Jul 26 (edited) in AI News
What a week for AI 🤯
- Meta releases Llama 3.1: Meta has unveiled its latest open-source AI model, Llama 3.1, with sizes up to 405 billion parameters. This model rivals other leading AI models and includes a context length of 128K tokens and support for eight languages​. Llama 3.1 is free and open source. It took MONTHS to train on 16,000 NVIDIA H100 GPUs, which likely costed hundreds of millions of dollars and used enough electricity to power a small country. According to benchmarks the model would beat ChatGPT-4o and Claude´s Sonnet. It comes in 3 sizes: 8B, 70B and 405B. Where B refers to Billions of parameters (the number of variables the model can use to make predictions). It is a relatively simple decoder only transformer. The code used to train this model is only 300 lines of python and PyTorch along with a library called FairScale used to distribute training across multiple GPUs. We will have a detailed post about the implementation and how the model work on the AI Quest. You can try the model for free on platforms like meta.ai, groq or NVIDIA´s playground.
2. OpenAI launched "SearchGPT": a new search engine aiming to revolutionize the way we find and synthesize information online. Unlike traditional search engines that provide a list of links, SearchGPT offers a more interactive and streamlined experience.
Key Features of SearchGPT:
- Chat Interface: Similar to ChatGPT, SearchGPT presents search results within a chat window. This setup allows for a more conversational and intuitive interaction with the search engine.
- Multimedia Search: In addition to text-based queries, SearchGPT supports image searches and includes various widgets for weather, calculators, sports updates, financial information, and time zones​.
- Summarization Capabilities: The search engine can summarize web pages in up to 300 characters, making it easier for users to quickly grasp the essential information without navigating through multiple links.
- Advanced Language Models: SearchGPT uses advanced language models such as GPT-4 Lite, GPT-4, or GPT-3.5. This ensures high-quality responses and improved comprehension of complex queries.
- Partnerships with Publishers: OpenAI has secured agreements with several publishers, including the Financial Times, allowing full news articles to be displayed directly within SearchGPT​.
OpenAI CEO Sam Altman emphasized that the goal is not to create a Google clone but to find a better way to help users synthesize and act on information efficiently. This approach aims to address the limitations of existing search engines by offering more relevant and concise answers, thus enhancing the overall user experience. This is also the mission of Perplexity..
3. Mistral AI Releases Mistral Large 2: Mistral AI has unveiled its latest language model, Mistral Large 2, designed to compete at the forefront of AI technology. This new model boasts significant advancements over its predecessor, enhancing performance across various domains including code generation, mathematics, and reasoning. Here are the key features and improvements:
- Enhanced Multilingual Support: Mistral Large 2 excels in multiple languages, including English, French, Spanish, German, and Italian. This makes it highly effective in generating nuanced and contextually accurate responses in various languages​.
- Extended Context Window: With a 32K token context window, the model can handle extensive documents more effectively, allowing for better information recall and context management.
- Advanced Function Calling: The model includes sophisticated function calling capabilities, which facilitate more interactive and advanced application development. This feature is particularly useful for developers aiming to integrate the model into complex workflows​.
- Competitive Performance: In benchmark tests, Mistral Large 2 has shown strong results, positioning itself as a top contender against leading models like GPT-4. It has demonstrated superior performance in reasoning, coding tasks, and mathematical problem-solving​.
- Availability and Access: Mistral Large 2 is available through various platforms, including Mistral AI's own hosting service "La Plateforme", and it is also accessible via Microsoft Azure, making it easy for developers to incorporate it into their existing systems​.
4. DeepMind AI earns silver medal at the International Math Olympiad: DeepMind's AI demonstrated exceptional mathematical problem-solving abilities, earning a silver medal at the International Math Olympiad, highlighting the growing competence of AI in complex academic fields​
5. Researchers at MIT have developed an AI model that can identify breast tumor stages likely to progress to invasive cancer. This advancement could significantly improve how clinicians assess and treat breast cancer, potentially reducing overtreatment​.
0
0 comments
What a week for AI 🤯
Leaderboard (30-day)
1
+1
powered by