Sung Kim

Free eBook: Machine Learning Systems

Principles and Practices of Engineering Artificially Intelligent Systems

This textbook bridges the gap between theoretical foundations and practical engineering, emphasizing the systems perspective required to build effective AI solutions.

www.mlsysbook.ai

Paper: arxiv.org/abs/2510.25781
Repo: github.com/AmirNoori68/...

A Practitioner's Guide to Kolmogorov-Arnold Networks

A systematic and comprehensive overview of the rapidly expanding KAN landscape, moving beyond simple performance comparisons to offer a structured synthesis of theoretical foundations, architectural variants, and practical implementation strategi

Google: TPU + JAX and killing it.
Meta: ??? + PyTorch ...and you had a plan for years, maybe even a decade; where's the execution? ...perhaps partner with Microsoft?

On the left, the refurbished Lincoln bathroom. On the right, picture I took in Saddam Hussein's palace in Basra in 2005.

Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning Large Language Models (LLMs) often struggle with problems that require multi-step reasoning. For small-scale open-source models, Reinforcement Learning with Verifiable Rewards (RLVR) fails when correc...

arxiv.org/abs/2510.25992

Google's Supervised Reinforcement Learning (SRL), a method designed to teach LLMs complex reasoning skills from expert demonstrations, for problems that are too difficult for conventional RL or SFT approaches.

"Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning"

Making Decisions under Model Misspecification We use decision theory to confront uncertainty that is sufficiently broad to incorporate "models as approximations." We presume the existence of a featured collection of what we call "structured model...

To address this "misspecification fear", they propose a decision criterion that evaluates outcomes by minimizing across alternative "unstructured models" while imposing a penalty based on the Hausdorff statistical set distance (C(p,Q)) from the original set Q.

arxiv.org/abs/2008.01071

In economics and science, "model misspecification" means that the specific models we use to predict the future or guide decisions are always simplifications and approximations, so they are inherently "wrong" or they don't perfectly represent the complex real world.

Project: emu.world
Model: huggingface.co/collections/...
Github: github.com/baaivision/E...
Paper: arxiv.org/abs/2510.26583

BAAI's Emu3.5

A large-scale multimodal world model that natively predicts the next vision-language state. Emu3.5 outperforms Nano Banana across image generation, editing, interleaved tasks and more.

Paper: arxiv.org/abs/2510.26788
Repo: github.com/sail-sg/Prec...

I like their simple conclusion.

Problem: The BF16 precision causes a large training-inference mismatch, leading to unstable RL training.

Solution: Just switch to FP16.

"Defeating the Training-Inference Mismatch via FP16"

I can’t help but wonder about this photo - I thought these two companies (Hyundai and Samsung) and families (Chung and Lee) HATE each other.

Moreover, Intel’s track record with past AI processor acquisitions has been disastrous.

Intel may purchase SambaNova.

Good: It’s a solid ASIC company - the only provider capable of running inference on DeepSeek, unlike Cerebras or Groq.

Bad: Lip-Bu Tan serves as the company’s Chairperson, and SambaNova is facing funding difficulties.

Rethinking Thinking Tokens: LLMs as Improvement Operators Reasoning training incentivizes LLMs to produce long chains of thought (long CoT), which among other things, allows them to explore solution strategies with self-checking. This results in higher accur...

"Rethinking Thinking Tokens: LLMs as Improvement Operators"

arxiv.org/abs/2510.01123

Reasoning training encourages LLMs to produce long chains of thought (CoT), improving accuracy via self-checking but increasing context length, compute cost, and latency. This work studies whether frontier models can achieve better trade-offs, higher accuracy with lower cost.

The Smol Training Playbook: The Secrets to Building World-Class LLMs

A practical journey through the challenges, decisions, and messy reality behind training state-of-the-art language models.

huggingface.co/spaces/Huggi...

Project: marin.community
Model: huggingface.co/marin-commun...
Repo: github.com/marin-commun...
Doc: marin.readthedocs.io/en/latest/
Discord: discord.com/invite/J9CTk...

Marin 32B Base - mantis (Open-source: Model, Code, and Data)

A key feature of Marin is reproducibility: every step, from raw data to the final model are recorded, not just the end result. This includes failed experiments, so the entire research process is transparent.

moonshotai/Kimi-Linear-48B-A3B-Instruct · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

usage and up to 6x decoding throughput at a 1M context length.

Model: huggingface.co/moonshotai/K...
Tech Report: github.com/MoonshotAI/K...

Moonshot AI's Kimi Linear (Open-weight)

A novel architecture that outperforms full attention with faster speeds and better performance—ready to serve as a drop-in replacement for full attention, featuring our open-sourced KDA kernels! Kimi Linear offers up to a 75% reduction in KV cache

LFM2-ColBERT-350M: One Model to Embed Them All | Liquid AI Today, we release LFM2-ColBERT-350M, a late interaction retriever with excellent multilingual performance. It allows you to store documents in one language (for example, a product description in Engli...

Full Blog: liquid.ai/blog/lfm2-co...
Model: huggingface.co/LiquidAI/LFM...
Demo: huggingface.co/spaces/Liqui...

LiquidAI's (MIT CSAIL offshoot) LFM2-ColBERT-350M

350M parameters embedding model that allows you to store documents in one language and retrieve them in many languages with high accuracy and inference speeds of models a fraction of its size.

🔗 Homepage: tongyi-agent.github.io
🔗 Technical Report: arxiv.org/pdf/2510.24701
🔗 Blog: tongyi-agent.github.io/blog/introdu...
🔗 Model HuggingFace: huggingface.co/Alibaba-NLP/...
🔗 Model ModelScope: modelscope.cn/models/iic/T...
🔗 GitHub Repo: github.com/Alibaba-NLP/...

Alibaba's Tongyi DeepResearch Technical Report

Dive deep into the technology and insights behind our 30B (A3B) open-source web agent that achieves SOTA performance: 32.9 on Humanity's Last Exam, 43.4 on BrowseComp, and 46.7 on BrowseComp-ZH.

A Geometric Analysis of PCA

What property of the data distribution determines the excess risk of principal component analysis? In this paper, they provide a precise answer to this question.

arxiv.org/abs/2510.20978

The Principles of Diffusion Models

It traces the core ideas that shaped diffusion modeling and explains how today’s models work, why they work, and where they’re heading.

www.arxiv.org/abs/2510.21890

Everything About Transformers: A visual story of how transformers came to life

www.krupadave.com/articles/eve...