As a practitioner for note taking (sometimes evergreen notes), I finally find it not working well when splitting note-taking in different dates with separate files. So I decide to use a holistic note-keeping “place” to hold all the notes and here it is. I hope I can keep writing here obeying the grammar of markdown.
Table of Contents:
25-08-11
<aside>
💡
Overtrained language models are harder to fine-tune.
- I want to thoroughly learn this paper because I want to know more about the mystery of LLM over-training, that is the question:
- Does training the model heavily with lower training loss always come to better downstream fine-tuning performance?
- Three parts construct the main content of the paper:
- Extended pre-training hurts post-training
- Experiments on olmo-1b/7b and llm360-amber-7b, 3 model families with intermediate checkpoints
- Two observations: 1). extended pre-training always benefit base model, 2). extended pre-training beyond certain budget might hurt task-specific post-training for both ID and OOD tasks
- Catastrophic overtraining
- A theoretical perspective of overtraining
</aside>
<aside>
💡
Diffusion language models are super data learners, released on Aug. 09 2025.
</aside>
<aside>
💡
ovatarl: training language model from scratch with pure reinforcement learning, Aug. 9 2025.
</aside>
<aside>
💡
Statistical suggestions for mech interp research and beyond, Aug. 6 2025.
- Read this blogpost to learn and review p-values.
</aside>
<aside>
💡
Assessing skeptical views of interpretability research, by Prof. Chris Potts.
- Interpretability might be my whole life belief in the science of data-driven intelligence like deep learning and its application in language understanding.
- I can read this blog to see several critiques of the current approach.
</aside>
<aside>
💡
System prompts of LLM vendors and Agent vendors
- system-prompts-and-models-of-ai-tools: FULL v0, Cursor, Manus, Same.dev, Lovable, Devin, Replit Agent, Windsurf Agent, VSCode Agent, Dia Browser, Xcode, Trae AI, Cluely & Orchids.app (And other Open Sourced) System Prompts, Tools & AI Models.
- system_prompts_leaks: Collection of extracted System Prompts from popular chatbots like ChatGPT, Claude & Gemini.
</aside>
<aside>
💡
yet-another-applied-llm-benchmark
- My own customized LLM benchmark https://github.com/carlini/yet-another-applied-llm-benchmarkMotivated by Carlini’s benchmark repo above, I also want to create my own customized LLM benchmark.
- This recent impulse is also resulted from recently newly published LLM models, for example, Kimi K2, GLM-4.5, Claude 4.1 and GPT-5.
- Since those models might be benchmaxed, it is getting harder to evaluate their intelligence, so how to properly evaluate them given a new task is the core question I really care these days. This is also the question of testing LLMs’ realistic generalization ability.
</aside>
<aside>
📃
Paper buf
xxx
</aside>
<aside>
📃
RL2: Ray less reinforcement learning: A concise library of reinforcement learning for large language models.
- Supports SGLang.
- RL2 is a production-ready library!
</aside>
25-08-13
<aside>
📃
The 2025 IMO: A blogpost that rethinks LLM’s IMO breakthrough.
</aside>
<aside>
💡
gpt-oss-reverse-engineering: A repo that analyzes prior knowledge in those LLM’s, by priming LLMs with several sentence leading token, such as ‘What’, ‘The’, ‘How’ etc. And then analyze the characteristics of the generated content.
</aside>
25-08-14
<aside>
💡
There I have a question: given a set of short phrases denoted as $\mathcal{P}$, and you know they have certain relationships such as:
- identical
- hypernymic
- part-of
- accessory-of
- situated-to-be
</aside>
25-08-15