The path to more powerful — and efficient — AI systems
Welcome to The Short, IBM Research's weekly recap of the latest innovations in AI, quantum computing, semiconductors, and the cloud. If you're liking what you see here, be sure to sign up for earlier access on LinkedIn.
In this week's edition:
- Charting a path to a more energy-efficient AI future
- Rapidly unlocking geospatial insights with IBM AI chips
- A new tool to unlock data from enterprise documents for generative AI
- IBM Quantum delivers on performance challenge made two years ago
- Learn how to build AI Agents with Bee
- Research Roundup
- This week's big question
While AI is revolutionizing the world, AI models are becoming increasingly complex, and costly to run. IBM researchers have been focused on developing new AI hardware architectures that could meet the performance requirements of today’s AI models. At the 2024 AI Hardware Forum at Yorktown Heights, IBM Semiconductors General Manager and Vice President of Hybrid Cloud Research Mukesh Khare outlined several research paths that are part of the IBM Research AIU family of chips.
These family members are in different stages of maturity, and represent the various ways that IBM is thinking about the future of chip design for AI — not just for tomorrow, but for years to come.
Our climate is changing faster than ever before, and being able to understand how it’s changing, and what we can do to mitigate issues, is key to our future. It’s partly why IBM Research has been working with NASA and the The University of Alabama in Huntsville (UAH), to build weather and climate models like Prithvi WxC. And it’s also why today, IBM and UAH are collaborating to install a cluster on campus containing IBM Spyre chips to run advanced AI models like Prithvi WxC.
IBM Spyre is the first AIU production accelerator born out of the IBM Research AIU family, and is part of a long-term strategy of developing novel architectures and full-stack technology solutions for the emerging space of generative AI.
Large amounts of valuable enterprise data lies buried in PDFs, annual reports, and other business documents. Docling, IBM's new open-source toolkit, is designed to extract and process this information so that large language models can digest it. Docling converts complex documents into JSON and Markdown files that can be used to fine-tune LLMs for enterprise tasks and to ground them on trusted data via retrieval-augmented generation. Docling can run on a standard laptop and takes just five lines of code to set up.
This week at the first-ever IBM Quantum Developer Conference (QDC), IBM researchers shared that they've successfully delivered a system capable of running accurate calculations employing circuits with 5,000 two-qubit gates. The new version of the IBM Quantum Heron quantum processing unit is what drives this — powered by 156 qubits in a heavy-hex layout. The team also announced a new coupling architecture that connects two multi-qubit chips, showcasing the path towards modular quantum computing.
🐈⬛ Read more here
Interested in learning how to build your first AI agent? Check out Maya Murad and Armand Ruiz's discussion on what AI agents are, why they are important, and how you can create your first agent in minutes using the Bee toolkit.
Highlighting new publications from IBM researchers that we liked the sound of:
What kinds of documents do you want to bring to fine-tuning with Docling?
ℹ️ If you enjoyed this week's edition of The Short, make sure to sign up on LinkedIn to get it in your inbox and feed even sooner.