Sung Kim
@sungkim.bsky.social
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
It seems memory and storage vendors, including Micron, Samsung, SK hynix, Sandisk, Seagate, and Western Digital, are effectively capping gross margins at around 85%. They are all trying to maximize cumulative free cash flow over the next five years while preventing another catastrophic downcycle.
Meta should just pay us to use their product.
Sung Kim@sungkim.bsky.social · 9h agoMeta's Muse Spark 1.2 I am not sure why he included GPT 5.6 terra, not GPT 5.6 Sol.
Meta's Muse Spark 1.2 I am not sure why he included GPT 5.6 terra, not GPT 5.6 Sol.
Jeff Dean, Google’s former chief scientist, is co-founding a new company called Discovery Loop. Its initial goal is to automate machine-learning research and engineering, essentially applying a bounded form of recursive self-improvement to AI development.
Discovery Loop — Continuous Exploration
Discovery Loop is building AI systems that automate the experimental loops of science and engineering. Continuous Exploration.
discoveryloop.com
Interesting... This happened faster than I expected. He may not be the best person to lead Google DeepMind as it competes with both Anthropic and OpenAI.
Axios@axios.com · 14h agoNEW: Google DeepMind CEO Demis Hassabis is stepping down
Why do I support the DSA as a fiscally conservative Democrat? Because AIPAC and its allies are bizarrely incentivized to drag us into another trillion-dollar Middle East quagmire every few years, all so Israel can “mowing the grass.” Compared with that, the DSA’s proposals are a rounding error.
Who knew it would be this hard to beat AIPAC? Anyway, congratulations on the win.
Abdul El-Sayed@abdulelsayed.bsky.social · 18h agoMoney out of politics. Money in your pocket. Medicare for All. Michigan... We won. I am so proud to be your Democratic nominee for US Senate. Let's win in November.
Hold-off on buying AMD stocks. AMD issued OpenAI and Meta warrants equivalent to nearly 20% of its current shares, before the EPYC server boom became obvious.
Google may acqui-hire Mechanize They build environments and evals for frontier coding agents. In these environments, models carry out software engineering work such as building a feature, deploying an application, or debugging an issue in an unfamiliar codebase.
Please help me understand why AMD, who is paying their customers to buy their AI hardware (not financing, but literally paying them) has a higher market value than Intel?
Apparently, Airtable executives spun off a profitable part of the business, called "Hyperagent", then dumped the rest to Bending Spoons.
Sung Kim@sungkim.bsky.social · yesterdayAirtable selling itself to Bending Spoons basically means: F*** the employees. Extract more from the customers. Keep just enough of the product working. We’re cashing out.
Airtable selling itself to Bending Spoons basically means: F*** the employees. Extract more from the customers. Keep just enough of the product working. We’re cashing out.
Someone should remind people that August is a vacation month. Anyway, Black Forest Labs releases FLUX 3 Video (open-weight soon) Web: bfl.ai/models/flux-3 Docs: docs.bfl.ai/flux_3/flux3...
What’s it like to have AI as your boss? "AIs are kind to their employees, to the extent that it sometimes affects their business in a bad way. They received 26 requests for time off and approved every single one, even those that came in with short notice."
Gavin Baker: "SSI says that they're going to come out with [you know] their model in August."
Mistral AI's Shieldstral (open-weight) A 3B open-weights, policy-adaptive multimodal safety classifier that matches models up to 7x its size on text safety and sets a new state of the art on multimodal moderation. mistral.ai/news/shields...
This is interesting. Citadel reviews academic finance papers and tests whether their findings persist out of sample because it has found that doing so is worthwhile. Of course, they used the agentic AI to automate the process.
Moondream's Photon 2.0: Inference engine for Physical AI Runs Moondream, Qwen 3.5, and Gemma 4 on NVIDIA H100. Photon outperformed vLLM and SGLang in every throughput test and brought every model online faster, via use of a single 'megakernel' that runs the entire inference on the GPU alone.
Cursor is open-sourcing Mixture-of-Kittens (MoK), their MoE training megakernel for NVL72s. It fuses all Mixture-of-Experts communication and computation into a single, fully deterministic kernel, and runs up to 2.37x faster than the strongest public baselines.
With Korea’s stock market swinging up or down by 10% almost daily, you’d think Koreans would be talking about the market. But no. 1. Weather — it’s miserably hot. 2. Property tax — the government has implemented a progressive property tax system.
Alibaba Ant's Ling-3.0-flash (open-weight) Official BF16 and FP8-quantized versions are now available. 🤗 Hugging Face Official: huggingface.co/inclusionAI/... FP8: huggingface.co/inclusionAI/...
Alibaba's Zvec's latest release introduces a native DiskANN index that stores raw vectors and the graph structure on SSD while keeping only a PQ-compressed "map" in memory. zvec.org/en/blog/2026...
DiskANN: Scaling Vector Databases to Hundreds of Millions of Vectors
Zvec introduces DiskANN in v0.6.0, reducing the memory cost of vectors by an order of magnitude by storing the raw vectors and graph structure on SSD and keeping only a PQ-compressed map in memory.
zvec.org
Google DeepMind's DiffusionGemma Technical Report They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
Two uncensored MiniMax-H3 - MINIMAX-H3 — NSFW: huggingface.co/Blackfrost-R... - Qwen3-VL-32B Heretic (MiniMax-H3 text encoder) — NVFP4: huggingface.co/sakamakismil...
Blackfrost-Research/MINIMAX-H3-NSFW · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
Liquid AI's LFM2.5-2.6B, an agentic model that runs entirely on-device. > Pre-trained on ~34T tokens > LFM2.5 flagship hybrid architecture > Context length: 128K > Vocab size: 128K > balanced intelligence per watt > customizable on a single GPU for any specialized task > LFM2 open-weight license
Williams-sonoma’s Taste of Korea When you make Chinese dinnerware and cookware, but decide to pivot to capitalize on the popularity of all things Korean. www.williams-sonoma.com/shop/new/tas...
Order Korean Sauces + Seasonings | Williams Sonoma
Explore authentic Korean sauces like ssamjang, Korean BBQ sauce, and versatile Korean BBQ dressing. Find your perfect Asian BBQ sauce for any dish. Shop now!
williams-sonoma.com
Understanding Transformers and Attention Mechanisms: An Introduction for Applied Mathematicians by Michel Fabrice Serret arxiv.org/abs/2604.00965
Looking at the news about Korea's weather. Did I mention, Korea's weather sucks.