Touhidul Islam

Yeah, even I am guilty of building my foundations on top of the UNIX philosophy, programming in C (the one without pluses) and even C++ with its highly optimized object-oriented data model. Just a few days ago, I learned about yet another petty Linux / UNIX design trap that can surprise you. It's the output stream buffering in the standard C library on UNIX systems that also affects C++. In the standard library, from the UNIX design point of view an integral part of the operating system, the decision on default buffering is made based on what the output of a program is connected to. For interactive programs it is important that you see the output lines as they are emitted by the application code. Therefore the default is line buffering where a newline character flushes the buffer and the data is passed to the terminal – but only when the output file descriptor is known to be a terminal. Programs that don't talk to a terminal do not flush every line by default and thus do not produce all output if they crash before flushing the buffer. I learned about this while using Compiler Explorer when I realized that it's not possible to debug crashing programs with simple `std::println` (from the C++23 <print> header) or similar tools (writing to `std::cout`) because the output is never shown. Of course there are solutions to the problem including those as simple as using `std::println` together with `std::cerr` or changing the buffering configuration of `std::cout`. However, my first suspect was the Compiler Explorer itself, not an age old “feature” of the libc. And this can easily bite you even as a non-developer if you pipe programs output to any interactive viewer with the assumption that the program has a text output that is line-buffered.

8

12 Kommentare

Seems parquet goes like AI embeddings encoding, hence compression!

1

It seems service deprecation is a thing now for AWS. I'll not argue here if that's a good or bad thing 😜. But if you want to keep track, here are two useful resources: 👉 https://lnkd.in/dCHGPiqK A curated list of breaking changes maintained by Scott Piper 👉 https://lnkd.in/d5rxybzi A list of service deprecation announcements (even if the post title is not that obvious) maintained by Luc van Donkersgoed #AWS #AWSCommunity

22

9 Kommentare

🔐💻 How to securely share your AWS S3 objects without making them public? Let’s talk about Pre-Signed URLs! 🚀 1) Generate a temporary URL to grant time-limited access. 2) Maintain full control over permissions and expiration time. 3) Perfect for secure file sharing and upload access. How are you leveraging Pre-Signed URLs in your AWS setup? Let’s discuss!👇 Read my latest blog post on Hashnode and also give your valuable feedback on this. #AWS #CloudComputing #CyberSecurity #TechTips #CloudSecurity #DataProtection #S3 #DevOps #TechInnovation #AWSCommunity

17

2 Kommentare

🧠 New article published! EventBridge Schedules provides a serverless approach to create one-time or periodic schedules to trigger events or actions in #AWS Services. And with #Terraform, it becomes easier to provision and maintain your schedules. 💭 What to expect in this article? ▪️ How to provision Schedules in AWS EventBridge with Terraform ▪️ Create periodic Schedules using `rate()` or `cron` expressions ▪️ Create one-time Schedules ▪️ Create one-time Schedules from a Lambda function with AWS SDK Give a read here https://lnkd.in/guHWkiJ6 And let me know what you think :) #Serverless #AWS #Terraform #EventBridge #Schedulers

17

1 Kommentar

🚨 Big News in the AI World! 🚨 #Google just made a groundbreaking move by spending an eye watering $2.7 billion 💰 to bring back #Noam_Shazeer, a brilliant #AI engineer who left the company three years ago. Shazeer originally developed a cutting edge AI chatbot named Meena, which he believed had the potential to revolutionize Google Search and generate trillions in revenue 💡. However, Google executives initially held back due to concerns over safety and fairness. In 2021, #Shazeer left #Google to co-found Character.AI, a company focusing on chatbots, and quickly gained more than 20 million active monthly users 🌍. Despite this success, #Character.AI faced challenges in sustaining its growth without a consistent revenue stream, leading Shazeer to negotiate with several tech giants, including Meta. Google co-founder #Sergey_Brin personally pushed for Shazeer’s return, recognizing that Google may have been too cautious when it came to releasing AI products. Now, with Shazeer back as Vice President, Google is doubling down on its efforts to lead in the AI race 🏃‍♂️, joining the fierce competition with #Microsoft and #Amazon. This move is sparking debates across the tech world: is this kind of spending on AI justified, or is Big Tech getting ahead of itself? 🤔 But one thing is clear: AI is here to stay, and it’s going to change the game forever. #AI #ArtificialIntelligence #MachineLearning #Google #TechNews #Innovation

5

💎 #ReAct, an interesting approach to enhance the #reasoning and #problemsolving capabilities of large language models (#LLMs) by integrating reasoning steps with task-specific actions that enable interaction with external environments such as knowledge bases or APIs. ✅ This approach allows LLMs to overcome limitations of existing reasoning methods like Chain-of-Thought prompting (#CoT), which relies solely on the LLM's internal knowledge and is prone to hallucinations. ✅ Empirical evaluations on various reasoning and decision-making tasks, including question answering, fact verification, and interactive games, demonstrate ReAct's effectiveness. ✅ The authors highlight the potential of combining #ReAct with #CoT to leverage both internal knowledge and external information access capabilities of LLMs for enhanced problem-solving. 👉 Paper https://lnkd.in/eiCJhGf8 ♻️𝒌𝒊𝒏𝒅𝒍𝒚 𝒓𝒆𝒑𝒐𝒔𝒕♻️

40

6 Kommentare

How well #LLMs generate #code? Paper "Analysis of Code and Test-Code generated by Large Language Models" tries to answer that by comparing chatGPT and CoPilot results. It states that engineers are afraid that LLMs will sunset programming as a profession. Which is frequently discussed topic. To see model code generation capabilities, prompts from picture below are used as inputs to the model and results were inspected, including compilation, etc. I would say if person is capable of writing such a prompt then he/she is also capable of googling implementation of the algorithm in the internet and simply copying it. In this case LLMs are just convenient tool to copy implementation from previous training data, real decision on which method to be used is made by real person. And this is critical aspect in my opinion. Such approach still requires programmer to know methods and make decisions. LLMs are reducing routine work somehow, but not capable of replacing a person. We should start to worry when prompts will be asking about "finding shortest path in weighted graph" for instance. That would require way more intelligence from models. If you are interested in reading the paper: https://lnkd.in/dAigPVvR

57

5 Kommentare

>> "Data Scientist" role doesn't exist in Germany any more During the last decade we created a lot of content explaining basics of ML and raised a generation of people, who would like to develop/improve/apply ML algorithms. They plan to spend their time on parameter tuning and squeezing yet another % of accuracy. Somehow many people are convinced that life is some kind of Kaggle competition and the more accurate models you create, the better you will be paid. To be honest, it was never the case. However in the past you could find a "Data Scientist" job, where squeezing accuracy was your main responsibility. Many companies learnt in a hard way that very deep knowledge of ML algorithms is not enough to create value out of them. Therefore typical job description of "Data Scientists" role in 2024 contains phrase like this: 𝘚𝘵𝘳𝘰𝘯𝘨 𝘬𝘯𝘰𝘸𝘭𝘦𝘥𝘨𝘦 𝘰𝘧 𝘔𝘓 𝘢𝘯𝘥 𝘦𝘹𝘱𝘦𝘳𝘪𝘦𝘯𝘤𝘦 𝘪𝘯 𝘣𝘳𝘪𝘯𝘨𝘪𝘯𝘨 𝘮𝘰𝘥𝘦𝘭𝘴 𝘵𝘰 𝘱𝘳𝘰𝘥𝘶𝘤𝘵𝘪𝘰𝘯 Mysterious "bringing models to production" essentially means that you should know Machine Learning Engineering (MLOps, Cloud, CI/CD, etc). All Data Science jobs become either Machine Learning Engineering, either Data Analytics + Data Engineering. Pure Kaggle-like Data Science roles do not exist any more (or to be more precise exist only in research in very limited amounts). There are 2 take-aways here: 1. If you're a Data Scientist and do not want to touch Engineering at all, it will significantly limit your career opportunities (Nowadays MLOps questions are asked even in interviews for management roles). 2. Many companies understand the need of MLE and are looking for people, but there is still a significant deficit of such specialists on the market. However, learning MLE on your own might be challenging: to learn how to "bring models to production" you need to talk with someone who did it, just online course is not enough. By the way, I'm a 𝐃̶𝐚̶𝐭̶𝐚̶ ̶𝐒̶𝐜̶𝐢̶𝐞̶𝐧̶𝐜̶𝐞̶ 𝗠𝗟𝗘 mentor ;)

115

20 Kommentare

Retrieval Augmented Generation (RAG) is a straightforward concept that any company can implement with little effort. You can think of RAG as an NN-based document summarization, like merging the results of a simple elastic search/apache solr query with a text summary. The original concept involved retraining the Llm to have a similar latent space, however in applied concepts the retraining has been dropped and replaced with a vector database and prompt engineering that can be paired with any Llm that exists. Based on Meta's LLAMA3, LlamaIndex, Qdrant, Ollama, Angular, and FastAPI, I built the RAG demo about funding lines for universities of applied sciences in hesse. The demo is available at (not well optimized for mobile use): https://lnkd.in/eBYQwKh4 This demo is a weekend project when the little ones were asleep, so it took me less than 3 days. Imagine your company creating a bot like this with more manpower and effort. Best of all, it's completely free and costs you nothing. Your data stays in your company without any security or GDPR violations. P.S.: On my own behalf: I also work as a speaker and consultant on AI topics. Through my research in context-aware fault diagnosis in industrial production lines, I know the state of the art in AI and have implemented various AI use cases myself using different neural networks under industry constraints. If you need some advice, just get in touch with me. I'm looking forward to talking with you about self-built solutions without buying expensive cloud services.

11

2 Kommentare

Over the past 5 years, I've worked extensively with Google Cloud Functions, growing with the platform from its first version to the second. Each update has introduced new features and improvements that have significantly changed the way we use serverless applications. This change isn't just a small update; it's a major improvement that combines the best of Cloud Functions with the power of Cloud Run. This means better performance, easier management, and more control over how we run our applications. For those of us who've been with Cloud Functions from the start, moving to Cloud Run Functions will bring some challenges but also great opportunities. It will require us to update how we work and to possibly rethink our current setups to make the most of the new features. This update is a big deal for us developers. it's a chance to make our applications faster, more scalable, and more efficient. For more detailed information on this transition, check out this link. https://lnkd.in/enmaWxzB #GoogleCloud #ServerlessComputing #CloudRunFunctions

6