Development and Cost Efficiency of Reasoning AI Model
The development of reasoning Artificial Intelligence (AI) models has become more feasible and cost-effective. The NovaSky research team at UC Berkeley’s Sky Computing Lab launched an open source reasoning model, Sky-T1-32B-Preview, that competes with earlier versions of OpenAI’s o1 in key performance indicators.
Open Source Model: Sky-T1-32B-Preview
The remarkable aspect of Sky-T1 is that it’s the first fully open source reasoning model that can be replicated from scratch, courtesy of the shared dataset and training code. The training of this model resulted in less than $450, debunking the notion that the development of high-level reasoning capabilities is expensive and inefficient.
Comparative Performance and Cost
Just until recently, millions were spent on training similar performance models. However, utilising synthetic data significantly reduced these costs.
Reasoning AI Model’s Reliability and Efficiency
Reasoning models, unlike their counterparts, fact-check themselves to bypass common pitfalls that often obstruct other models. Although slightly slower, these models prove to be more reliable in physics, science, and mathematics.
Sky-T1’s Competitive Performance
Sky-T1 outperforms an initial version of o1 in handling “competition-level” math challenges and a set of difficult problems from a coding evaluation. However, it slightly trails behind the o1 model on GPQA-Diamond, containing advanced science-related queries.
The NovaSky team claims Sky-T1 as the maiden step in their journey towards developing open source models with superior reasoning capabilities with a focus on efficiency and accuracy in forthcoming models.
Fonte original: Leia a matéria completa no TechCrunch
