Gabriel Wong

GO COMMENTARY WEEKLY: FRAMEWORK & TOOLS FOR THE LLM ERA Welcome to Go Commentary by Phat Nguyen (backend engineer), exploring the latest developments in the Go programming ecosystem. Here we take a closer look at how Go is adapting to the rise of Large Language Models (LLMs). New frameworks like LangChainGo and Genkit are emerging to help developers integrate LLMs into their applications, expanding Go’s role in AI. - Go Survey: The Go team is seeking feedback from developers on how Go is used, challenges, and desired features to improve community support. - LLMs in Go: Go is being positioned as a solid choice for building LLM-powered applications, thanks to its strengths in concurrency and networking. - RAG Server Example: A simple Retrieval-Augmented Generation (RAG) server implementation demonstrates Go’s straightforward approach to LLM integration. - LangChainGo, Genkit, Ergo: Introducing tools for LLM tasks, production-ready features, and distributed systems, raising questions about balancing simplicity and innovation in Go's ecosystem. Read Dwarves Go Weekly at: https://lnkd.in/g5VKE9vf #GoLang #LLM #dwarves #software _ Dwarves Notes (https://memo.d.foundation/) combine our team’s collective know-hows, R&Ds, and operation approaches. Connect and learn alongside other tech fellows: - Discord: discord.gg/dwarvesv - Github: https://lnkd.in/gZZ2eZMu - Website: https://d.foundation

3

[Image linked] Buckle up, tech enthusiasts! 🚀 Our understanding of neural networks is getting a whole new framework, thanks to Jiaming Luo. This graduate student's intriguing discovery sets the stage for Liquid #AI - a game-changer that's turning static neural networks dynamic. Let's applaud this innovation that could revolutionize data interpretation.👏🎓 But, predictably, this shift brings unique challenges regarding transparency and accountability. Want to dive into the depths of this compelling breakthrough? Follow the link to satiate your curiosity: 👉 https://lnkd.in/en9A-UiZ #ArtificialIntelligence #NeuralNetworks #TechnologyNews

2

Check out how GovTech Singapore worked with the Ministry of Manpower to deploy AI solutions, including an automatic job posting classifier and an AI-powered Sensemaker that can extract and summarise insights from large volumes of documents! Henry Chang recently spoke about these innovative projects that combine generative AI and traditional AI at the Amazon Web Services (AWS) summit yesterday (https://lnkd.in/g5xxEMaA) and GovInsider has previously showcased the innovative Analytics.Gov here (https://lnkd.in/gbuV4tPp). How else can government organisations combine traditional and generative AI to make lives easier for staff and citizens? 💡

24

1 Comment

Digital Promise's Xin Wei unveils new insights in a study on AI and mathematics education, now published in Frontiers. In “Evaluating ChatGPT-4 and ChatGPT-4o: Performance Insights from NAEP Mathematics Problem Solving," Wei’s assessment of the capabilities of ChatGPT-4 and ChatGPT-4o in solving mathematics problems from the NAEP exams across grades 4, 8, and 12 are highlighted. The findings show that AI has the potential to support instruction in specific mathematical areas, but there remains a need for teacher-mediated strategies in areas where AI is less effective. https://bit.ly/3ZwyFej

14

1 Comment

🎉 Great News from Entropica Labs! 🎉 We are happy to share some of our latest work on #FaultTolerance and #QubitRouting. Congratulations to the entire team - Chiew Shao Hen, Vishal Sharma, Chai Jing Hao, Ezequiel Ignacio Rodríguez Chiacchio and Hui Khoon Ng - for bringing this project to life! 🙌 Our team has demonstrated that, by using reasonably well-behaved Swap gates, one can preserve the fault-tolerance design of an abstract quantum circuit when mapping it to a physical circuit with a different topology. 🚀 This property is particularly relevant when working with the planar surface code on hardware with heavy hexagonal lattices (like the IBM Quantum platforms). We show how our routing schedules preserve the fault-tolerant nature of the surface code error-correction circuit without the need to modify the surface code circuits to have a hexagonal interaction graph. 📄 Find the paper here: https://lnkd.in/gGyexACH ✨ At Entropica Labs we are working hard towards building the software tools necessary to make fault-tolerant quantum programs a reality across all quantum computing platforms. Stay tuned for more results in the coming months as we continue to improve the design of fault-tolerant quantum programs. #QuantumTech #Innovation #TechContent #Milestone #TeamWork #QuantumComputing

63

Last week, Machine Learning Engineer Yue Liu facilitated our weekly internal 'Learning Breakfast' about model distillation. As large language models (LLMs) like Llama 3.1 become open-source, an exciting question emerges: How can we leverage these massive models to create smaller, efficient ones without losing performance? 🤔 📉 Model distillation and why it matters 📉 • Model distillation offers a solution by transferring knowledge from large models (teachers) to smaller models (students). This addresses the need for faster, more lightweight models that can run on lower-powered hardware while delivering high-quality results. • Distillation allows smaller organizations and research institutes to leverage open-source large models as teachers, enabling them to train smaller, efficient models that suit their specific needs. 💡 Key insights 💡 • Dark knowledge: Training smaller models on soft labels from larger models captures complex patterns, allowing student models to achieve results closer to their teachers. • On-policy distillation: This method allows student models to improve by learning from their own outputs under the guidance of the teacher model, which makes them more robust to errors and adaptive over time. However, it’s also the most computationally demanding form of distillation, as it requires continuous generation and evaluation of student’s outputs throughout training. Thank you Yue for your valuable insights! #MachineLearning #ModelDistillation #AIOptimization #KnowledgeTransfer

32

4 Comments

Being deeply interested in AI and Machine Learning, I realized I needed to strengthen my understanding from the ground up to build a strong foundation for my future in the industry. After coming across the Specialized Machine Learning course from Stanford University taught by Andrew Ng,I came to realize why he is considered a legend in the industry. The mathematics behind the apparently mysterious machine learning was actually fascinating and intriguing. The Specialized Course was divided into 3 parts, 1.Supervised Machine Learning: Regression and Classification. Where I built robust models using Linear and Logistic Regression and explored techniques like Gradient Descent for optimization. Implemented Regularization strategies to prevent overfitting and ensure models generalize well. Also gained hands-on experience using scikit-learn and NumPy for efficient data handling and model training. 2.Advanced Learning Algorithms Where I designed and trained Artificial Neural Networks (ANNs) using TensorFlow for multi-class classification tasks. Developed Decision Trees and optimized them with ensemble methods like Random Forests and XGBoost to tackle complex decision-making scenarios. Focused on applying best practices for model development, including error analysis and data-centric approaches to enhance model performance. 3.Unsupervised Learning, Recommenders, and Reinforcement Learning Where I implemented Clustering algorithms and Anomaly Detection to uncover hidden patterns and outliers in datasets. Built personalized Recommender Systems using collaborative filtering and deep learning methods. Explored Reinforcement Learning by creating models that learn through trial and error, demonstrating the powerful intersection of ML and decision-making. Also through this specialization, I got to work on industry-relevant projects that simulate challenges faced in AI development, such as: Optimizing Retail Sales Predictions: Leveraged regression models to forecast sales based on historical data, inventory levels, and other factors, helping optimize stock and improve supply chain efficiency. Customer Segmentation and Targeting: Applied clustering techniques to segment customers based on purchasing behaviors, allowing for targeted marketing strategies and personalized experiences. Anomaly Detection in Financial Transactions: Developed models to identify fraudulent transactions and minimize financial risks, using unsupervised learning to catch anomalies without extensive labeled data. Building Robust Recommendation Engines: Created a hybrid recommender system combining collaborative filtering and content-based approaches, improving recommendation accuracy in streaming platforms.

5

#ICYMI: At the #DARESymposium earlier this year, A/Prof Tongliang Liu shared his research on causal discovery! Since it is typically expensive and time-consuming to conduct randomised experiments, there has been significant attention towards revealing causal relations through the analysis of purely observational data, commonly known as #CausalDiscovery. Over the past few years, with the rapid development of big data, causal discovery is facing great opportunities and challenges. In this presentation, Tongliang introduces some classical causal discovery methods, which have been successfully applied to cases without latent variables. However, in complex systems, we typically fail to collect and measure all task-relevant variables. He also discusses causal structure recovery in the presence of latent variables. Engineering and Computer Science - University of Sydney | #MachineLearning #ML #DataScience #causality

2

Unlocking Insights 📊 with Python! In this two-part series, Ong Yi Shen, our intern from NTU, shares some nifty tricks to clean, analyse, and report data from CSV files. Click the link below to learn more! #DataScience #PythonProgramming https://lnkd.in/gn2TUgRy

1

Am currently learning NN, and created a Notion pages where I'll re-explain it on my why ( which is going to be so simple ). Feel free to join/correct me. Notion: https://lnkd.in/gmQMrbZ5 My learning sources: - https://lnkd.in/gkZwqyP6 - https://lnkd.in/gdUJXX_p - Hands On Machine Learning with Scikit Learn and TensorFlow #neuralnetwork #ml #nn

6

Reinforcement learning (RL) is commonly used to finetune LLMs based on human feedback, but did you know that we can also use RL to automatically improve our prompts? Prompt engineering is the act of discovering (via trial and error) prompts that perform well. Small changes or rewordings to our prompt (or using a different LLM) can make a massive difference. For this reason, prompt engineering requires a lot of manual–and tedious–effort. Optimizing prompts: To automate this manual labor, researchers have proposed techniques for learning prompts from data. For example, prompt and prefix tuning [1, 2] add new (continuous) tokens to the prompt and finetune them. Methods like AutoPrompt [3] follow a similar strategy but keep learned tokens discrete so that they are human interpretable. Discrete optimization of tokens: The main difficulty with optimizing a prompt is that the prompt’s tokens are discrete. If we use the gradient to update a token, the likelihood of this update producing another valid token within the model’s vocabulary is effectively zero. Optimizing discrete tokens is not possible via traditional gradient-based algorithms. LLM generating prompts: One way we can optimize a prompt is by having a (separate LLM) output the prompt. Then, we can train the weights of the LLM–which are continuous–to produce better prompts instead of training the discrete prompt directly. But, how do we optimize this LLM? There’s no differentiable training signal that can be used for this. Using RL: To optimize the LLM generating the prompt, we can use RL! Our policy network is the LLM that generates the prompt. This policy network will output a prompt using next token prediction. Once a full prompt / sequence is generated, we derive the reward for this generation by measuring the prompt’s performance on a downstream task. Practical examples: Both RLPrompt [4] and TEMPERA [5] use RL to optimize discrete, human interpretable prompts. RLPrompt optimizes prompts in an offline fashion (i.e., optimize a prompt once and use it for inference many times), while TEMPERA optimizes prompts dynamically at inference time. “The resulting optimized prompts are often ungrammatical gibberish text; and surprisingly, those gibberish prompts are transferrable between different LMs to retain significant performance, indicating that LM prompting may not follow human language patterns.” - from RLPrompt paper Despite performing well, the prompts discovered by these techniques tend to be gibberish / ungrammatical, leading us to wonder whether prompting is a language of its own. The best prompts do not always follow standard rules applied to natural language.

230

8 Comments

Finished reading the 3rd chapter of Chip Huyen's Designing Machine Learning Systems which entitles Data Engineering Fundamentals. It's very eye opening for me as I generally do the machine learning part and depends on my data engineeer teammates. This chapter gives me some trouble as I needed to go through several other resources while reading this chapter to brushen up my understandings. My key findings are, 1. It's a footnote that sounded very interesting to me as once replied the same. Like, one of the author's friends told him that they uses user's historial product browsing and purchases to make recommendation. When author asked why they don't use personal data. He looked confused and replied if she meant demographic data. They thought browsing and perchasisng activities are extremly personal. So which data to use, is really a very key questions. 2.Row major formats are better for when a lot's pf data to be written whereas column major ones are better if there is many column based readings. 3.File format is a crucial thing. 4. Relational and No SQL db both jave their one virtues. I would love to remember Aminur Rashid, who sometimes become very creative and once thought of rearranging a flat transectional table into a graph structure to show his more creativity. 5. I was a bit confused of data warehouse and data lake this types of jargons. It hase a very clear answer. Repository having structured data is data warehous and data lakes are usually used to store raw data before processing. 6. Transectional databases must have ACID property, that's I have been learnt so far. But I didn't know there is another school of thought, BASE, Basically available, Soft state, and Eventually consistent. Though this defination is more vague than the definations of ACID. 7. Different use cases of ETL and ELT. 8. 2 ways of data passing through databases, data passing through services lke using RSET abd data passing through real time like PUb-Sub model and message queue. 9. This chapter ends with a detailed discussion on Batch Processing vs Stream Processing. I would love to read back this chapter again in a few days. #mlsystemdesign #machinelearning #machinelearningengineer #mlops #ai #aicommunity

15

What Is Computation - Does Life Compute? = I am doing a series of posts which started as my reaction to a recent post by Ilya Fastovets: lnkd.in/gJ3Bw_uX First I needed to explain the hard problem of Consistency. This is known as: Hilbert's Second Problem In mathematics, Hilbert's second problem was posed by David Hilbert in 1900 as one of his 23 problems. It asks for a proof that arithmetic is consistent – free of any internal contradictions. Hilbert stated that the axioms he considered for arithmetic were the ones given in Hilbert (1900), which include a second order completeness axiom. In the 1930s, Kurt Gödel and Gerhard Gentzen proved results that cast new light on the problem. While the theorems of Gödel and Gentzen are now well understood by the mathematical logic community, no consensus has formed on whether (or in what way) these theorems answer Hilbert's second problem. Simpson (1988) argues that Gödel's incompleteness theorem shows that it is not possible to produce finitistic consistency proofs of strong theories. Kreisel (1976) states that although Gödel's results imply that no finitistic syntactic consistency proof can be obtained, semantic (in particular, second-order) arguments can be used to give convincing consistency proofs. Detlefsen (1990) argues that Gödel's theorem does not prevent a consistency proof because its hypotheses might not apply to all the systems in which a consistency proof could be carried out. Dawson (2006) calls the belief that Gödel's theorem eliminates the possibility of a persuasive consistency proof "erroneous", citing the consistency proof given by Gentzen and a later one given by Gödel in 1958. So Mathematics crashed due to Hilbert's Program and could not recover fully till today. = Then in the heat of the arguments, I claimed AI is not based on Logic and Mathematics. Immediately I was under attack and I had to do a defensive post: How I Managed To Get My Foot Out From My Mouth (link in comment) In that post I doubled down on my initial assertion that: AI is not based on Logic and Mathematics So what is AI based on? AI is based on ML hacks which apparently works. I could do that post since definition of Artificial Intelligence is very loose. Sometimes AI acts very smart, sometime that same AI acts infinitely dumb. It is only mimicry going on with AI. AI is illogical and not mathematical. If AI is based on Logic and Mathematics then how can AI be illogical and not mathematical?? My writing journey continued with the next post: Do you have a Rock Brain? (link in comment) Finally people started asking questions like: Hey what is Computation? This question was asked in the context of: Does AI Think or Compute? What is Learning? Thanks to people like Sean Kempton, Matthew Kamerman, Rodney Beard et al. In this post I am revealing that I am addressing: What is Computation? See the PDF below from 2023 Turing Award winner!!

24

22 Comments

How does OpenAI learn the Strawberry (o1) model to think? Reinforcement Learning (RL) and Chain of Thought (CoT) 1. Reinforcement Learning (RL): - RL with Human Feedback (RLHF): This technique involves using human feedback to guide the learning process of the model. Ilya Sutskever's involvement suggests that RLHF is still a significant part of the training process for the Strawberry model. - Reward Model: In RLHF, the model's outputs are evaluated by a reward model, which assigns scores based on human preferences. These scores are used to update the model to align better with desired behaviors. 2. Chain of Thought (CoT): - CoT Tokens: These are intermediate steps or thoughts that the model generates before arriving at a final response. Jason Wei's involvement indicates that CoT is crucial in the alignment process. - Integration with RLHF: In the enhanced RLHF+CoT approach, CoT tokens are also fed to the reward model. This helps the model refine its thinking process, not just the final output. Training Process - Training Phase: During training, the model learns to generate CoT tokens and uses the feedback from the reward model to improve its alignment with human preferences. - Inference Phase: At inference time, the model starts by generating CoT tokens, which can take up to 30 seconds, before producing the final response. This ensures the model spends more time "thinking" and refining its answers. Missing Technical Details (OpenAI report) - Reward Model Training: The specifics of how the reward model is trained and how human preferences for the thinking process are elicited are not detailed in the report. - Human Feedback: The process of collecting and integrating human feedback for the thinking process is also not fully explained. #openai #reinforcement_learning #AI #LLM #deep_learning

9

1 Comment

This new #FoundationsofDataScience paper aims to empower statisticians and data scientists to engage with #QuantumComputing literature and collaborate with quantum algorithm designers, and ultimately drive the development of the next generation of statistical and data analytics tools. The review, by DARE researcher A/Prof Minh-Ngoc Tran and affiliates Anna Lopatnikova and Scott Sisson, provides an introduction to quantum computing tailored to statisticians and data scientists. The objective is to equip them with a comprehensive framework, the basic language, and building blocks of quantum algorithms, as well as an overview of existing quantum applications in #statistics and data analysis. Quantum computers promise to surpass the most powerful classical supercomputers in tackling critical practical problems, such as designing pharmaceuticals and fertilisers, optimising supply chains and traffic, and enhancing #machinelearning. Read the paper here: https://lnkd.in/gTnxWYqm American Institute of Mathematical Sciences | The University of Sydney Business School | UNSW Data Science Hub (uDASH) | ARC Centre of Excellence for Mathematical and Statistical Frontiers (ACEMS)

19

As AI and ML continue to revolutionize various industries, ethical concerns around bias in AI systems are at the forefront. Our latest article, with insights from Jenny Chun Chi Lien 連君綺, Data Science Manager at AltaML, navigates the pressing issue of harmful bias in AI and the role of open-source toolkits in addressing it. Learn how these toolkits offer solutions for identifying, explaining, and mitigating bias in AI systems: https://bit.ly/3UyoK3w #machinelearning #artificialintelligence #AI #ML #ethicalAI #responsibleAI #toolkits #innovation #data #technology

19

This is a modern and comprehensive take. From language representation, sequence models, transformers, LLMs, agents and tool use. Highly recommended.

21