Intel® Liftoff Member, Moonshot Uses Intel Technology To Predict the Next Big Thing

By now, a large chunk of the world’s population is plugged into LLMs. They’re using it to write everything from blog articles to automation scripts. 

But what if you could use that processing power to look forwards as well as backwards? That’s the challenge Intel® Liftoff member Moonshot took on for this year’s Intel® Liftoff Hackathon. Their mission is to create autonomous LLMs that can accurately predict future signals - everything from how long you’ll live to GDP growth. Innovations like Moonshot’s represent the leading edge of natural language processing, finding new ways to use this technology for good. Their algorithms are already being used in a wide range of applications including skin cancer prediction, DNA sequencing, mRNA vaccine development, environmental conservation and space research.

For the Intel® Liftoff Hackathon, they focused on one question in particular. 

Can You Train LLMs To Predict Recessions From News Articles?

Recessions ruin societies, and it’s often hard to tell when one is coming. If they could forecast recessions, institutions would be in a position to plan ahead to reduce the impact, or steer the economy in the right direction before it even happens. 

But that’s harder than it sounds. Take the infamous yield curve for example. It’s a fairly reliable harbinger of recession. But it isn’t useful, since each recession has different time lags. News articles from 1851 onwards exist, but an LLM’s minimal context size of 8K or 16K limits its capabilities. Moonshot’s goal for the Intel® Liftoff Hackathon was to find a way to use news articles effectively to predict recessions.

Moonshot’s Methodology

They used the New York Times API to extract all news article titles and synopses from 1851 until the present time. Then, they carefully created a 1K human labeled CoT style dataset. This dataset could standardize news headlines to make them time invariant, and then annotate how the headline would affect the economy. 

This was essential, because LLMs are trained on past data. By removing all time-dependent information like names of places or names of events, Moonshot effectively sidestepped this limitation. With that done, they used LoRA to finetune OpenLLaMA to predict how a news headline would affect the economy.

Moonshot used Intel Max 1100 GPUs, Intel Xeon CPUs, and Intel Pytorch Extensions.

Moonshot’s Results: Turning LLMs Into Crystal Balls

Moonshot’s results show that LLMs can be finetuned carefully to infer recession signals from news articles. The trick? You have to carefully standardize time variant features from headlines. The next step is a painstaking annotation of a thousand examples that show how each headline will contribute to a recession. By using LoRA, fine tuning an LLM in this way is a breeze. 

Defining the Future of AI Tech (or the Future Itself) With Intel® Liftoff

Whatever the future holds, one thing is certain: many of the most innovative and game-changing solutions will have passed through Intel® Liftoff’s tech startup accelerator programs. We’re proud to play our part in propelling revolutionary platforms like Moonshot to new heights. If you’ve got a visionary idea, we’d love to hear from you. Take the next step and apply to Intel® Liftoff for Startups.