Published on

This Week in AI: New Models, Features, and Education

Authors
  • Name
    Callum van den Enden
    Twitter

Overview

A look into the advancements and challenges within the world of Artificial Intelligence and its impact on the world. Some cool new tech is emerging.

The Push for Smarter AI

OpenAI and its competitors are facing hurdles in scaling up large language models (LLMs). The traditional approach of "bigger is better" seems to be hitting a plateau. Researchers are now exploring "test-time compute," which focuses on enhancing AI models during the "inference" phase, when the model is actually being used. This allows for more complex reasoning, similar to human thought processes. This shift has significant implications for the AI hardware landscape, potentially impacting demand for Nvidia's chips and opening doors for competitors in the inference market. (Reuters)

  • My Take: This is a fascinating development. It highlights that simply throwing more data and computing power at a problem isn't always the solution. The focus on inference is a clever way to squeeze more performance out of existing models, potentially leading to more efficient and effective AI.

New AI Models Emerge

Several companies are releasing new AI models tailored for specific industries. Microsoft, in collaboration with partners like Bayer, Cerence, and Rockwell Automation, is introducing adapted AI models designed to address unique industry needs. These models, trained on industry-specific data, aim to improve everything from crop protection to automotive digital assistants and manufacturing processes. (Microsoft Blog) Los Alamos National Laboratory has developed EPBDxDNABERT-2, a multimodal deep learning model that improves the prediction power for genomics related to disease. This model leverages DNA breathing dynamics to understand the relationship between transcription factors and gene activities, offering potential for drug development. (Phys.org)

  • My Take: The trend towards specialized AI models is a game-changer. It allows businesses to leverage AI in ways that are directly relevant to their operations, leading to more tangible benefits and ROI. The Los Alamos model is a prime example of how AI can unlock breakthroughs in scientific research.

A Graph-Based Approach to Innovation

MIT researchers have developed a graph-based AI model that maps the future of innovation by connecting seemingly disparate fields. This model, inspired by category theory, analyzes relationships between objects and their interactions, allowing for a deeper understanding of complex systems. The model was used to analyze scientific papers on biological materials and even found similarities between these materials and Beethoven's 9th Symphony. It also suggested a new mycelium-based composite material inspired by a Kandinsky painting. (MIT News)

  • My Take: This is a truly innovative approach to using AI - I actually had a dream to build something similar when considering starting YouQ. Having the ability to explore related concepts and glean new information is spectacular. Harks back to the idea that there are no new ideas, just combinations of old ones leading to something new. It demonstrates the potential of AI to not just analyze data but to inspire creativity and drive innovation in unexpected ways. The ability to connect seemingly unrelated fields opens up exciting possibilities for cross-disciplinary research and development.

I would love to use a graph to explore the relationships between different fields and ideas. I'm not sure if it's possible to do this with OOTB LLM's/vector embeddings beyond a surface level semantic similarity, but it would be a fascinating experiment. LLMs are really good at this kind of thing; spotting patterns in text is common, but spotting patterns in ideas is a super interesting case I've been exploring.

Microsoft Tailors AI for Every Industry

Microsoft isn't just building general-purpose AI; they're laser-focused on tailoring it to specific industries. They've unveiled "adapted AI models," trained on industry-specific data, in collaboration with partners like Bayer (agriculture), Cerence (automotive), Rockwell Automation (manufacturing), and Siemens (software). These models, accessible via the Azure AI model catalog, promise to solve unique industry challenges. (Microsoft Blog)

My take: This is a smart move by Microsoft. Generic AI is cool, but tailored AI is where the real value lies. By focusing on specific use cases, they can deliver more impactful solutions for businesses.

OpenAI Rethinks Its Approach

OpenAI, the creators of ChatGPT, is facing the reality that simply making models bigger isn't always the answer. They've acknowledged that scaling up pre-training has hit a plateau. Now, they're exploring "test-time compute," a technique that enhances models during the "inference" phase (when the model is actually being used). Their new o1 model showcases this approach, focusing on more human-like reasoning and using expert feedback. (Reuters, The Information)

My take: The AI world is evolving rapidly. OpenAI's shift reflects a broader trend: we're moving beyond brute-force scaling and towards more nuanced, efficient AI development. This could also shake up the AI hardware market, as inference requires different types of chips.

AI in Education

New AI tools are being marketed as study aids for students, raising concerns among educators. While some see these tools as a way for students to offload the hard work of learning, others recognize their potential benefits, particularly for neurodivergent students. The accuracy of AI summarizing tools is also a concern, due to the potential for hallucinations. (EdSurge)

  • My Take: The use of AI in education feels like an obvious use case - after all, it's a tool that can help students learn faster and more effectively. That said, there are genuine issues with the accuracy of AI tools. In my experience with AI in education, I've found it's great for exploring ideas, and tailoring lessons to learners' existing knowledge to improve knowledge acquisition when grounded with quality source data (shameless plug for my own AI-powered learning platform), but it's not great where precise answers are important (yet), like medicine, science, and math.

Quick Hits

  • DataRobot: Unveils new AI app-building capabilities, emphasizing solutions over dazzling tech specs. (CRN)
  • AI Stocks: The boom continues, with Meta shining and Microsoft facing some headwinds. The focus is shifting from hype to tangible results. (Investors.com)
  • Salesforce: Plans to hire over 1,000 workers to support the growth of its new generative AI product, Agentforce. (Yahoo Finance)
  • 8x8: Adds new AI-powered transcriptions and customer engagement features to its cloud CX platform. (CX Today)
  • Final Cut Pro: Gets a major update with AI masking tools, automated captions, spatial video editing, and more. (The Verge)
  • US News: Lists its top 10 AI stocks to buy, highlighting major players like Microsoft, Nvidia, Google, and Amazon. (US News)