Airdroplet AI summary

AI News: Vibe Jam, The BEST Small LLM, Claude Search, OpenAI Audio Models, and more!

March 21, 2025Matthew BermanAI score 9863,171 views

AI-generated summary

Here's a rundown of the latest happenings in the AI world, covering everything from incredibly fast-paced game development competitions using AI to powerful new language models and creative tools. It's exciting to see smaller models performing exceptionally well, major players adding essential features like web search, and surprising new entrants like LG releasing advanced AI models. The focus is clearly shifting towards making AI more accessible for local use and empowering creators.

Here are the key updates discussed:

Vibe Coding Jam is in full effect: This is a competition for building web-based multiplayer games using "Vibe Coding." The idea seems to have exploded since a demo of a vibe-coded flight simulator came out recently.
- Some early submissions are described as "insane" and "legitimately fun."
- Examples include a Fortnite-style game with Minecraft looks (multiplayer working), a simple safari driving game, a Line Rider style game, a Tetris-like puzzle game built in hours, an 80s aesthetic tank game, a food fight simulator, and an air traffic control game that evolved from 2D to 3D.
- It's surprising how fast impressive multiplayer games are being created with this approach.
- You can actually try the games right now, which is pretty cool.
- It's definitely worth checking out Vibe Jam on Twitter to see all the submissions as the presenter cannot wait to see them.
Mistral released an incredible small model: This new model, Mistral Small 3.1, is open source and surprisingly outperforms similar larger closed-source models.
- It does particularly well on the Knowledge GPTQA benchmark, showing very low latency per token and a high GPTQA Diamond score.
- It beats models like Gemma 3, Claude 3.5 Haiku, GPT4o Mini, and Cohere Eye of Vision.
- It's relatively small at just 24 billion parameters, making it runnable on a single RTX 4090 or a Mac with 32 GB of RAM, which is great for local use.
- It's multimodal, meaning it can handle more than just text, and is described as a "Foundation for Advanced Reasoning," suggesting it's good for training into a "thinking model."
- It has a decent context window of 128,000 tokens.
- It's encouraged to download and play around with it.
Claude finally gets Web Search: This has been a long-awaited, and frankly, essential feature ("table stakes").
- Claude 3.7 and 3.7 Thinking models now have web search capability.
- This makes Claude another strong alternative to Google Search, joining Grok3, ChatGPT, Perplexity, and Mistral.
- It's particularly useful because Claude is considered one of the best coding models, and now it can reference current API documentation, library info, and web bugs, which is super cool for developers.
OpenAI released three new audio models: There are significant updates to their text-to-speech (TTS) and transcription models.
- Two new transcription models, GPT-4o Transcribe and GPT-4o Mini Transcribe, both outperform the previous Whisper model in every language tested.
- They also introduced a new text-to-speech model that allows for detailed instructions on how the text should be spoken, not just the text itself.
- You can specify voice affect, tone, pacing, and emotion, similar to giving a prompt or system message.
- There's a demo site, openai.fm, where you can try out the TTS with different styles like "choral" or "dramatic," and it sounds really good.
- The interface design is reminiscent of Teenage Engineering, and OpenAI is actually partnering with them.
- They are holding a competition for the best text-to-speech creations, with the top three most creative winning a $550 Teenage Engineering OB4 speaker. This is a great opportunity to play around with the new model.
- The code to implement the TTS is also very simple, just a few lines.
Windsurf Wave 5 is here: This is an update for more traditional coding environments, focusing on improving tab completion.
- Wave 5 integrates autocomplete, super complete, tab to jump, and tab to import into one "seamless tool."
- It can write new code, make multi-line edits, and help navigate files.
- It's seen as a significant leap in quality and speed for passive coding assistance.
- The good news is that even free users get unlimited Windsurf Tab completion.
Krea AI released video training: Krea AI now allows users to train their WAN 2.1 model on their own videos.
- This gives users much more control over their AI video creations.
- Training the model helps it learn specific styles, objects, and even motions from your uploaded videos.
- After training, you can create new AI videos that incorporate those learned elements. This is a cool update for video creators wanting a specific aesthetic or subject in their AI generations.
Notebook LM got a big update with Mindmaps: Notebook LM, a tool for processing documents and podcasts, can now automatically generate a mind map based on the content you provide.
- This is highlighted as a great way to learn and explore the information extracted from your documents.
- Even AI personality Jimmy Apples was impressed by this feature.
Hunyuan announced a major upgrade to their 3D modeling AI: They released two new open-source versions of their 3D generation model, 3D 2.0 MV (multi-view generation) and 3D 2.0 mini.
- These are open source and available to download and use now.
- This is considered powerful stuff for creators who want to generate 3D characters for motion graphics, games, or videos.
Stability AI released Stable Virtual Camera: This new feature allows users to upload 2D images and create immersive 3D videos with controlled camera movements.
- You can do things like zoom out or move around within the scene generated from the 2D image.
- This capability is seen as moving closer to a future where entire movies or TV shows could be created by AI by anyone.
- The weights for the model are open source and free to use for non-commercial projects.
Gemini finally adds Canvas: Gemini now has the ability to write code and run it directly within the browser.
- This is a feature already available in tools like Claude and ChatGPT.
- You can write HTML or JavaScript code, edit it in the browser, run it immediately, and iterate quickly, which is great for "vibe coding" web projects.
LG just released an open-source thinking model called EXAONE: Surprisingly, LG, a company not typically associated with advanced AI models, released an open-source model focused on enhancing reasoning capabilities and particularly on agentic AI.
- It comes in three versions: a 32 billion parameter version (EXAONE Deep 32B) that topped the AIMI benchmark, outperforming competitors at just 5% of its size, and smaller 7.8 billion and 2.4 billion parameter versions suitable for local use.
- The 32B version is shown to be very comparable to models like Deep Seek R1 on benchmarks.
- It's exciting to see a company like LG entering the space with a performant, open-source model aimed at local usage and agentic capabilities.
- It's encouraged to download and try it out.

Video transcript

Open transcript

Lots of AI news to go over this week. First, Vibe Coding is in full effect. Just a few weeks ago, Levels.io put out a demo for a completely Vibe Coded flight simulator, and Vibe Coding games really took off since then. Then a few days ago, he announced Vibe Jam, which is a competition for building web-based multiplayer Vibe Coded games. And some of the submissions already are insane. Here's a Fortnite style game built with Minecraft aesthetics. This was completely Vibe Coded into reality. And it looks so good. It already has multiplayer functionality. They're doing play testing right now. This game looks legitimately fun. And for those of you who are saying this might be fake, you can actually try it right now. I'll drop a link to this post down below. Here's another much simpler game where you're driving through a safari. Here's a Line Rider style game. Here's a puzzle game that was Vibe Coded in just a few hours. I mean, look at the shadows. Everything feels really nice. It's kind of like Tetris, but puzzle style. Here's one that has an 80s aesthetic that is called Vibe Tanks. Here's one being created. It's a food fight simulator. And here's the progression of an air traffic control game. You could see it starts very simply top down 2D. Then they start adding 3D elements and now they actually fill out the map. So really cool stuff. Definitely check out Vibe Jam on Twitter. I cannot wait to see all of the submissions. All right, next, Mistral released an incredible small model. This is an open source small model that actually exceeds the performance of similar larger closed source versions. Check this out. This is Knowledge GPTQA. We can see the latency per token right here on the x-axis and it is very, very low on this. And on the y-axis, the GPTQA Diamond score. And it outperforms Gemma 3, Claw 3.5 Haiku, GPT40 Mini, and Cohere Eye of Vision. This comes in at just 24 billion parameters. And it's multimodal. As it says right here, the model can run on a single RTX 4090 or a Mac with 32 gigabytes of RAM. Great for local inference. And it says Foundation for Advanced Reasoning, meaning you can train it to be a thinking model. And it has a context window of 128,000 tokens. So pretty decent. So check it out, download it, play around with it, and let me know what you think. All right, next, Claude finally gets Web Search. This has been a long awaited feature. It is kind of table stakes at this point. But now, Claude 3.7, 3.7 Thinking, they all get Web Search. So now you have another replacement for Google Search. Grok3, ChatGPT, Perplexity, Mistral, and now Claude can all search the Web. This is also super useful because Claude tends to be the best coding model out there. So now you can actually have it reference up-to-date API docs, library information, and any current bugs that might be floating around on the web. So really cool. Check it out. Next, OpenAI released three new updates to their audio models. I covered this in a full video. Definitely check it out after you watch this video. I'll link it down below. First, they have two new text-to-speech models, both of which outperform their previous Whisper model. One is called GPT-40 Transcribe. The other is GPT-40 Mini Transcribe. And in every single language they tested it with, it outperformed every other model. And they also have a new text-to-speech model that you can give instructions, direction, on how to speak. So not just the text, but specific direction in how you want the model to actually say the words that you're giving it. You could try out the text-to-speech right now at openai.fm. So here's an example of what this looks like. We're going to use choral, dramatic, and you can have voice affect, tone, pacing, emotion. You describe very similar to a prompt or a system message exactly how you want it said. And let's listen. It was thick with fog, wrapping the town in mist. Detective Evelyn Harper pulled her coat top. So really good. And it's super easy to build too. If you click this little button up in the top right, you get the code for it. Just a few lines of code and you can get going. Now, if this design feels familiar, it's because it kind of feels like Teenage Engineering. And if you're not familiar with that company, they have the most incredible designs, at least in my opinion. And they are partnering with Teenage Engineering on this, holding a competition for the best text-to-speech creations. And the top three most creative ones will win a Teenage Engineering OB4. And here's what that looks like. This is $550. So definitely take a few minutes and go create something with OpenAI TTS. Next, Windsurf Wave 5 is here. If you've been vibe coding, you're probably familiar with Windsurf. In Wave 5, they've made big improvements to their tab completion. This is more of like the passive coding, not quite vibe coding, but more traditional coding where you're doing tab completion. And they added a lot of really cool functionality. Let me play a brief clip for you so you know exactly what you're going to get with Wave 5. Wave 5 has only one feature, Windsurf Tab. We've rolled up autocomplete, super complete, tab to jump, and tab to import into one seamless tool that can write new code, make multi-line edits on your existing code, and navigate around your files. Compared to our previous passive experience, Tab is a leap in quality and speed. The good news is that everybody, including free users, get unlimited Windsurf Tab completion. So really cool updates if you're doing more traditional coding. Next, Krea AI has released a big update. Now they have video training, which gives you much more control over your AI video creations. You can train the WAN 2.1 model on your own videos, giving it the ability to learn specific styles, specific objects, and even specific motions. Once that training is done, you can create new AI videos based on that style. So really cool update. Try it out. Next, Notebook LM got a big update. Notebook LM will now generate a mind map based on the documents that you provided in addition to the podcasts that you're familiar with. A great way to learn, a great way to explore the knowledge that is created by the documents that you're giving it. And even Jimmy Apples is impressed. Not bad at all. Nice. Next, Hun Yon, I hope I'm pronouncing that correctly, has announced a major upgrade to their 3D modeling AI. So we are thrilled to announce a major upgrade to our open source 3D generation model, including two groundbreaking new versions, 3D 2.0 MV multi-view generation and 3D 2.0 mini. These are open source, you can download them, you can play around with them right now. So this is great if you want to create 3D characters for motion graphics, for a game, for videos, whatever you want. Definitely check this out. Really powerful stuff for creators. And speaking of creative AI, Stability AI released some new features. They now have stable virtual camera. This allows you to upload 2D images and create 3D immersive video from a simple 2D image. It's pretty incredible. This takes us one step closer to entire movies, entire TV shows being able to be created by AI by anybody. So with this capability, you upload a 2D image and you can do things like zoom out, you can move around. And again, these are all 2D images, nothing more. You can download the weights right now. It is open source and it is free to use for non-commercial use. Next, Gemini finally adds Canvas. You can now write code with Gemini and run it directly in the browser, very similar to what Claude and ChatGPT can already do. So write HTML or JavaScript code, edit it directly in the browser, run it directly in the browser and iterate and vibe code away. And a company that I did not think was working on AI, LG just released an open source thinking model that you can download and run right now. It is called Exo 1 Deep, next generation AI model designed to enhance reasoning capabilities. And specifically, they focused on agentic AI. It comes in three versions. They have a 32 billion parameter version, which achieved number one on the AIMI benchmark, outperforming competitors at just 5% of its model size and a 7.8 billion and 2.4 billion that are more appropriate for local usage. So look at this, you can see some of the comparable models right here. Here's Exo 1 Deep 32B versus Deep Seek R1, the full version. And what do we see? It is very comparable across the benchmarks. So download it, try it out, let me know what you think. So that's all the news for today. If you enjoyed this video, please consider giving a like and subscribe and I'll see you in the next one.