[8 September 2025] Interesting Things I Learnt This Week

 1. Multi-Token Prediction Reshapes LLM Training Paradigms - This research proposes that training large language models to predict multiple future tokens simultaneously, rather than just the next one, improves sample efficiency. The approach involves the model predicting the next 'n' tokens at each position in the training data using separate output heads on a shared model base. When used as an auxiliary task, multi-token prediction enhances performance in downstream tasks for both code and natural language models without increasing training time. The benefits are more significant with larger models and persist across multiple training epochs. Notably, this method excels in generative tasks like coding, where a 13B parameter model outperforms next-token models by solving 12% more problems on HumanEval and 17% more on MBPP. Additionally, it aids in developing induction heads and algorithmic reasoning on small tasks, and offers up to 3x faster inference speeds with 4-token prediction, even at large batch sizes.

My Take: The technology behind large language models is improving by leaps and bounds. While each day brings smaller, incremental steps, the cumulative effect is a forward surge that makes the tech more efficient and drives down costs—a true game-changer. In my opinion, the real race is not among all the players but specifically between Google and Chinese providers like Deepseek and Qwen. Everyone else is just trying to keep up. OpenAI, once seen as the undisputed leader, seems to have stumbled. The lukewarm reception of GPT-OSS and the challenges faced by GPT-5 suggest that the recent departure of key founding team members may be having a real impact on their releases. This talent exodus is compounded by companies like Meta offering what feel like unsustainable salaries to top AI researchers. Ultimately, LLMs are becoming a commodity, much like databases did decades ago. Open-source models will eventually become just as good as their commercial counterparts—think Oracle versus Postgres. The real winners of this new era won't be the companies that build the models, but the ones that create excellent services using these generative AI tools. While today’s costs for running LLMs are high, the winners will be the innovators who use them to solve problems and deliver valuable services to end-users.

 

2.  How much extra is the gift wrap? - The core message is that presenting a product or service with care and thoughtful presentation elevates it beyond a mere transaction. By "gift-wrapping" your offering, whether through a summary, useful packaging, or any added touch, you transform it into a memorable experience or "story," which is what customers truly value, rather than just the functional aspect.

My Take: Gift wrap doesn't have to be a physical thing. It’s any small, thoughtful act that enhances the value of your core offering. Just like choosing the perfect gift wrap requires careful thought, the small details you give your customers demand the same attention. A simple act, such as offering a cup of tea to visitors at your store, can be a great "gift wrap" for your service. The goal isn't to get a huge reward from every single gesture. Instead, these small acts dramatically increase the perceived value of your primary offering. They show you care, and that appreciation is often the most powerful gift of all.

 

3.  Scientists reverse Parkinson’s symptoms in mice - Scientists at the University of Sydney have uncovered a malfunctioning version of the SOD1 protein that clumps inside brain cells and fuels Parkinson’s disease. In mouse models, restoring the protein’s function with a targeted copper supplement dramatically rescued movement, hinting at a future therapy that could slow or halt the disease in people. 

My Take: As the life span of the general public increases, it is sad to expect that diseases like Parkinson's and Alzheimer's may become more common. This is especially difficult when you see your elders fighting these conditions, and it’s natural to hope for a cure. While it may not always be in the news, many Indian institutions are at the forefront of this critical research. Leading centers like the Indian Institute of Science (IISc) and the National Institute of Mental Health and Neurosciences (NIMHANS) in Bengaluru, along with other IITs, are actively engaged in groundbreaking work, from understanding the genetic basis of these diseases to developing novel diagnostic tools and new therapeutic drugs. This research offers real hope for better treatments and a cure in the future.

 

4. Longitudinal associations between vegetarian dietary habits and site-specific cancers in the Adventist Health Study-2 North American cohort - This study, based on the Adventist Health Study-2 cohort of 79,468 North American Seventh-day Adventists, investigated the link between vegetarian dietary patterns and site-specific cancer risk. Vegetarians (vegans, lacto-ovo-vegetarians, pesco-vegetarians) showed significantly lower risks for total cancers and medium-frequency cancers compared to nonvegetarians. Key protective associations were found for colorectal cancer, stomach cancer, lymphoma, and all lymphoproliferative cancers. Vegan diets were particularly linked to reduced breast and prostate cancer risk in younger individuals. Adjusting for BMI slightly weakened these associations, suggesting a partial mediating role for lower body weight. The study highlights the importance of specific dietary subtypes over broad "vegetarian" categories and supports a protective effect of plant-based diets, especially against gastrointestinal and certain hematologic cancers. While observational limitations exist, the large sample size, rigorous validation, and health-conscious population strengthen the findings. Results suggest that vegetarian diets are associated with lower overall cancer risk, with public health implications, though causality cannot be definitively proven.

 My Take: That's a very interesting article, and the results are even more compelling. My personal experience as a lifelong vegetarian, informed by the Jain philosophy of "you are what you eat," makes me believe in the power of diet. However, we can't ignore the findings of this study. It would be important to see a broader study undertaken, not just in the USA, but in other countries as well. The lifestyle of those studied should also be a significant consideration. I believe that in the US, being a vegetarian is often a conscious choice, and those who make that choice may also lead a healthier overall lifestyle. It would be valuable to see these results replicated across different countries and populations to be sure of the findings. The other key factor would be for the research to uncover the exact biological mechanism of how a vegetarian or non-vegetarian diet affects cancerous tissues.