# I Tested This Week’s AI Tools (And Broke Them)

Source: https://www.youtube.com/watch?v=BCW-wrDzWCs
Recap page: https://rapidrecap.app/video/BCW-wrDzWCs
Generated: 2025-07-19T14:37:29.283+00:00

---
## Quick Overview

This week's new AI tools, including ChatGPT's agent mode, InVideo's AI Twin, Hume AI's personality cloning, Runway's Act-Two motion capture, Decart's MirageLSD, Google's AI business calling, Mistral's deep research and audio models, and Amazon's Kiro IDE, demonstrate impressive capabilities but often exhibit significant bugs, slowness, or incomplete functionality upon release, indicating they are not yet fully polished for practical, complex tasks.

**Key Points:**
- ChatGPT's new agent mode, designed for complex multi-step tasks, failed to complete a detailed date night planning request after 16 minutes and produced a basic, outdated presentation for a website.
- InVideo's AI Twin feature successfully created a digital avatar of the presenter and integrated it into a product advertisement, demonstrating its utility for quick video creation and reshoots.
- Hume AI's personality cloning, which aims to mimic speaking style, resulted in an AI that talked excessively and did not perfectly replicate the user's voice, despite its quick cloning process.
- Runway's Act-Two motion capture model showed impressive facial animation but struggled with full-body tracking, leading to visual glitches like extra limbs on animated characters.
- Decart's MirageLSD is a novel live-stream diffusion AI model that reskins live video in real-time based on text prompts, offering various stylistic transformations, though it still exhibits visual artifacts.
- Google's new AI-powered business-calling feature, designed to make calls on users' behalf, was not available for the presenter to demo, indicating limited public access at launch.
- Mistral AI released Voxstral, an open-source audio model capable of transcribing up to 40 minutes of audio for a very low cost (0.0001 cents per minute), making it a highly affordable transcription solution.

![Screenshot at 5:07: A text overlay showing 'Stopped after 16 minutes' with a crying emoji, indicating the failure of ChatGPT Agent's complex task.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-05-07.png)

**Context:** In a rapidly evolving AI landscape, new tools are constantly emerging with ambitious claims. The presenter aims to cut through the hype by personally testing these latest AI advancements, providing honest feedback on their current functionality and practical utility. This week's focus is on a diverse range of tools, from autonomous agents to voice cloning and real-time video manipulation, highlighting both their impressive potential and their current limitations.

## Detailed Analysis

This week saw a flurry of new AI tool releases, with the presenter testing several to evaluate their real-world performance. OpenAI's new ChatGPT agent mode, designed to perform complex tasks like booking reservations or creating presentations on your computer, proved to be slow and buggy, failing to complete a multi-step date night planning task after 16 minutes and producing poorly designed, outdated slides for a website presentation. OpenAI also released a record mode for ChatGPT Plus users on macOS, allowing audio recording for summaries. InVideo introduced an AI Twin feature, enabling users to create digital clones or use pre-made avatars for videos, including product integration, which worked well for a Glassy-O's cereal ad. Hume AI launched personality cloning, aiming to mimic not just voice but also speaking style; however, the demo showed the AI talking excessively and not perfectly matching the user's voice. Runway's Act-Two, a next-generation motion capture model, demonstrated impressive facial tracking but struggled with full-body animations, producing glitches like extra limbs. Decart's MirageLSD, a live-stream diffusion AI video model, showcased real-time video reskinning based on text prompts, offering various visual styles like 'Frozen Wonderland' and 'BlockCraft,' though it also displayed visual artifacts. Google rolled out an AI-powered business-calling feature, intended to make calls on users' behalf, but the presenter was unable to access this feature for a live demo. Mistral AI introduced a 'deep research' mode for its Le Chat chatbot, offering 5 free deep searches per month, and also released Voxstral, an open-source, highly affordable audio-to-text transcription model. Amazon launched Kiro, an AI-powered IDE designed for project planning and coding, which aims to streamline development by mapping out projects before coding. Overall, while these tools highlight rapid advancements, many are still in early stages, requiring further development to achieve seamless, reliable performance.

### OpenAI's ChatGPT Agent

- Can perform complex tasks on your computer
- Sam Altman warned about risks with sensitive data
- Failed to complete a multi-step date night planning task after 16 minutes
- Generated a poorly designed, outdated presentation for a website
- Agent mode is currently only available to Pro users ($200/month) but will roll out to other paid plans.

### OpenAI's Record Mode

- Now available for ChatGPT Plus users globally on macOS desktop app
- Records audio from computer (Zoom, YouTube, Google Meet)
- Provides detailed summaries after recording.

### InVideo AI Twin

- Allows uploading videos to create digital clones or using pre-made avatars
- Can integrate products into videos for advertising
- Successfully created a video of the presenter's avatar talking about 'Glassy-O's Glass Shard Cereal'
- Useful for last-minute reshoots, simple ads, and creative fun
- Free trial available at invideo.io.

### Hume AI Personality Cloning

- Aims to mimic voice, speaking style, and language
- Requires 30-90 seconds of audio for cloning
- AI talked excessively and didn't perfectly match the user's voice in the demo
- User found the AI's constant talking annoying and not sounding like him.

### Runway Act-Two

- Next-generation motion capture model for head, face, body, and hand tracking
- Requires a driving performance video and reference character
- Facial motion tracking was impressive
- Full-body animation showed glitches like extra limbs and floating objects.

### Decart MirageLSD

- First live-stream diffusion AI video model with zero latency
- Reskins live video feed in real-time based on text prompts
- Offers various visual styles like 'Frozen Wonderland,' 'Versailles Palace,' 'BlockCraft,' and 'Midnight'
- Can be controlled by typing world ideas in real-time
- Demo showed visual artifacts and pixelation.

### Google's AI Business Calling

- AI-powered feature to call local businesses on your behalf for information
- Aims to help users access info without picking up the phone
- Feature was not available for the presenter to demo live.

### Mistral AI Updates

- Le Chat chatbot now has a 'deep research' mode (10x speed, 5 free searches/month)
- Released Voxstral, its first open-source AI audio model
- Voxstral can transcribe up to 30-40 minutes of audio
- Voxstral costs $0.0001 per minute, making it very affordable
- Mistral models are competitive with OpenAI and Google models on benchmarks.

![Screenshot at 0:20: OpenAI's ChatGPT Agent landing page showing 'Introducing ChatGPT agent: bridging research and action'](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-00-20.png)
![Screenshot at 0:40: ChatGPT Agent's capabilities listed, including calendar briefing, ingredient planning, and competitor analysis.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-00-40.png)
![Screenshot at 0:50: ChatGPT Agent's ability to navigate websites, filter results, run code, and deliver editable slideshows/spreadsheets.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-00-50.png)
![Screenshot at 1:50: ChatGPT Agent's 'Introducing agent mode' pop-up, highlighting its ability to work on its own computer for research, spreadsheets, and bookings.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-01-50.png)
![Screenshot at 2:42: The detailed prompt given to ChatGPT Agent for date night planning, clothing purchase, and birthday gift ideas.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-02-42.png)
![Screenshot at 3:23: ChatGPT Agent searching the web for San Diego restaurants on OpenTable, showing its virtual browser activity.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-03-23.png)
![Screenshot at 5:07: ChatGPT Agent's complex task failed, showing 'Stopped after 16 minutes' with a crying emoji, indicating frustration.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-05-07.png)
![Screenshot at 5:59: ChatGPT Agent's generated presentation slides for 'Advertising with FutureTools.io', showing basic design and content.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-05-59.png)
![Screenshot at 6:32: ChatGPT Agent's generated traffic overview chart for FutureTools.io, showing monthly visits in thousands.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-06-32.png)
![Screenshot at 7:10: ChatGPT Agent's generated 'Future Tools Newsletter' and 'YouTube Channel' statistics, some of which are outdated.](https://ss.rapidrecap.app/screens/BCW-wrDzWCs/00-07-10.png)