focus on solving the challenging problem of generating accurate human anatomy and movement. Many other video generation models struggle with this aspect, but Hickfield AI seems to have cracked the code.

Additionally, the tool's integration of various camera movement presets allows users to create dynamic and visually engaging videos without the need for specialized equipment or expertise.

Overall, Hickfield AI appears to be a powerful and versatile video generation tool, particularly for those working in creative industries or requiring high-quality human-centric visuals. Its ability to generate realistic and anatomically correct human subjects, combined with its custom camera movement capabilities, make it a standout in the AI video generation space.

This is an update to Microsoft Copilot, the AI assistant available to many corporations. The main selling point of Copilot is its focus on data privacy, making it suitable for use in many organizations.

The key updates to Copilot include:

1. Memory: Copilot is adding a memory feature, similar to the "custom instructions" in ChatGPT. This allows Copilot to learn about you from your past conversations and provide more personalized responses.

2. Web Browsing: Copilot is integrating web browsing capabilities, which can help reduce hallucinations by allowing it to reference online sources for information, rather than just relying on its own knowledge.

3. Actions: Copilot is adding the ability to connect to and perform actions in other applications, making it more integrated into users' workflows.

While these updates sound promising, there has been some miscommunication about their availability. Based on our testing, these new features are not yet widely accessible. Some outlets have reported that they are already rolled out, but we were unable to access them despite trying various methods.

It's likely that these Copilot upgrades will be rolled out gradually over the coming months. For current Copilot users, these new features are something to look forward to, as they will significantly improve the functionality and capabilities of the AI assistant.