Features

 Banana bread has become a timeless homemade treat, loved for its comforting flavor and simple ingredients. My version, which is a twist of the recipe from Natasha’s kitchen,  is especially loved because it consistently produces a soft moist loaf with rich banana taste and just the right amount of sweetness.

Before we jump into the recipe, here is a quick history: banana bread became widely popular in the United States during the Great Depression in the 1930s. At a time when people were careful not to waste food, overripe bananas were no longer thrown away but instead used in baking.Around the same time, baking soda and baking powder became more common in households, making quick breads like banana bread easier to prepare without yeast. This amazing treat is commonly eaten now as well, and once you learn how to make it, its addictive flavor will make you not want to stop.

To begin, preheat the oven to 350°F (175°C). Prepare a standard 9x5-inch loaf pan by greasing it lightly or lining it with parchment paper to prevent sticking.

In a large mixing bowl, mash 3 very ripe bananas until smooth. The riper the bananas, the stronger and sweeter the flavor will be. Stir in ½ cup (1 stick) of melted unsalted butter, mixing until fully combined. Next, add ¾ cup granulated sugar, followed by 2 large eggs (lightly beaten) and ½ teaspoon vanilla extract. Mix everything together until the batter is smooth and well-blended.

In a separate step, add the dry ingredients directly into the wet mixture. Sprinkle in 1 teaspoon baking soda and ½ teaspoon salt, then add 1 ½ cups all-purpose flour. Stir gently until just combined. It is important not to overmix the batter, as this can make the banana bread dense instead of soft.

Pour the batter evenly into the prepared loaf pan and smooth the top. Bake for 55 to 60 minutes, or until a toothpick inserted into the center comes out clean. Baking time may vary slightly depending on the oven, so checking near the end is important.

Once finished, allow the banana bread to cool in the pan for about 10 minutes before transferring it to a wire rack to cool completely. This helps it set properly and makes slicing easier.

This incredible dessert is moist without being heavy, sweet without being overwhelming, and simple enough for beginners to make successfully. Whether enjoyed warm with butter or eaten plain, it remains a reliable and crowd-pleasing recipe that never disappoints.

Image credit: natashaskitchen.com

   An expectant mother asks a large language model for advice concerning swollen legs. If she asks in English, she would most likely be warned about preeclampsia, which is a medical complication that is responsible for over 70,000 maternal deaths every year. However, if she asked the same question in Swahili, she would merely be told not to worry. In the scenario, the disparity is plausible given the research and potentially detrimental to health. The stakes of such a difference are only becoming more frequent by the day. 

Large language models are predominantly trained on English-language text, and as a result, they simply know more in English. To elaborate on the aforementioned, LLMs parse inputs that affect their responses; in particular, before the model directly answers the text, it breaks it into small chunks called tokens. Since tokenization is designed for English, non-English text is broken into inefficient fragments. Consequently, more tokens are needed to convey the same amount of information written in other languages. The matter is significant, as it indicates the model is less efficient at understanding the language, taking more energy to do so, and as a further factor into consideration, developers pay per token. So the same query would cost significantly more in a language other than English. 

Most overlook the multilingual gap and fail to understand how vastly divergent responses may be. As suggested by The Lancet Digital Health, researchers recently tested various frontier models for reasoning and medical knowledge across eleven African languages and discovered that, nevertheless, the top-scoring models performed between twelve and twenty percentage points worse than in English. In the worst cases, results were quite dramatic; one model that correctly answered around 75% of questions in English had dropped to approximately 23% in other languages. The performance was evidently at the level of an English language model from five years ago in terms of accuracy, said Deena Mousa, a Yale graduate and program officer for the Global Health & Wellbeing Cause Prioritization (Wang et al.). Individuals who have tracked AI models over the past five years know the significant difference compared to the latest versions established this year. 

Although individuals may argue that AI models are improving, they assume the multilingual gap is closing; in fact, they are wrong. Researchers at Stanford used formal benchmarks to evaluate the latest released models. Those newer systems indeed performed higher than the others in the original study, so progress may be present. Regardless, the important caveat to consider is that the models still did not perform nearly as well as in English. So, GPT 5.2 would be roughly on par with models, such as GPT-3, which was released around 2020 (“Technical Performance - The 2025 AI Index Report”). Improvement is, therefore, existent, yet the multilingual gap has not closed and proven to be significant in the past years. 

Withal, models designed to be multilingual are, unfortunately, not immune to the issue either. A study directed towards Metta LLaMA’s model, which was primarily built for multilingual use, found that, nevertheless, when asked a question in another language, it often retrieves the answer internally in English first and then proceeds to translate at the last step. These additional steps add opportunities for errors, and as the same study adds, the model scored a quarter of the factual questions correct, although the internal representation showed that it indeed found the right English answer (“How Multilingual is LLaMA?”). Simply put, the model had the knowledge, yet it was adrift from the proper translation.

The multilingual gap in LLMs is, therefore, not merely a minor technical quirk; rather, it is a persistent and growing issue impacting human health. As these models become more ubiquitous in global health, education, and everyday decision-making, the cost of unequal performance across languages should be well aware of. Until developers address these imbalances, such disparities will put millions of users at detrimental risk. 

Works Cited

“How Multilingual Is LLaMA?” 19 June 2025.

“Technical Performance - the 2025 AI Index Report.” Stanford.edu, 2025, hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance. Accessed 3 Apr. 2026.

Wang, Xiaofei, et al. “Reasoning-Driven Large Language Models in Medicine: Opportunities, Challenges, and the Road Ahead.” The Lancet Digital Health, vol. 8, no. 1, Jan. 2026, p. 100931, www.thelancet.com/journals/landig/article/PIIS2589-7500(25)00113-X/fulltext, https://doi.org/10.1016/j.landig.2025.100931. Accessed 3 Apr. 2026.

Cynthia Lam with the H2Pro device