AI: 1 in 3 VC Dollars

Sign Up | Advertise

The AI World Today

• Depth Pro Revolutionizes AI

• Meta Unveils Movie Gen Model

• China Develops TeleChat2-115B

• AI Startups Dominate VC Funding

  +

• Heads Up

• AI Solution

Apple’s Depth Pro Elevates AI Depth Mapping

Apple’s AI team has introduced Depth Pro, a groundbreaking model for monocular depth estimation. It generates detailed 3D depth maps from single 2D images in just 0.3 seconds, without needing traditional camera data. Depth Pro excels in precision and speed, producing sharp 2.25-megapixel depth maps, even capturing fine details like fur and vegetation. It also supports zero-shot learning, making it adaptable across different scenarios without extensive training. With real-world applications in AR, e-commerce, and autonomous vehicles, Depth Pro offers metric depth, enabling accurate real-world measurements. Open-sourced on GitHub, the model is set to impact industries requiring spatial awareness, from robotics to healthcare.

Meta’s Movie Gen Creates Realistic AI Videos

Meta's new Movie Gen model generates realistic videos with sound from text prompts, aiming for practical, natural-language-driven video creation. The model produces up to 16 seconds of video at 768 pixels wide, upscaled to 1080p, and supports basic text-based editing for adjustments like changing backgrounds or outfits. Audio is added to match video content, such as engine noises or music, though voice generation is not included. Movie Gen was trained on a mix of licensed and publicly available data, though specifics are not disclosed. While not publicly released due to potential deepfake concerns and technical limitations, Meta highlights safety and practicality as priorities, with the model serving as a research concept for now.

TeleChat2-115B Trained Using Chinese Infrastructure

TeleChat2-115B, a 100-billion-parameter AI model from China Telecom's AI Research Institute, was trained entirely on domestic infrastructure. China Telecom claimed that the model was trained using 10 trillion tokens of Chinese and English data and is compatible with Huawei’s Ascend Atlas 800T A2 training servers, powered by Kunpeng 920 processors based on the Arm 8.2 architecture. Though smaller in scale compared to models like Llama or OpenAI's latest, TeleChat2-115B's parameter count suggests it required less computational power. Despite lacking access to cutting-edge GPUs, China Telecom leverages its vast resources to continue advancing AI technology. The model has been open-sourced on GitHub.

AI Startups Secure One-Third of VC Dollars

A new CBInsights report reveals that AI startups have secured 1 in 3 venture capital dollars in Q3 2024, highlighting AI’s dominance in the VC landscape. Despite a cautious investment environment, the average deal size has grown to $13.9 million, with AI startups capturing 31% of all funding, the second-highest on record. Silicon Valley remains a key AI hub, with AI companies exiting faster than those in other sectors. Notably, Safe Superintelligence (SSI), an early-stage AI startup, raised a $1 billion round, underscoring investor confidence. Despite bullishness, many AI startups may struggle to meet high expectations, while recent tech IPOs have shown positive performance, potentially encouraging more public listings.

Heads Up

Meta has released CoTracker 2.1 on Hugging Face, an enhanced Transformer-based model for video motion prediction, capable of tracking 70,000 points on a single GPU.

WiLoR, a cutting-edge system for real-time 3D hand localization and reconstruction, has just been released. Trained on 2M+ images, it outperforms previous methods with its transformer-based model and refinement module. 

Google DeepMind and BioNTech are collaborating to develop AI lab assistants, helping researchers plan experiments and predict outcomes, aiming for specialized applications of AI in science.

Google's Gemini chatbot is now available for select Gmail users on iOS, allowing Google One AI Premium and Workspace subscribers to chat about their inbox within the app.

Cohere unveiled updates to its fine-tuning service, enhancing flexibility and transparency for enterprises using its latest Command R 08-2024 model, accelerating AI model customization and adoption.

AI Solution

MIT’s AI Tool Enhances Future Self-Continuity, Wellbeing

"Future You" is an AI-powered intervention by MIT researchers designed to improve future self-continuity, helping users feel more connected to their future selves. Developed through research on mental health and wellbeing, the system allows users to chat with a virtual version of their future self, personalized based on their goals and qualities. The AI-generated "Future You" character, combined with an age-progressed image and a unique synthetic memory, creates a realistic, relatable conversation about the user's potential life at age 60. After a brief session, users reported reduced anxiety and increased future self-continuity. The intervention is based on a large language model, personalized through a pre-intervention survey, and aims to support long-term thinking and mental health improvements.