Kazi Hasan Zubaer

The policies being driven forward by Thierry Breton around the EU AI Act have been utmost damaging for the European tech industry, and they need to be revised as soon as possible. We need to better balance driving European innovation through data collection and modeling with conflicting needs such as data protection and safety. We need to ensure we can turn these developments into businesses that can be the pillar of the next generation of companies that provide productivity growth in Europe. We have no other choice - either we do this here, or we will have to take over products developed in the US and in China, as it already happened with search engines, cell phones, and social networks. We wouldn't want to let this happen with AI applications in robotics or digital assistants. #euacc #euaiact #aiact

9

One thing I’ve learned in my first 10 months at Meta… Don’t ask a senior eng what to do or how to do something. Instead, explore different approaches and ask the senior eng which they would recommend. For example, when devising the tech plan for a new feature that requires storing and retrieving new data, don’t simply ask what data solution you should use. Instead, scour internal wikis, existing code patterns, etc, to learn about the data storage solutions available. Then list out the pros and cons specific to your use case in terms of latency cost, capacity usage, reliability, implementation complexity, etc. Sending that breakdown to the senior eng (along with your own proposal for what you think the best choice is) will not only make their lives easier, but prove to them that you’re capable of working at the next level. Of course, there are exceptions. For example, if there’s a quick question that a senior engineer could answer in 30 seconds, but would take you 2 hours to figure out on your own, it makes sense to ask. What do others think? What types of questions should you ask senior eng and how should you decide whether to ask a question or figure it out on your own? #softwareengineer

1,329

138 Comments

Principal Engineer Staff Engineer Distinguished Engineer (my favourite) I’m going to say it and it’s likely a little rhetorical: I dislike all these terms immensely. Please can someone give me the canonical description for what they are supposed to mean? (I know I can use a search engine or AI to do it… I want to know what people think they mean). The industry I work in is in danger of disappearing up its own backside. What’s wrong with: Software engineer Why do we need to differentiate? The only reason I can see is signposting to people with differing levels of experience or different sets of responsibilities. If we need to do that then, I’m not sure our teams are set up correctly in the first place. #software #softwareengineering #softwaredevelopment #productdevelopment

98

105 Comments

Falling back in with a side-quest Colab about LLM evals (high level) - its not just about prompting. Including my absolute favorite: Cosine Similarity. I love this algorithm (or is it technically just a function? ratio? idk/care). This colab is a bridge to some concepts I'm hoping to get to soon: like how to automate quality evals, and auto-reject outputs that do not closely follow the prompt instructions. These are things that really should be handled by code when creating LLM-backed services for production purposes. Again, this is using the smallest Llama 3.2 1B model, and there's no training step this time, so you can run it on just a CPU runtime without much fuss. LMK what you find (anyone who runs this should get different results, and those results are probably always going to be interesting)

17

3 Comments

We're headed down the complex path of "replacing and improving upon activities you might do with your smartphone" with natural conversation. Try creating your own AI conversational agents at iyo.studio and let us know if you have feedback!

1

A blog on (very) nerdy topic that I worked on with our (amazing) intern Xiangpeng Hao this year: https://lnkd.in/eQDfvNfH. We quantify how much of the discussion about changing Apache Parquet metadata was shiny-new-object syndrome and how much was actual technical limitations. We think there is easily a factor of 4 to be had from existing implementations before new formats are needed

290

12 Comments

Exciting times ahead at Autonomys (formerly Subspace)! 🚀 We're pioneering the Age of Autonomy with our identity-based decentralized AI (deAI) stack for human + AI (H+AI) collaboration. Our Autonomys ID (Auto ID) protocol ensures secure, verifiable identities for both humans and AI agents, fostering a trustful, ethical AI environment. With distributed storage, compute, and a robust dApp suite, we're setting the stage for scalable, AI-powered applications. Join us in this journey towards a decentralized, autonomous future! https://lnkd.in/eUCn4yZY #AI3 #Web3 #DeAI #Blockchain #Identity #Autonomy #AIRevolution #DigitalIdentity

11

Hello Everyone, Welcome to Part 2 of our series on Building a Cache Iteratively. In this instalment, we'll delve into the concept of introducing threads into the equation, understanding the challenges of concurrent access, and devising a solution within a data structure. We'll tackle these issues in the most basic yet effective manner, gradually building a reliable system. You can find the detailed discussion in the following Medium article: https://lnkd.in/gsKYTyzN For those interested in exploring the code base, it's available on GitHub: https://lnkd.in/gp-dTVdU Feel free to dive in, explore, and share your thoughts! #LRUCache #Cache #LLD

14

How about contrastive preference optimization (CPO)? It looks similar to ORPO that incorporates both preference and nll

2

What a honour to take part in building resources to teach primary school students about AI! 🦾 Well done to the Code for Life team!! Resources are available for all primary teachers and you can find out more below ☺️ Here’s a teaser: https://lnkd.in/eeAvt3C9

22

1 Comment

On doing Prompt Engineering, rather than guessing I've been asked recently to explain how I craft the prompts that work more often than not, want to share the approach with you as well. Context - we are often using LLMs to extract structured data, but the approach applies to any system interacting with LLM. 1. I build iteratively until getting the results I want, so it helps to know what I expect in the response. example - you are doing categorization using LLM and you know the ground truth classes. so you iterate until the quality is good. usual ML stuff. 2. try different approaches and try to find patterns in how the model responds, so your search becomes more efficient. example - if you expect (semi) structured output, the order of attributes makes a difference, model looks to build on what it generated before. CoT is about that as well. so, find the patterns. 3. use templating and dynamically build prompt content (task definition, N-shot examples), to facilitate iterations and speed. example - dynamically generating ordering of attributes, their possible values, and descriptions based on the metadata. changing metadata is easier to do and automate than editing templates directly. 4. use separate LLM calls to give ideas for prompt adjustments. example - we have classification task - input data, expected result, original prompt and LLM-classifier response. then do another call where pass some (or all) of the context and ask to tune the original prompt. 5. if possible, store prompt / completion / metadata and context, because I often come up with post-analysis ideas. usually, that is a part of some inference pipeline, and combined with context-aware visualization, allows me to understand quickly what is happening. study Experiments Tracking in ML. 6. scale gradually but consistently. first make a single request (sample) work, expand to class, then a group of classes. then make it easy to test different prompt variations, inference parameters, then LLMs. when your analysis is automated as well, you are spending more time experimenting and (hopefully) close-to-zero effort on manual work of changing configs etc. 7. make sure to have fun. treat prompt engineering as a puzzle and game. make sure you can iterate quickly to get feedback instantly and keep you in the loop. if you can do anything prompt engineering for long enough - you'll come up with something good. Q: what are your tricks and lessons?

17

Hello, colleagues, this Yahoo article is interesting and relatable for professionals. Here is a teaser: "Amazon has an incredibly granular list of criteria that need to be met before you can be considered for promotion, so I worked with my manager to identify the gaps in my performance, ... and I created a project for each gap that, if completed, could serve as evidence of my ability. I worked backward to create a strict timeline to finish my projects ... that helped me get systematic about my approach moving forward." https://lnkd.in/dsfvPeAk Take a lesson from this person. I seems Amazon gets very analytical when it comes to promotions. And in a company that size, it is impossible to not promote without rigid check-off type policies or whatever XYZ job promotion software. But, hey, there is a lot I will do for a $500,000 salary. Please take note of this last sentence, if anyone works for Amazon and wants a Quality Engineering Manager/Director anywhere in the world.

1

1 Comment

I am excited to share the latest blog, co-authored by myself, Junhwi Kim and Vyas Sekar, detailing how we managed to successfully localized and improved LLM by combining Fine-Tuning and RAG. https://lnkd.in/gKDkgGPq

70

3 Comments

I echo Pere Martra' s urgency for the European Union to advance in artificial intelligence and other key sectors. While Europe boasts incredible educational systems and social safety, its conservative approach often limits the entrepreneurial spirit seen in places like the United States and parts of Asia. These regions not only embrace but celebrate risk as a pathway to innovation, which starkly contrasts with the European caution that can hinder our global competitiveness in rapidly evolving fields such as AI and blockchain. To truly lead in technological innovation, Europe must strike a balance between its commendable focus on security and a more dynamic approach to business and technological advancements. Promoting a startup-friendly culture, reducing regulatory constraints, boosting venture capital funding, and fostering an ethos of 'intelligent risk-taking' are critical. Such steps are vital not just for keeping pace, but for steering global innovations, leveraging our rich intellectual and educational heritage to play a pivotal role in the next technological revolution.

3

New partnership between Stack Overflow and Open AI. I will be super interested in seeing how this impacts the short and long term accuracy of LLMs in the future. In this video Theo shares a very thought-provoking picture of what this could mean for LLMs in the future. https://lnkd.in/gph_wGPs

1

2 Comments

This is a great blog from Gian Marco Iodice . Arm and Meta have a fabulous collaboration to empower developers to deploy quantized Llama models on Arm CPUs with ExecuTorch and KleidiAI. That's 4 great areas of progress in one sentence: quantization improvements, new Llama models, ExecuTorch beta release, and the great and growing KleidiAI backend. Whether you're one of over 20 millions developers targeting Arm, or one of the billions of users of Arm-powered devices, discover how you can leverage this powerful and innovative collaboration for mobile! Check it out here: https://lnkd.in/eepURWdX

13

🚀 Ever felt overwhelmed by the sheer number of tech tools out there? 😅 You’re not alone! I have been getting a lot of questions about how I approach learning new technologies, and I realized it’s time to share my insights! ✍️ In my latest article, I outlined my personal strategies for mastering new tech tools efficiently. 💡 https://lnkd.in/dxDTwA-C Whether you are a seasoned developer or just starting out, these tips can help streamline your learning process and boost your productivity. 📈Curious about my journey? 🌟 Check it out and let’s discuss how you tackle learning in this fast-paced tech world! 🤔💬 #Learning #TechTools #DeveloperJourney 

30

2 Comments