Metacognitive Reuse: Turn recurring LLM reasoning into concise, reusable “behaviors”. The model learns named skills from its own chains-of-thought and reuses them to think faster & cheaper. Paper: arxiv.org/abs/2509.13237
A business analyst at heart who enjoys delving into AI, ML, data engineering, data science, data analytics, and modeling. My views are my own. You can also find me at threads: @sung.kim.mw
6,814 followers 1,143 following 4,485 posts
view profile on Bluesky
Metacognitive Reuse: Turn recurring LLM reasoning into concise, reusable “behaviors”. The model learns named skills from its own chains-of-thought and reuses them to think faster & cheaper. Paper: arxiv.org/abs/2509.13237
Let's say you want to improve LLM on AIME exam. You can use a sort of cheat/cram sheet by asking the LLM to use reasoning traces for old exams to extract a list of useful behaviors. Providing this list in-context while solving AIME questions improves "reasoning" (less bloat, better accuracy).
So You Want to Be an Academic? A couple of years into your PhD, but wondering: "Am I doing this right?" Most of the advice is aimed at graduating students. But there's far less for junior folks who are still finding their academic path. My candid takes: anandbhattad.github.io/blogs/jr_gra...
which is useful when counting objects in an image. Model: huggingface.co/moondream/mo...
Moondream 3 A 9B param, 2B active MoE vision hybrid reasoning vision language model that supports both reasoning and non-reasoning mode. It focus on visually grounded reasoning, where the model references objects and spatial positions in the image while doing said reasoning,
research.google/blog/deep-re...
Google's Test-Time Diffusion Deep Researcher (TTD-DR) A framework that uses a Deep Research agent to draft and revise its own drafts using high-quality retrieved information. This approach achieves new SOTA results in writing long-form research reports and completing complex reasoning tasks.
Paper: Pre-training under infinite compute ( arxiv.org/abs/2509.14786 )
The Stanford University researchers think the future of pre-training lies in the algorithms that will best leverage infinite compute. They find simple recipes that improve the asymptote of compute scaling laws to be 5x data efficient, offering better perf w/ sufficient compute.
Paper: LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures ( arxiv.org/abs/2509.142... ) Repo: github.com/rbalestr-lab...
'One small step for JEPAs, one giant leap for LLMs'." "They provided a novel solution to materialize a predictor from a (pretrained) LLM through a novel predictor token, and empirically validate LLM-JEPA on various tasks and models -> significant benefits in performance and training dynamics."
@yannlecun and his collaborators (Hai Huang and Randall Balestriero) have released their first iteration of Joint Embedding Predictive Architectures (JEPAs) for LLMs. "Learning by latent space prediction has revolutionized vision models, and it will revolutionize LLMs!, which they termed as
Dataset: huggingface.co/datasets/Byt...
, the first fully open-source agent benchmark for realistic, open-domain financial search and reasoning. Paper: FinSearchComp: Towards a Realistic, Expert-Level Evaluation of Financial Search and Reasoning ( arxiv.org/abs/2509.13160 ) Code: randomtutu.github.io/FinSearchComp/
For those of you who are developing an AI-based finance research app: I’m not sure this is useful to evaluate your apps, but they have released the paper, the code, and the dataset. ByteDance in collab with Columbia Business School released the FinSearchComp
Infrastructure: Async RL achieves a 3x speedup over Sync frameworks huggingface.co/meituan-long...
Meituan, China's largest food delivery company, is releasing more open-weight LLMs. LongCat-Flash-Thinking: Smarter reasoning and leaner costs! Performance: SOTA open-source models on Logic / Math / Coding / Agent tasks using 64.5% fewer tokens to hit top-tier accuracy on AIME25
xAI's Grok 4 Fast with a 2M context window. FREE on OpenRouter, Vercel, Cline, OpenCode, Orchids App, and more for a limited time. Grok 4 Fast is also generally available via the xAI API, with pricing starting at $0.20 / 1M input tokens and $0.50 / 1M output tokens. x.ai/news/grok-4-...
Today I’ve learned that the most successful Japanese live-action movie of 2025 not based on manga is Kokuho (『国宝』, National Treasure), which is adapted from a novel by Shuichi Yoshida. m.imdb.com/title/tt3523...
The hilarious part is that this market and department store are owned by Hanwha Group, a conglomerate best known as one of Korea’s largest defense contractors, producing everything from tanks and artillery to aerospace systems.
I always find Korean conglomerates hilarious. For example, Erewhon, a LA-based market, sells ice for $30. In Korea, there’s a comparable market called Gourmet 494, located inside the Galleria Department Store, where some of the prices are higher.
But I have this quirk: whenever I take up a new sport, I end up buying outfits and gear that make me look like a seasoned enthusiast. My latest sport is running, and I didn’t realize just how expensive it could get, especially given my personality quirks.
My clothing style is simple - I usually go for outfits with no logos or only minimal branding, except when it comes to outdoor gear (since those brands love to plaster their logos everywhere).
To those who argue this will push companies to offshore software development: it’s not as if they were ever going to stop. Offshoring will continue regardless of this new fee.
• Fortune 1000 companies: You’ve replaced seasoned professionals with body shops. Either you or the body shop is absorbing the fee - let the body shop absorb it.
My continued thoughts on Trump’s new mandate imposing $100,000 fees on H-1B visas: • Big Tech: Many experienced software engineers are unemployed. • Startups: Many graduates from top-tier CS programs are unemployed.
Marathon in Kenya - A cutoff time: 2 hour 30 minutes.
Let them open offices in other countries. H-1B visas enabled TCS, Infosys, Cognizant, and HCL to take over the IT departments of Fortune 1000 companies. What do we lose by shifting these resources to Canadian firms? It’s not as if they’re employing Americans anyway.
I do love the so called liberal pundit and technologists criticizing Trump’s new mandate on charging $100,000 fees on H1-B visas. Why wasn’t this implemented earlier. Go to any companies’ IT departments - we have outsourced all tech jobs to employees from TCS, etc, because of H1-B visas.
China - PRC. RedNote is Chinese app.
I still use RedNote once in a while. It’s a good app for wasting time, as long as you don’t engage with Chinese netizens - they are the most annoying group of people.
Parts of the world where people use a phone as a placeholder for a table or their spot in line. - Korea - Japan - Singapore and probably some UAE countries.
I wonder if carrying a phone with a lanyard will catch on in safer parts of the world. In some places, it would just be an invitation for violent mugging.
The most popular sports in Korea right now in running. It is so popular, there is a literally Runner "Subway" Station in Seoul. english.visitseoul.net/area/RunnerS...
• The Ultra-Scale Playbook: huggingface.co/spaces/nanot... • The Case for Co-Designing Model Architectures with Hardware: arxiv.org/abs/2401.14489 • Matrix Multiplication on Blackwell: Part 1 - Introduction: www.modular.com/blog/matrix-...
• Domain specific architectures for AI inference: fleetwood.dev/posts/domain... • A postmortem of three recent issues: anthropic.com/engineering/... • How To Scale Your Model: jax-ml.github.io/scaling-book/
• Defeating Nondeterminism in LLM Inference: thinkingmachines.ai/blog/defeati... • Making Deep Learning go Brrrr From First Principles: horace.io/brrr_intro.h... • Transformer Inference Arithmetic: kipp.ly/transformer-...
Supplementary information for this article as collected by Fleetwood: • How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: siboehm.com/articles/22/... • Outperforming cuBLAS on H100: a Worklog: cudaforfun.substack.com/p/outperform...
Did you know there’s supplementary information for this peer-reviewed paper (a kind of unique concept these days)? static-content.springer.com/esm/art%3A10...
Yes, ChatGPT or a similar LLM can be useful for sentiment scoring, provided you supply a well-crafted instruction prompt with around 10 example shots, and run it with a low temperature and a fixed random seed for consistency.
This was March 31, 2025 position. It is a widely known secret that U.S. government was backing Intel. Outsourcing our future to a Taiwanese company is just dumb and stupid.
The biggest holiday in Korea, Chuseok, is coming up in October. Be sure to stock up on cases of SPAM as gifts for the occasion.
46% of this guy's hedge fund was INTC call options.
Thinking of flag-jacking (e.g., an American stiching a Canadian flag on the backpack) while traveling abroad? Do your homework first — you don't want to come across as a flag-jacking fraud. www.huffpost.com/entry/flag-j...
CPU RAM and GPU VRAM. 2. For the industry: If the packaging deal proves successful, Nvidia may expand its partnership with Intel beyond this initial scope.
Bottom line: 1. For consumers: Expect very fast laptops that combine Intel CPUs with Nvidia GPUs - offering more raw power than Apple’s best M-series processors. Apple’s chips still run cooler and use a unified memory architecture, while Intel + Nvidia laptops will continue with separate
2. Nvidia investment: Nvidia plans to invest $5B in Intel, pending regulatory approval. 3. Intel exiting consumer dGPU? No. 4. Intel exiting data center GPU? No. 5. How about Data Center CPU and GPU? ???
My understanding of the Nvidia and Intel deal: 1. Will Nvidia use Intel Foundry Services (IFS)? Yes, but only for packaging. Intel has years of experience packaging both its own CPUs and TSMC-fabbed GPUs. I believe this applies mainly to mobile CPU/GPU combos, not desktop products.
Huh? EVGA essentially resold Nvidia chips with more cooling.
My take. My take: Nvidia doesn’t want to be entirely beholden to TSMC, especially with U.S. government intervention in the mix. To diversify, Nvidia may shift some lower-risk products, such as RTX GPUs, to Intel’s fabs. We will find out at 10AM PT.
I am so happy! 🥰🥰🥰
Is this AI? How is this move possible?
Oh.
Was this a Zoom meeting?
🚀 We released version 0.3 of the DuckLake specification and the DuckDB ducklake extension today. It includes interoperability with Iceberg, support for geometry types and more. Check the announcement blog for more details ducklake.select/2025/09/17/d...
The world runs on machine learning models, not AI models, like LLMs.
How We Built the First AI-Generated Genomes Here they detail some of the technical innovations that enabled us to generate viable bacteriophage genomes with substantial evolutionary novelty. arcinstitute.org/news/hie-kin...
Ladies, when you visit Korea and wonder where all the K-Drama oppas are... Just head to a department store like Shinsegae or Hyundai. Go into a luxury boutique - you’ll find them working there: tall, slim, and dressed in sharp suits. You are welcome.
Oh well… heading back to Seoul, then home. My experience: there are a lot of beautiful people, but overall it felt kind of meh - just like working at studios.
Visiting Busan, Korea. I went to the Busan International Film Festival 2025 (well, more like visited the site). It felt surreal because it’s the first film festival I’ve ever been to - even after living in Los Angeles most of my life and working at 3 film studios (Disney, Sony Pictures, and WB).
Introduction to Machine Learning Systems by Vijay Janapa Reddi It is a really long book - 2042 pages. www.mlsysbook.ai/assets/downl...
Anthropic doesn’t care if the U.S. loses its AI chip supremacy, as long as they’re making billions in stock options. Anthropic CEO criticizes administration’s chip policy, calls for stronger restrictions thehill.com/policy/techn...
In LLM pretraining, overfitting works well. 😁
I would have added a display to the Oakley Meta Vanguard, not the Meta Ray-Ban Display. The display is perfect for stat-obsessed people, who wears Garmin and uses Strava. www.meta.com/ai-glasses/o...
To those of you who are not using Docling for document conversion, processing, and parsing. Why aren't you?
Alibaba released a paper on their Tongyi DeepResearch. They incorporated Agentic Continual Pre-training (Agentic CPT) into the deep research agents training pipeline to build powerful agentic foundational models. Paper: Scaling Agents via Continual Pre-training arxiv.org/abs/2509.13310
Alibaba's subsidiary - Ant Group has released Ling-flash-2.0 as open-weight. 100B MoE LLM with 6.1B active params - 3x faster than 36B dense (200+ tokens/second on H20) - Beats ~40B dense LLM on complex reasoning - Powerful coding and frontend development
while fully preserving their layout, tables, equations, lists and more. Blog: www.ibm.com/new/announce... Models: huggingface.co/collections/...
For those of you who are using IBM Docling (github.com/docling-proj...) for document conversion, processing, and parsing; IBM has released an open-weight VLM - Granite-Docling-258M, a 258M-parameter vision-language model (VLM) for converting documents to machine-readable formats
I was wondering why store clerks kept recommending size Large… now I get it.
I am pretty sure I went a little overboard with the weight training. My jackets are fitting me very tight. I will have to shrink the muscles so I can keep wearing them.
DeepSeek published an article on Nature magazine. "DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning" www.nature.com/articles/s41...
Model: huggingface.co/unitreerobot...
This world-model provides two key functions: (a) Simulation Engine – operates as an interactive simulator to generate synthetic data for robot learning; (b) Policy Enhancement – connects with an action head and, by predicting future interaction processes with the world-model
Unitree, Chinese Robotics company, released the open-weights of UnifoLM-WMA-0 It is Unitree‘s first open-source world-model–action architecture spanning multiple types of robotic embodiments, designed specifically for general-purpose robot learning.
I am not sure about this, but some researchers may find this helpful. “My Boyfriend is AI”: A Computational Analysis of Human-AI Companionship in Reddit’s AI Community Paper: arxiv.org/abs/2509.11391
😀
Paper: arxiv.org/abs/2509.09677
Per-step accuracy of models degrades as the number of steps increases. Is this due to long-context limitations? No, the models become more likely to make mistakes when their previous errors remain in the context. "The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs"
Blog: www.spectrallabs.ai/research/SGS-1 Demo: huggingface.co/spaces/spect...
SGS-1: A foundation model for physical engineering design. SGS-1 enables the creation of manufacturable CAD geometry for real engineering workflows. This example shows SGS-1 in Fusion360 CAD software creating a bracket for a roller assembly.
I’ve noticed one thing in life: some people don’t like change, and persuading them to change is just painful. For example, AOL still has thousands of dial-up Internet customers. AOL is on sale for $1.5B. www.wsj.com/business/dea...
@unsloth.ai 's train Vision LLMs with Reinforcement Learning in free notebook! Unsloth VLM RL via GRPO: 1.5× faster, 90% less VRAM, 15× longer context & no accuracy loss. Guide: docs.unsloth.ai/new/vision-r... GitHub: github.com/unslothai/un... Colab: colab.research.google.com/github/unslo...
Designing NotebookLM by Jason Spielman He led design for NotebookLM, shaping the product’s core user experience, brand identity, and visual system from experiment to launch. jasonspielman.com/notebooklm
I upgraded to macOS 26. Now my iPhone apps are showing in the Apps view. Is there any way to hide them?
Homepage: tongyi-agent.github.io Blog: tongyi-agent.github.io/blog/introdu... Model HuggingFace: huggingface.co/Alibaba-NLP/... GitHub Repo: github.com/Alibaba-NLP/...
Alibaba launched Tongyi DeepResearch, the open-source web agent to achieve similar performance as OpenAI's Deep Research with only 30B (Activated 3B) parameters!
cloud.google.com/blog/product...
A few years ago, someone crudely built an app that let an AI agent pay a person to do work. Now, Google has formalized the idea with the launch of the Agent Payments Protocol (AP2)—an open, secure standard for AI agents to conduct transactions, debuting with support from over 60 partners.
The AI data annotation pivot continues - this time at Google.
Orchids, yet another AI Full Stack Engineer. Try it if you want. www.orchids.app
Funnily enough, Apple Watch Ultra 3 uses LEO satellites for its Emergency SOS and satellite messaging capabilities. Apple also offers two years of free satellite service for those features. Apple was expected to begin charging for iPhone satellite services, but they haven’t implemented any fees.
In the satellite Emergency SOS space, Garmin’s inReach devices are generally considered the gold standard. These use Low Earth Orbit (LEO) satellites.  Garmin’s new Fēnix 8 Pro, however uses geostationary (GEO) satellites. 
Moonshot AI's engineer also wrote about the thinking and technical story behind Checkpoint Engine. "How Kimi K2 Achieves Efficient RL Parameter Updates" moonshotai.github.io/checkpoint-e...
Do you also know that both Ilya Sutskever and Geoffrey Hinton are Canadians, and they worked together at the University of Toronto, where Hinton was Sutskever’s Ph.D. advisor.
NVIDIA's market capitalization, at $4.32 trillion as of September 15, 2025, is now greater than Canada's estimated GDP, which was around $2.24 trillion in 2024.
Someone posted a Google Research paper heralding AI can think step-by-step, using just a prompt. It received thousands of likes and hundreds of re-posts. ...but that paper is over 4 years old. Did I say there is nothing interesting in the AI sphere...
4. Nvidia sold $6.3B worth of GPUs to CoreWeave with a guarantee that they will buy any unsold CoreWeave capacity through 2032. It's a kind of a real estate lease-back deal. Again, nothing interesting.
There is nothing interesting in AI sphere. In other news, 1. HD makers like Seagate are experiencing an AI moment because of a sudden shortage of HDDs. I thought they had an abundance of capacity. 2. OpenAI releases a new version of Codex. 3. Google is popular again because of Nano-Banana.
