NOTE: This site has just upgraded to Forester 5.x and is still having some style and functionality issues, we will fix them ASAP.

Learning diary › Year 2025 › May, 2025 › 2025-05-06 [2025-05-06]

#agent #benchmark #blogging #context #formal #game #gpu #news #optimization #os
- LM
    - survey papers
        - A Survey on Inference Engines for Large Language Models: Perspectives on Optimization and Efficiency
        - Low-Precision Training of Large Language Models: Methods, Challenges, and Opportunities
            - sadly, NormalFloat used by QLoRA is excluded from the paper, as it is "specifically designed for pretrained fixed parameters, which only participate in the forward inference stage during training."
    - improvements on reasoning
        - Think on your Feet: Adaptive Thinking via Reinforcement Learning for Social Agents
            - the model is trained to switch between different complexity level of thinking mode based on context for an optimal balance between correct answer and reasoning length
            - the thinking modes are inspired by Hierarchical Cognitive Control Theory (HCCT)
        - RM-R1: Reward Modeling as Reasoning
            - provides good insight into reward modeling and how each training stage boosts the model's performance
    - other
        - Unlearning Sensitive Information in Multimodal LLMs: Benchmark and Attack-Defense Evaluation
            - I'm still interested in unlearning and model editing
- Benchmarking Crimes Meet Formal Verification
    - learn about Systems Benchmarking Crimes
- #zig
    - Implementing a Struct of Arrays
        - Zig has easy-to-use support of SoA via MultiArrayList
        - learn about Data-Oriented Design
        - I can't help but feel that new features of C++ such as reflection is really becoming syntax cancer
    - Reflecting on a year of Gamedev in Zig
    - Jai, the game programming contender
- bringing data to AI, not the other way around
    - The Data Surrender Trap: How Enterprises Are Losing Control in the AI Gold Rush—and the Simple Fix
- lifestyle
    - Protect The Habit
    - skimmed blogging in isolation
- skimmed modern-latex: A short guide to LaTeX that avoids legacy cruft
- The Beauty Of Having A Pi-hole
    - own your own DNS
- An appeal to Apple from Anukari regarding GPU frequency scaling
    - I'm amazed that the author is using GPU to simulate physics for realtime audio