AI advancements show specialized performance in carb counting and efficient decision modeling

Recent developments in artificial intelligence highlight both the limitations of large language models in specialized tasks and the rise of highly efficient, niche decision models.

In a test regarding carbohydrate counting for diabetes management, Anthropic’s Claude Fable 5.1 and OpenAI’s Astra demonstrated varying levels of accuracy. While Fable 5.1 averaged an error of 7g per meal—comparable to human performance in some studies—Astra averaged a 13g error. The testing noted that these models do not replicate human error patterns, such as the tendency to undercount.

Separately, the Brazilian lab Supersonic Labs released Julia 1, a 144.3-million-parameter decision model. Unlike generative models, Julia 1 is designed to select labels or scores from provided options without generating text. The model was trained for approximately US$104 and is capable of running on low-power hardware like tablets or older laptops. In benchmarks, it outperformed existing reference models in categories such as emotion classification and typed decisions.