DeFi: Shaping the Future of Finance

Explore how DeFi Technologies Inc. (CAD: DEFI & US: DEFTF) bridges traditional finance and the $3T digital asset market. With innovative strategies and global expansion, DeFi is redefining the investment landscape. Gain exposure to Bitcoin, Web3, and beyond with regulated, secure solutions.

Read The Report

Today:

• Google Reclaims Top AI Developer

• OpenAI Upgrades ChatGPT for Mac

• Google Launches Gemini 2.0 Flash

• Perplexity Valuation Triples to $9bn

• Google DeepMind Collaborates with Apptronik

• Ex-Twitch CEO Launches Stem AI

Google AI DESTROYS OpenAI... Google Deepmind is UNSTOPPABLE…

Google has reclaimed its position as the leading AI developer this week. Their latest tools, Veo 2 for creating high-quality videos and Imagen 3 for generating detailed images, are better than any others available. These models can understand and follow complex instructions, making their outputs more accurate and visually appealing. 

Additionally, Google's Gemini language models now top the rankings, outperforming OpenAI’s GPT-4. They introduced Project Astra, an AI assistant that helps with everyday tasks like cooking and navigating, and Project Mariner, which lets AI interact with websites like a human.

WATCH THE VIDEO ON YOUTUBE

OpenAI Upgrades ChatGPT for Mac

OpenAI has upgraded its ChatGPT desktop apps for macOS and Window, enhancing their ability to connect with programs like Xcode, Notion, and Apple Notes.Users can now use voice commands and connect ChatGPT with various applications securely. These improvements let ChatGPT automate tasks and understand document content without users needing to do it manually. 

Earlier updates included phone access to ChatGPT, making it usable without internet, and advanced search features for all users. OpenAI also introduced "Projects" to organize chats and launched a smart video tool called Sora. These updates aim to make ChatGPT more versatile and user-friendly for everyone.

Google Launches Gemini 2.0 Flash

Google has launched Gemini 2.0 Flash Thinking, a new artificial intelligence model designed to compete with OpenAI’s offerings. This advanced AI can handle both text and images, making it versatile for various tasks. It processes large amounts of information quickly and provides clear, step-by-step explanations for its answers, helping users understand how it reaches conclusions. 

Gemini 2.0 is free to use on Google AI Studio and has been rated the top model by independent reviewers. Although it currently doesn’t integrate with Google Search or other tools, its ability to combine different types of data and offer transparent reasoning makes it a strong contender in the AI market.

Perplexity Valuation Triples to $9bn

Perplexity, an AI-driven search engine, tripled its valuation to $9bn in its latest funding round, aiming to compete with Google and OpenAI. The $500mn round was led by Institutional Venture Partners, with participation from Nvidia, New Enterprise Associates, B Capital, and T Rowe Price.

Perplexity, based in San Francisco, has rapidly grown this year, attracting 15mn monthly active users mainly from the US and receiving hundreds of millions of queries a month. The funding will enable Perplexity to compete for AI engineering talent, enhance its search advertising system, and stay competitive in the AI industry.

Google DeepMind Collaborates with Apptronik

Google DeepMind and Apptronik have joined forces to develop AI-powered humanoid robots capable of working alongside humans in dynamic settings. Apptronik, renowned for creating robots like NASA's Valkyrie and the current Apollo model, brings extensive robotics expertise to the partnership. Google DeepMind contributes its advanced artificial intelligence, including the Gemini 1.5 Pro model, which enables robots to navigate complex environments by learning from simple smartphone videos. 

These robots can identify small objects and understand tasks, such as finding a suitable area to draw on a whiteboard. This collaboration aims to enhance robot adaptability and functionality in industrial workplaces. Meanwhile, OpenAI is also advancing its own robotics initiatives.

Ex-Twitch CEO Launches Stem AI

Emmett Shear, the former CEO of Twitch, is launching a new AI startup called Stem AI, supported by the venture firm Andreessen Horowitz (a16z). Stem AI aims to develop artificial intelligence that understands and aligns with human behavior, preferences, biology, morality, and ethics. 

One of the co-founders, Adam Goldstein, previously founded the travel site Hipmunk and has experience in biological research. Shear has voiced concerns about AI safety and the need for better AI regulation. Stem AI is currently operating quietly, with more details expected in the future. 

🧠RESEARCH

No More Adam: Learning Rate Scaling at Initialization is All You Need

SGD-SaI, an improved training method for deep learning that adjusts learning rates at the start based on each parameter’s signal quality. This approach matches or outperforms the popular AdamW optimizer while using half the memory. It works effectively across various models and tasks, making training more efficient.

Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference

ModernBERT, an enhanced version of the BERT model for language tasks. Trained on extensive data and able to handle longer inputs, ModernBERT achieves top performance in various tasks while using less memory and running faster. It is optimized to work efficiently on standard graphics cards.

TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks

TheAgentCompany, a test to evaluate AI agents on real work tasks like browsing the web and writing code. They found that the best AI could handle 24% of these tasks on its own. This shows AI can manage simpler jobs but still struggles with more complex, long-term projects.

RAG-RewardBench: Benchmarking Reward Models in Retrieval Augmented Generation for Preference Alignment

RAG-RewardBench, a new test to evaluate how well AI models align their responses with human preferences. They created challenging scenarios and used AI judges to assess 45 models. Findings show current models struggle with preference alignment, highlighting the need for improved training. The benchmark is publicly available.

LLaVA-UHD v2: an MLLM Integrating High-Resolution Feature Pyramid via Hierarchical Window Transformer

LLaVA-UHD v2, an improved AI model that better understands images by using detailed layers of visual information. This model outperforms existing systems on various tests, providing clearer image interpretation and more accurate language responses. It is efficient and freely available for further research.

🛠️TOP TOOLS

MathGPT - AI math solver and calculator that provides instant solutions and step-by-step explanations for a wide range of mathematical problems

Imagica - No-code AI development platform that empowers users to create AI applications quickly and easily

Photosonic - AI-powered art generator developed by Writesonic that transforms text descriptions into unique digital images.

HitPaw Watermark Remover - AI-powered tool designed to efficiently remove watermarks, logos, text, and other unwanted elements from both images and videos.

v0 dev - Generative user interface system developed by Vercel Labs that leverages AI to create React code compatible with Shadcn UI and Tailwind CSS. 

📲SOCIAL MEDIA

🗞️MORE NEWS

• Microsoft is testing live translation on Intel and AMD Copilot Plus PCs. Windows 11 Insiders can now translate audio from over 44 languages into English subtitles in real time, supporting videos, calls, and recordings.

• Patronus AI unveils Glider, a compact 3.8 billion-parameter model that rivals GPT-4 in evaluating AI systems. Glider provides real-time, on-device assessments with detailed explanations, enhancing cost-efficiency, privacy, and accessibility for developers.

• Instagram will launchAI video editing tools next year using Meta’s Movie Gen model. Users can modify almost any video aspect with simple text prompts, such as changing outfits, backgrounds, and appearances, making content creation easier without advanced skills.

• Researchers developed FastGlioma, an AI tool that accurately detects residual brain tumors during surgery in seconds. Faster and more precise than current methods, it operates on-device, ensuring privacy and improving surgical outcomes for brain cancer patients.

• Abu Dhabi’s Technology Innovation Institute launches Falcon 3, powerful open-source AI models trained on 14 trillion tokens. Outperforming competitors on single GPUs, Falcon 3 offers various sizes and languages, easy integration, and a free chatbot, with multimodal features planned for 2025.

• OpenAI is facing an outage disrupting email verifications for ChatGPT account signups and logins for several hours. However, new users can still register using Google, Microsoft, or Apple accounts, and some can log in successfully after multiple attempts.

• TuSimple rebrands as CreateAI, shifting from autonomous trucking to AI animation and gaming amid plans to move U.S. assets to China. Shareholders oppose the pivot, citing conflicts of interest and pushing to liquidate the company.