AI
This past weekend, I participated in Hackathon@Sea 2025, an internal event hosted by my company. Unsurprisingly, the theme was the "AI Agent," a topic surrounded by immense hype and speculation. The experience provided a clear snapshot of the current landscape of AI development, crystallizing my thoughts on
AI
Lilian Weng's recent article, "Why We Think", provides an excellent overview of the current state-of-the-art in AI reasoning. Intriguingly, and perhaps unsurprisingly, the methods employed to enable AI to "think" closely mirror human cognitive processes. These parallels are summarized in the table below. Table
AI
While forecasting how AI will impact my career in Consumer Lending Credit Risk Management (Article Link), I explored the timeline forecast methodology used in AI 2027. Prerequisites Time Horizon The concept of time horizon, proposed by the Model Evaluation & Threat Research (METR) team, states that the "50%-task-completion
AI
Recently, I've been exploring the theory of computation. With the rapid advancement of artificial intelligence—essentially a vast collection of algorithms and computational instructions designed to process inputs and generate outputs—I find myself increasingly curious about the fundamental capabilities and limitations of computation itself. Concepts such as
AI
Building neural networks often feels like magic. We import libraries like PyTorch or TensorFlow, stack some layers, call .fit(), and voilà! But what's happening under the hood? To truly understand concepts like backpropagation, I decided to follow Andrej Karpathy's fantastic 'nn-zero-to-hero' series and build
AI
Link to the article: https://www.anthropic.com/research/building-effective-agents One Paragraph Summary Agents are systems where Large Language Models (LLMs) dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. They can start with a command or interactive discussion, and then gain "ground
AI
Yuval Noah Harari's Nexus offers a compelling new lens for understanding human history and our current predicament: the analysis of information networks. Harari argues that understanding how information connects people and things, rather than focusing solely on technological advancements, is key. He emphasizes that information's power
AI
When I decided to create a Slack bot to remind me of my due Anki flashcards, I didn’t find the concept intimidating. The steps—accessing Anki’s data, posting to Slack, and setting up a cron job—were all within my capability. However, I’m not a seasoned software
AI
AI tools have opened up a world of possibilities, enabling us to accomplish tasks more efficiently and creatively. Here are some of the exciting things I've tried recently with AI: 1. Create a Webpage with Just a Chatbox Using tools like Cursor, you can: * Build a webpage by
AI
In essence, "Low Training Error + Sufficient Data Relative to Model Complexity → Low Test Error". Supervised learning must work on the SAME distribution. Key Concepts * Training Error (\(\text{Train}_S(f)\)): The error of a model \(f\) evaluated on the training dataset \(S\). * Test Error (\(\text{Test}_D(f)
![[Study Note] V and Q Function in RF](/content/images/size/w600/2024/10/Snipaste_2024-10-27_17-50-53.png)
AI
UC Berkeley's CS285 course: lecture 4. The goal of reinforcement learning is to find an optimal policy \(\pi^*\) that maximises the cumulative rewards over time. Define the trajectory \(\tau = (s_0,a_0,s_1,a_1,...) \), which is a sequence of states and actions. The probability under