Idle/Random/whatever Thoughts of a Demented/Idle/Whatever Mind

 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 predi...

1.  The Illusion of Thinking  - A recent paper from Apple's Machine Learning Research team challenges the idea that Large Reasoning Models (LRMs) can truly "think." The researchers argue that current methods for evaluating these models, which often use standard math and coding problems, are flawed due to potential data contamination and a focus on final answers rather than the reasoning process. Using controllable puzzles to test the models, the study found that even the most advanced LRMs experience a complete collapse in accuracy when the problem's complexity passes a certain point. Counterintuitively, as problems became more difficult, the models' "thinking" effort decreased, suggesting they "give up" rather than reason more deeply. The research identifies a significant gap between mimicking thought-like patterns and genuine, scalable reasoning, questioning the true capabilities of current AI.   My Take : Thinking is a complex p...

1. How Major Indian Apps Are Bypassing Google’s Policy to Spy on Your App Drawer? - Report reveals that many major Android apps on the Google Play Store are allegedly bypassing Google's privacy policies to access a list of other apps installed on users' devices. Apps like Swiggy, Zepto, and HDFC are among those seemingly "peeping" into users' app drawers. While some argue this access might be needed for core functionality (like fraud detection in banking apps), the report suggests it's being used for fingerprinting and user profiling. Analysis of AndroidManifest.xml files shows apps use methods like querying by package name or intent signature to extract this data. Concerns are raised about apps like KreditBee, which appear to be profiling users based on the presence of specific apps, including those related to religion and astrology. The report questions Google's enforcement of its policies and the lack of user consent or awareness regarding this data co...

  1. How Many Decimals of Pi Do We Really Need? - This article explains that even for NASA's highest accuracy calculations, such as interplanetary navigation, only a small number of pi's decimals are needed. For example, JPL uses pi to the 15th decimal place (3.141592653589793). The article illustrates this by calculating the circumference of Voyager 1's distance from Earth, the circumference of Earth, and even the circumference of the known universe, demonstrating that only a limited number of decimal places are necessary for accurate calculations, even at vast scales. My Take : We often get super excited about accuracy and going to crazy lengths to achieve it. But, honestly, for any everyday functional use of pi, having 3 digits after the decimal gives you really good accuracy, which is usually within the range of error from other stuff anyway. In fact, that article mentions 37 decimals would be enough to calculate the circumference of the observable universe with the ac...

After a brief hiatus I am back at publishing this. Had been heads down at getting Sudama Clinic operational. Not that I did not learn during this time, but was struggling to find time to pen down all that. Got a little breathing time now, so here is it is.    1. Dillo Browser : Dillo is a remarkably lightweight and fast web browser, prioritizing speed and efficiency above all else. Designed for minimal resource usage, it excels on older or less powerful systems. By focusing on core web functionality and avoiding complex features like JavaScript and advanced CSS, Dillo achieves a streamlined experience. This minimalism also contributes to enhanced security by reducing potential vulnerabilities. Available across various platforms, including Linux and BSD, Dillo caters to users seeking a basic, no-frills browsing experience. Its simplicity makes it ideal for specific use cases where speed and resource conservation are paramount, offering a practical solution for those who requi...

1. Why you should never use !! in Kotlin - Using double-bang operator(!!) is a bad practice; multiple alternatives cover every other possible case. Using it is a hack and a shortcut that will backfire while not saving you much time at the time of writing using it. You shouldn’t use it unless you're writing a short personal script. My Take : The double-bang operator (!!)  which throws null pointer exception in Kotlin might very well be considered " Billion Dollar Mistake ." While it offers a concise way to assert that a variable is not null, I believe it's preferable to have slightly more verbose code that explicitly handles nullability and avoids potential runtime exceptions. Null pointer exceptions have plagued programmers for decades, and introducing a feature that can easily lead to them seems counterproductive. I'm uncertain whether all Kotlin linters currently check for excessive or inappropriate use of the !! operator, but if yours don't, I strongly re...

1. Llama 3.3 -  Llama 3.3 70B model offering similar performance to the Llama 3.1 405B model, allowing developers to achieve greater quality and performance on text-based applications at a fraction of the cost.  My Take : Meta continues to impress with each new release. Their recent claim that the 3.3 70B model performs as well as the 3.1 405B model is particularly noteworthy. This rapid progress suggests that these powerful models could soon be running on personal computers or even directly in web browsers.Meanwhile, OpenAI's recent move to introduce a $200 monthly subscription for advanced access to its models raises questions about its long-term strategy. While the initial novelty factor might justify such a premium, it seems excessive, especially for users in regions like India. It's likely that we'll see highly capable language models running on mobile devices within the next six months, potentially even before desktop deployment. My experience with a Llama 3.2 1B mod...

1. C++ Pitfalls - this article, explains some of the pitfalls you can fall into when programming in C++ and shares their experience with them. This includes Operator precedence, Arithmetic conversion rules, Right bit-shift, Implicit conversions and Order of evaluation.  My Take : We've all read about JavaScript's quirks, leading to unexpected outputs. But let's remember that such behaviors aren't unique to JavaScript. Even languages like C and C++ have their share of ambiguities and compiler-specific behaviors. While it's tempting to dive deep into these intricacies, I've found that focusing on clear, consistent, and unambiguous code is often more practical. However, understanding these nuances is essential for any experienced programmer. It helps us avoid pitfalls and write robust, reliable code. Remember, mastering a language isn't just about writing "Hello World." It involves delving into its intricacies, including its counterintuitive and ambi...

1. AI Makes Tech Debt More Expensive - The core argument is that generative AI significantly widens the gap between codebases with low and high tech debt. Companies with clean code can leverage generative AI to write code much faster, while companies with messy legacy codebases struggle to adopt these new tools. This widens the gap between the two codebases, making tech debt even more expensive for companies with outdated systems. My Take : Legacy systems inevitably accumulate technical debt. A common analogy for this is the state of urinals over time: missed targets lead to increasing mess, keeps pushing people away from urinals and leading to more mess in turn. AI-powered coding tools could potentially accelerate the accumulation of tech debt. However, skilled engineers can leverage AI to reduce and even eliminate it. Ultimately, AI is just a tool, and human judgment remains crucial for making informed decisions. I hope AI will be used responsibly to drive positive outcomes. 2. Make...

 1. Booleans are a trap : Booleans, while seemingly simple, can introduce complexity and limitations into domain modeling. When used to represent complex states, booleans can lead to unclear code and potential errors. To avoid these pitfalls, consider using enums or enum sets. These approaches provide a more expressive and maintainable way to model complex states in your applications. My Take : I wish I had learned this earlier in life. When working with Google Docs, I realized the limitations of booleans. Expanding boolean-based behaviors often led to adding more fields, making the system complex. Enums would have provided a simpler solution. The problem wasn't just the booleans themselves, but also the need to pass them through multiple layers of the stack, often requiring additional parameters. Another important aspect of enums is the inclusion of an " enum not specified " state. This state explicitly indicates that an enum value hasn't been set, which can be usef...

1. One Company A/B Tested Hybrid Work. Here’s What They Found. - Recent research has revealed that a hybrid work model, where employees work from home three days a week, can significantly boost employee satisfaction and reduce turnover rates. This not only improves employee morale but also saves companies millions of dollars in recruitment and training costs. To successfully implement a hybrid model, businesses should prioritize rigorous performance management systems, coordinate team and company-wide schedules, and secure leadership support. Moreover, executives should experiment with different management practices to optimize their approach and maximize the benefits of hybrid work. My Take : There have been too many debates about remote work versus office work. I would have loved to see data comparing the effectiveness of remote, hybrid, and office-based work. I've worked at companies that mandated office work, like Tata Elxsi, where we didn't have laptops and had to rely on...

1. Transforming CSV Files into Graphs with LLMs -  This is a step by step guide including code to transform csv files into graphs. There are prompts, python code and samples provided in it.  My Take : This is an excellent approach to streamlining data cleaning, a task that often consumes a significant portion of a data engineer's time. By leveraging the power of LLMs, we can potentially reduce this effort significantly. However, it's important to note that not all LLMs are equally effective for this purpose. Fine-tuning LLMs specifically for data cleaning tasks could further enhance their capabilities. I'm eager to learn more about this promising technique and its potential applications. Do let me know if have something interesting to share.   2. SOFA Start Often F*** Achievements - The SOFA principle advocates for starting numerous projects and finishing few. This approach prioritizes gaining diverse experiences and skills over completion. By frequently initiating new ...

1. Why Founders Need to Focus More on Sales and Marketing - Harvard Business School senior lecturer Mark Roberge argues that every aspect of being an early-stage founder involves sales. But many founders lack an understanding of how to incorporates sales into their ventures. This talks about how to hire for early sales roles, design compensation, and lay a strong foundation for a growing sales team. Hiring the right salesperson is crucial, as they should be skilled in both product knowledge and customer interaction. Scaling sales requires a strategic approach, considering factors like product-market fit and customer retention. The sales compensation plan can be a powerful tool for aligning sales behavior with business goals. While AI is making strides in automating certain sales tasks, the human element of building relationships and understanding customer needs remains indispensable. My Take : IMHO a startup is like an elephant from the story of blind men and an elephant . There are d...

1. ROS : ROS (Robot Operating System) is a versatile open-source framework that simplifies robot development. It provides a modular architecture, robust communication mechanisms, and a vast ecosystem of tools and libraries. This comprehensive platform empowers developers to create efficient and innovative robotic systems, from hardware abstraction to high-level behavior control. ROS's flexibility and adaptability make it suitable for a wide range of robot applications, while its active community fosters collaboration and knowledge sharing. My Take:  I recently discovered ROS, which has been around since 2007. It's remarkable that such a powerful tool has existed for so long. Over the years, ROS has evolved into a comprehensive platform, offering a vast array of algorithms to simplify robot development. The ability to run ROS on Raspberry Pis further enhances its capabilities, making it a versatile choice for a wide range of robotic projects. 2. Should we be worried about vega...

 1. Stem cells reverse woman’s diabetes - Scientists have achieved a breakthrough in treating type 1 diabetes by using reprogrammed stem cells. A woman was successfully treated by receiving a transplant of these cells into her abdominal muscles. Following the transplant, she was able to maintain stable blood sugar levels without insulin for over a year. This is the first time such an approach has shown to be effective in reversing type 1 diabetes. While this is a promising development, more research is needed to confirm its long-term safety and efficacy in a larger group of people. My Take : This is a groundbreaking discovery. My father, along with millions of others, suffers from diabetes. Unfortunately, due to genetics, I'm likely to join them soon. However, this news offers hope not only for me but also for countless individuals who endure daily injections. I sincerely hope that the cost of this treatment remains affordable. Perhaps the government could intervene to ensure acce...

 1. Google says it’s made a quantum computing breakthrough that reduces errors - Google researchers have made a breakthrough in quantum error correction, which could pave the way for useful quantum computers. The main challenge in quantum computing has been the high error rate of the hardware, which prevents them from executing long and useful algorithms. The new research demonstrates that it is possible to add components to reduce these errors, which was previously thought to be impossible due to engineering limitations. This bolsters the idea that error correction is a viable strategy for building a useful quantum computer. My Take : While these advancements may seem incremental, they are crucial for the development of quantum computers. However, it’s important to avoid overhyping their implications, as was the case with the early hype surrounding GenAI being AGI. Despite this, the progress made with GPTs represents a significant leap forward in the field of AI. Similarly, these...

1. The Genius Behind the Quantum Navigation Breakthrough - This is the World's First Unhackable Quantum GPS. Using the coldest atoms in the Universe, it may revolutionise the way we navigate. This Youtuber to got to go behind the scenes and see their device up close to try and understand just how important this breakthrough is. My Take : This video goes into more details about how quantum navigation works as well hits the home run in explaining why its really needed. I had also mentioned about Quantum Navigation in a previous post where IEEE had covered it. This video goes to explain more details of it and if you are a physics nerd, you would enjoy it.   2. Forget ChatGPT: why researchers now run small AIs on their laptops -  This article talks about the different small and locally runnable LLMs which various scientists are using for not just poetry but also for medicine. The locally runnable LLMs have the distinct advantage of preserving privacy, fine tuning using the open...