Daniel Cheng

Great tips about what to expect during the SWE interviews at Bloomberg!

9

These are some great tips, especially using a tool like Pramp to practice interviewing. I think once you have the foundation, you get good at interviews by doing more interviews, not by doing more problems.

8

Wondering how our entry-level interview works? Check this out!

14

Excited to share this paper about the unintended energy ramifications of the AI race between the U.S. and China! The project was spearheaded by Joyce Opal Guo, a fellow student in the Schmidt Program on Artificial Intelligence, Emerging Technology, and National Power at the Yale Jackson School of Global Affairs, taught by Edward Wittenstein. Our team was also mentored by Lucianoi Ferrari and his lab at the Yale Digital Ethics Center. Finally, I want to thank my team members: Zoey Duan, Ethan Hsu, and Chance Flanigan for the long nights and collaborative dinners. Together, we identified a "race to the bottom," in which each nation's attempts to cut each other's artificial intelligence (AI) computing resources through protectionist policies. My contribution was quantifying and visualizing the difference in energy consumption between protectionist and non-protectionist policies for case studies in the areas of chip production, data center usage, and algorithm design. Specifically, we found that energy consumption stems from six key areas: 🚨Limited access to the latest and most energy-efficient hardware; 🚨Unintended spillover effects in the consumer space due to the dual-use nature of AI technology and processes; 🚨Duplication of production processes, particularly in areas lacking comparative advantages; 🚨The relaxation of environmental standards for onshore production; 🚨The potential for use as a weapon of the renewable energy supply chain, which supports artificial intelligence infrastructure, hampering the pace of the transition to renewable energy; 🚨The loss of synergy in the advancement of artificial intelligence, including the development of more energy-efficient algorithms and hardware, due to the transition to a more autarkic information system and commerce.

28

2 Comments

Over the past few weeks, I've had the incredible opportunity to dive deep into the TextGrad framework, an innovative tool developed by Stanford University.This framework has truly opened my eyes to the power of AI-driven prompt optimization. Unlike traditional LLMs, TextGrad iteratively refines prompts based on detailed, step-by-step feedback, leading to significantly better performance. By focusing on detailed feedback, it fosters a deeper understanding and more creative solutions to complex problems. 🎯 What I Accomplished: I successfully optimized complex coding problems, including intricate balance calculation logic, by implementing advanced algorithms.This hands-on experience not only improved my coding skills but also deepened my understanding of AI-driven solutions, showcasing the framework's versatility and potential across various applications. #AI #MachineLearning #TextGrad #Innovation #Coding #DataScience

52

2 Comments

I'm excited to share that our recent research on COVID on trees and infinite grids is now hosted on Cornell University's arXiv! Grateful for the opportunity to contribute to understanding complex graph theory patterns. Check it out here: https://lnkd.in/deU9q7-r #Research #Epidemiology #Mathematics #arXiv

3

🎯 Hi everyone, I need 2 𝗜𝗻𝘁𝗲𝗿𝗻𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗖𝗼𝗱𝗲𝗿𝘀/𝗦𝘁𝘂𝗱𝗲𝗻𝘁𝘀 (𝗽𝗿𝗲-𝗰𝗼𝗹𝗹𝗲𝗴𝗲 𝘀𝘁𝘂𝗱𝗲𝗻𝘁𝘀) with programming background to Join UC Berkeley 𝗖𝗔𝗟𝗜𝗖𝗢 Fall '24 programming contest with me 🚀 I'll be competing in the CALICO Fall '24 programming contest hosted by UC Berkeley This competition is all about solving tricky algorithm problems in a limited time. I'm looking for 2 skilled competitive programmers to join my team. If you enjoy solving complex problems and want to collaborate on this challenge, let's team up! 𝗛𝗲𝗿𝗲'𝘀 𝗮 𝗾𝘂𝗶𝗰𝗸 𝗼𝘃𝗲𝗿𝘃𝗶𝗲𝘄 𝗼𝗳 𝘁𝗵𝗲 𝗰𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝗼𝗻: - There will be 6-12 problems of different difficulty levels, and we have 3 hours to solve them as a team. - You can code in C, C++, Java, or Python3. - The problems will test both your coding skills and how efficiently your code runs. - The team with the most points wins, and in case of a tie, the fastest correct solutions will decide the winner. 𝗚𝗲𝗻𝗲𝗿𝗮𝗹 𝗰𝗼𝗻𝘁𝗲𝘀𝘁 𝗶𝗻𝗳𝗼: - Date and time: November 16, 2024 at 4pm-7pm PT - Up to 3 members per team - Contest will be offered online and in person (free for online contestants) - Eligibility and contest rules: https://lnkd.in/djpG5cpg 𝗧𝗵𝗲 𝗴𝘂𝗶𝗱𝗲𝗹𝗶𝗻𝗲𝘀 𝗳𝗼𝗿 𝘁𝗲𝗮𝗺 𝗲𝗹𝗶𝗴𝗶𝗯𝗶𝗹𝗶𝘁𝘆 𝗮𝗿𝗲 𝗮𝘀 𝗳𝗼𝗹𝗹𝗼𝘄𝘀: • CALICO is open to everyone! However, teams with only pre-college students are eligible to win prizes • We define "pre-college student" as anyone who has not entered college/university If you're experienced in competitive programming on platforms like #Codeforces, #LeetCode, or #CodeChef, I'd love to work together! This is a great chance to improve our skills, learn from each other, and have fun competing. Send me a message if you're interested in joining the team. Let's make this competition one to remember! 𝗥𝗲𝗮𝗱 𝘁𝗵𝗶𝘀 𝗽𝗱𝗳 𝗳𝗼𝗿 𝗺𝗼𝗿𝗲 𝗶𝗻𝗳𝗼: https://lnkd.in/djpG5cpg 𝗖𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝗼𝗻 𝘄𝗲𝗯𝘀𝗶𝘁𝗲: https://lnkd.in/d-jgVKw2 #CompetitiveProgramming #CALICO #CodingCompetition #TeamBuilding #ProblemSolving #UCBerkeley #programming #python #c #c++ #contest #codingchallenge #calico

12

2 Comments

I'm excited to say that my two latest publications are now live on TD Commons! These papers introduce advancements in Large Language Model (LLM) fine-tuning and deployment methodologies. The first focuses on a new QLoRA-Blend methodology, which enables small LLMs to outperform larger state-of-the-art models in target domains with minimal financial investment. By integrating multiple domain-specific QLoRA adaptations using Spherical Linear Interpolation (SLERP), this technique significantly enhances accuracy and efficiency in Retrieval-Augmented Generation (RAG) systems. This work can be used to extend other full SFT, DPO, and RLHF training methods. QLoRA-Blend lowers the barrier to fine-tuning LLMs effectively by making high-performance AI more accessible and cost-effective—especially for those who are GPU-poor. The second publication proposes LLM deployment approaches that facilitate the use of LLMs on mobile, edge, and low-resource hardware. By leveraging techniques such as selective pruning, QLoRA-Blend fine-tuning, and our new approach called SynthRAG, which increases both the speed and accuracy of retrieved sources compared to traditional RAG pipelines, these methodologies achieve impressive accuracy and speed in real-time applications on consumer-grade hardware for local CPU-only inference. This work is important as it democratizes access to cutting-edge AI technology, allowing for high-level performance on widely available devices. We are excited to use these advancements in upcoming projects that will push the boundaries of what's possible with AI/ML technology. Stay tuned for the official archived version of our paper to be released soon, showcasing how we used these breakthroughs to develop an LLM that can work in enterprise networking hardware to help diagnose, analyze, and remediate problems within networks. Check out the full publications here: https://lnkd.in/g2qANSdc https://lnkd.in/g-HvKk_m

25

2 Comments

Ever wondered why LLMs struggle with math? In this piece, I take a look under the hood of LLMs and math-related AI study tools, to uncover the reasons LLMs are not well-suited for math, what innovators are doing to get ahead of the problem, and some exciting companies in the space. I explore several approaches to improve LLM math performance: process supervision, integration with symbolic computing systems (i.e. ChatGPT + Wolfram|Alpha LLC), and novel ways of capturing unique and structured data ( Thinkverse (formerly Learnie AI), Photomath, Thetawise, among many others). One startup in Korea — QANDA(Mathpresso) — has built its own math-specific LLM that can outperform many of the incumbents. If you are building AI to promote math learning and understanding, please reach out! Very grateful for the opportunity to go deeper in the EdTech space at Reach Capital this summer before kicking off my MBA/MA in Education at Stanford in the fall! Shoutout to Tony Wan and James Kim for the support on this piece! https://lnkd.in/g8xk_hqJ

49

This week at CUNY Tech Prep, we dove into the world of Retrieval-Augmented Generation (RAG) systems, utilizing Hugging Face and MongoDB Atlas Vector Database. My team used poems from Project Gutenberg on HuggingFace as our dataset. We also used Streamlit to create a quick interface. Users can enter prompts and receive quick, relevant poems based on the dataset! #CUNYTechPrep #MongoDB #HuggingFace

8

A few short days ago, I wrapped up my research position at UC Davis Health with Professor Farzad Fereidouni. My role at the lab entailed a specialization in machine learning and computer vision and I'm excited to share what I was able to accomplish with everyone. I worked on two projects during my time at the lab. The first one was something I called "Kernel Generation for Image Similarity" (https://lnkd.in/ghsY73ts). It was to process an image such that it would look like it was H&E stained. I learned a lot about the basics of histology and digital histology through this process. The most novel part of this approach was using Wiener deconvolution to generate a kernel that would be used to enhance image similarity (typically using SSIM, MSE or a combination of both). A big part of this process was optimization. I learned how to use Python libraries such as Cupy and other tools to GPU accelerate a lot of these image-processing optimization problems. My second project and my most meaningful contribution was called "FastSAM-needle-biopsy" (https://lnkd.in/gXXydq-X). The objective of this project was to use the FastSAM image segmentation network to accurately generate image segmentation masks of live tissue such that a microscope would only scan the relevant parts of an image. Here is step by step walkthrough of the image processing pipeline: 1) VB.NET (Darius) application calls a DLL (compiled from csharp.NET) 2) DLL calls FastSAM image segmentation code using Python.NET 3) FastSAM segments image and returns image segmentation bitmask and as well as the fastest path to visit every bit (Traveling Salesman Problem) 4) Python returns values are returned to DLL using JSON as a reliable data transfer solution  5) DLL returns an object that can be parsed by VB.NET FastSAM is a faster version of Meta Research's Segment Anything Model and I used it to greatly speed up development time on this project. It was a big challenge to integrate such cutting-edge, machine-learning driven, image processing pipelines into a legacy codebase that was written in Visual Basic .NET but thanks to tools such as DLLs, JSON, and C#, I was able to figure it out! I am very grateful for the time I spent in the lab. I feel like I developed some great programming skills, especially in computer vision, that will prove useful for the rest of my career. I'd like to give special thanks to Professor Fereidouni, Ruby Mascareno, Willy Ju, Dena Sayrafi, Ruben Gnanaruban, Inyola, Antara Mallick, and Jayla. The link is here if you'd like to see my lab report going over everything I did in more technical detail and with a few pictures! https://lnkd.in/gRpZVmjG #computervision #machinelearning #AI #histology

67

10 Comments

🌟 This summer, I've been diving deep into the fascinating world of AI, ML, DL, NLP, and LLM! 🚀 I've had the incredible opportunity to work on building chatbots and explore cutting-edge technologies using OpenAI, Llama, and Gamma models. The experience has been nothing short of amazing! Starting from scratch, I learned Python and built a solid foundation, thanks to the guidance of a fantastic tutor, Najaf Zawar. His expertise and support have been instrumental in my journey. I'm excited about the future and eager to continue exploring and creating innovative solutions in this rapidly evolving field. Stay tuned for more updates! #AI #MachineLearning #DeepLearning #NLP #LLM #Chatbots #OpenAI #Llama #Gamma #Python #TechJourney

9

2 Comments

Hi LinkedIn Community,   Short version:  Grad student building low-code tool for ML data pipelines. Goal: Make ML dev work enjoyable again, not just endless debugging. Ask: Shadow sessions (see Calendly link below) Longer version: As ML devs, we often leave the office thinking, “Do I actually love doing my work?” because we had spent most of the day on inefficient whack-a-mole debugging and maintenance work. I want to get us back to where we can love coding for those reasons we originally chose to pick up coding in the first place.   A lot of AI tools are restrictive and end up mangling codebases, and I believe the current state of AI coupled with strong design principles can enable much better code native developer experiences. I’ve gone through prototypes/testing and have some competing hypotheses. I’d love to observe coders working on ML data pipelines (happy to sign NDAs if necessary, can also be personal projects instead). If you're open to sharing your expertise, please schedule a shadow session: https://lnkd.in/eBW6p6qE Let's make ML data development more efficient and actually enjoyable! #MachineLearning #DataPipelines #LowCode #DevTools #AI #SoftwareEngineering #GradStudent #Research #MLOps #DataScience #ArtificialIntelligence #ProductivityTools #DeveloperExperience #CodeQuality #TechInnovation #ComputerScience #DataEngineering #CloudComputing

84

8 Comments

Week 3 at Headstarter Fellowship: Diving Deep into AI Chatbot Development 🤖 https://lnkd.in/gQs-w42f Wrapped up an intense week building a sophisticated AI-powered chatbot. We started with a Next.js foundation, leveraging its API routes for our backend needs. The frontend came to life with Shadcn UI components styled with Tailwind CSS for that sleek, modern look. But the real magic happened when we integrated the AI. We implemented Retrieval Augmented Generation (RAG) using Langchain and Pinecone, which allows our chatbot to pull relevant information from a specified knowledge base (Right now it just has some old information about me). Then came the fun part: we integrated both Groq's Mixtral-8x7B and OpenAI's GPT-4o, building a custom routing mechanism to choose the best model based on query complexity. The chatbot isn't just smart; it's multilingual too! We added support for English, French, and German responses. And to keep things dynamic, we implemented streaming responses - it's pretty cool to see the AI "think" in real-time. While each of these components might seem straightforward on its own, the real challenge was orchestrating everything into a seamless, cohesive application. It's one thing to implement RAG or integrate multiple AI models, but to make them work together smoothly? That's where the real fun begins! There's still room for improvement, of course. We're looking at refining the query complexity determination and optimizing RAG performance. But I'm proud of what we've accomplished in just a week. It's a testament to how quickly we can build complex AI applications with modern tools and frameworks. This project has been a fantastic learning experience, thanks to Bill Zhang for the great help, pushing me to think holistically about AI application architecture. Can't wait to see how we'll apply these skills in future projects! #HeadstarterFellowship #AIEngineering #ChatbotDevelopment #HeadstarterAI Faizan Ahmed

5

🚨🚨 Preprint alert! 🚨🚨 I'm excited to announce that my internship work at NVIDIA on alignment for text-to-image (T2I) models has culminated into the preprint "Elucidating Optimal Reward-Diversity Tradeoffs in Text-to-Image Diffusion Models" TL;DR: We propose "Annealed Importance Guidance", an inference-time method to improve both image generation diversity and image quality for text-to-image models finetuned on human preference data. Long version: Open source text-to-image models (like StableDiffusion) are trained on noisy internet-scale image-text pair data, necessitating further alignment using human preference datasets and reward models. However, finetuning algorithms are vulnerable to a phenomenon called "reward hacking" where the model produces high-reward images but loses all its initial diversity. We provide a simple and elegant mathematical proof for why this occurs, followed by an analysis of existing regularization methods that are adapted from the LLM-finetuning literature. We observe that unlike LLMs where the entire output token sequence constitutes the response (requiring regularization on the full token sequence), the intermediate noisy images from the diffusion model play asymmetric roles in image generation (diversity in the earlier timesteps versus finegrained details in the later timesteps), rendering existing regularization techniques highly suboptimal. To improve the reward-diversity tradeoff, we tackle this issue using "Annealed Importance Guidance", an inference time regularization that uses a finetuned models and dynamically adjusts its regularization over the sampling chain to obtain high-diversity and high-quality images. An extensive user study shows overwhelming preference of our method over state-of-the-art finetuning method DRaFT, that comes with more than a 2x preference in image quality with no loss in text-to-image alignment. Moreover, we reach a significant Pareto-improvement over other regularization methods on four diversity metrics. For more details, check out the links below 👇 I'm extremely grateful to Nima Tajbakhsh, Ali Taghibakhshi, and Ashwath Aithal and the entire NeMo team for this fantastic opportunity! The code for this project is tightly integrated into the NeMo and NeMo-Aligner libraries, adding on to the arsenal of versatile fine-tuning strategies for LLMs and T2I models. 📂 Project page: https://lnkd.in/gjDNXVsP 📝 ArXiv: https://lnkd.in/g488n3KR 🖥️ Code: https://lnkd.in/gpYNZXUj

202

7 Comments

I recently released a paper folks may find interesting. We study how well various editing, compression, and unlearning techniques affect one another when applied to the same LM. Having LM interventions that compose well is important since, in practice, we may want to apply multiple interventions to the same model, such as updating facts, unlearning knowledge, and compressing (quantizing/pruning).

38

I’m thrilled to share my resume for the 2024 placement season! I'm actively seeking full-time roles in Machine Learning Engineering, Software Development, and Quantitative Analysis. If you know of any opportunities or would like to connect, feel free to reach out! #fulltimeopportunity in #softwaredevelopment and #quantitativefinance #MachineLearning #MLEngineer #DeepLearning #AI #SDE #IITKanpur #SoftwareDevelopment #SoftwareDeveloper #QuantitativeFinance #QuantitativeTrading #QuantitativeResearch #QuantitativeStrategist #Quant #placements

140

30 Comments