Accelerating discoveries with new AI tools, powerful chips — and turkeyAIAI HardwareExploratory Science

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:

• A new tool for accelerating the discovery of new materials
• IBM Granite has new experimental features for developers to test
• How to test a new AI chip
• You have Thanksgiving turkey to thank for LASIK
• News bytes
• Research Roundup
• This week's big question

A new tool for accelerating the discovery of new materials

Chemistry patent documents can be difficult to search because they rarely include the names of molecules. A new, open-access database from IBM Research lets users search by molecular structure images, which will help scientists and businesses stay apprised of the competitive landscape and devise new molecules.

⚒️ Read the blog here

IBM Granite has new experimental features for developers to test

Traditional software has a fast development cycle, with updates sometimes appearing in a matter of days. AI model development, however, moves at a much slower pace. New versions of models sometimes can take a year to come out. To speed up the process, IBM Research has launched a new playground called Granite Experiments for developers to test new ideas before a new Granite model is released. Researchers released two open-source experimental LoRAs designed to give developers more control over AI content generation. These adapters can detect hallucinations in RAG applications and estimate the uncertainty of a model's response. If these capabilities prove useful, they could be incorporated into the next version of Granite.

Read the blog here

How to test a new AI chip

Take a behind-the-scenes look at JohnDavid Lancaster's lab in Yorktown Heights, where he and his team test and validate the "power budget" - a "wattage-to-inference" ratio - of IBM's latest AIU (Artificial Intelligence Unit) chips, which were specifically designed and optimized for AI applications.

You have Thanksgiving turkey to thank for LASIK

Thanksgiving is here again in the US, and for most people, that means a week of turkey sandwiches, turkey hash, and turkey in too many other configurations. But for a few IBM Research scientists back in the early 80s, it meant something to shoot a laser at.

The team had been using a new excimer laser to test out whether they could use lasers to etch out circuit boards, but they guessed it may be accurate enough to use on organic material - without damaging anything around where the laser was pointed. To test out their curiosity, the researchers used their laser on the turkey bones from their Thanksgiving leftovers. Their intuition was correct: there was no damage beyond where they shot. A few years later, their findings inspired a group of ophthalmologists to use these lasers in corrective eye surgery, which eventually became LASIK. So don't be so quick to throw away your leftovers - you never know what they might inspire!

Read the blog here

News bytes

IBM at the Grammys

Check out Jonathan Adashek's post to learn more about how The Recording Academy is using watsonx AI to transform the digital experience for members and music fans.

LEGO Z16 Mainframe

Take a guided tour of the impressive LEGO Z16 mainframe. Discover its intricate features, innovative design, and the creativity that brings this remarkable structure to life.

IBM Research History

On November 21, 1966, IBM researchers announced a new advancement in chipmaking. They created experimental memory cores that were slightly wider than a human hair, using techniques similar to candle making. Check out "This Day" in IBM Research History to learn more about this milestone.

Research Roundup

Highlighting new publications from IBM researchers that we liked the sound of:

• A Mamba-Based Foundation Model for Chemistry
• Agnostic Causality-Driven Enhancement of Chemical Foundation Models on Downstream Tasks
• Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors

This week's big question

What's the funniest LLM hallucination you've ever seen?

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.