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
You can also find me at threads: @sung.kim.mw
Reposted by Sung Kim
Holy f… I thought _I_ had a Skills obsession. See his claude_scientific_skills: github.com/K-Dense-AI/c... !!
November 19, 2025 at 12:10 PM
Holy f… I thought _I_ had a Skills obsession. See his claude_scientific_skills: github.com/K-Dense-AI/c... !!
Distribution Aware Speculative decoding framework that accelerates RL rollouts without altering model outputs. DAS integrates two key ideas: an adaptive, nonparametric drafter built from recent rollouts using an incrementally maintained suffix tree,
November 19, 2025 at 8:52 AM
Distribution Aware Speculative decoding framework that accelerates RL rollouts without altering model outputs. DAS integrates two key ideas: an adaptive, nonparametric drafter built from recent rollouts using an incrementally maintained suffix tree,
I’m not sure if it’s worth looking at, but he did name his project in a way that’s guaranteed to go viral.
Karpathy
An agentic ML Engineer that trains state-of-the-art ML models using Claude Code SDK and Google ADK. This is a very simple implemenation demoing the power of Claude skills for ML.
Karpathy
An agentic ML Engineer that trains state-of-the-art ML models using Claude Code SDK and Google ADK. This is a very simple implemenation demoing the power of Claude skills for ML.
GitHub - K-Dense-AI/karpathy: An agentic Machine Learning Engineer
An agentic Machine Learning Engineer. Contribute to K-Dense-AI/karpathy development by creating an account on GitHub.
github.com
November 19, 2025 at 6:10 AM
I’m not sure if it’s worth looking at, but he did name his project in a way that’s guaranteed to go viral.
Karpathy
An agentic ML Engineer that trains state-of-the-art ML models using Claude Code SDK and Google ADK. This is a very simple implemenation demoing the power of Claude skills for ML.
Karpathy
An agentic ML Engineer that trains state-of-the-art ML models using Claude Code SDK and Google ADK. This is a very simple implemenation demoing the power of Claude skills for ML.
Stanford's Hazy Research released the ParallelKittens
An update to ThunderKittens that lets you easily write fast computation-communication overlapped multi-GPU kernels, along with new kernels for data, tensor, sequence, and expert parallelism!
An update to ThunderKittens that lets you easily write fast computation-communication overlapped multi-GPU kernels, along with new kernels for data, tensor, sequence, and expert parallelism!
November 19, 2025 at 6:06 AM
Stanford's Hazy Research released the ParallelKittens
An update to ThunderKittens that lets you easily write fast computation-communication overlapped multi-GPU kernels, along with new kernels for data, tensor, sequence, and expert parallelism!
An update to ThunderKittens that lets you easily write fast computation-communication overlapped multi-GPU kernels, along with new kernels for data, tensor, sequence, and expert parallelism!
Reposted by Sung Kim
Today we’re releasing Deep Research Tulu (DR Tulu)—the first fully open, end-to-end recipe for long-form deep research, plus an 8B agent you can use right away. Train agents that plan, search, synthesize, & cite across sources, making expert research more accessible. 🧭📚
November 18, 2025 at 3:31 PM
Today we’re releasing Deep Research Tulu (DR Tulu)—the first fully open, end-to-end recipe for long-form deep research, plus an 8B agent you can use right away. Train agents that plan, search, synthesize, & cite across sources, making expert research more accessible. 🧭📚
If you’re feeling down because you made a mistake, take a moment to listen to the sage advice of Masood Boomgaard.
www.youtube.com/@selfhelpsin...
www.youtube.com/@selfhelpsin...
November 19, 2025 at 5:55 AM
If you’re feeling down because you made a mistake, take a moment to listen to the sage advice of Masood Boomgaard.
www.youtube.com/@selfhelpsin...
www.youtube.com/@selfhelpsin...
Alibaba released AgentEvolver , an open-source, self-evolving agent system.
AgentEvolver integrates three synergistic mechanisms: Self-Questioning, Self-Navigating, and Self-Attributing; to systematically address critical bottlenecks in Agent RL training, including task scarcity,
AgentEvolver integrates three synergistic mechanisms: Self-Questioning, Self-Navigating, and Self-Attributing; to systematically address critical bottlenecks in Agent RL training, including task scarcity,
November 19, 2025 at 5:46 AM
Alibaba released AgentEvolver , an open-source, self-evolving agent system.
AgentEvolver integrates three synergistic mechanisms: Self-Questioning, Self-Navigating, and Self-Attributing; to systematically address critical bottlenecks in Agent RL training, including task scarcity,
AgentEvolver integrates three synergistic mechanisms: Self-Questioning, Self-Navigating, and Self-Attributing; to systematically address critical bottlenecks in Agent RL training, including task scarcity,
Just image Transformers (JiT)
JiTs are simple large-patch Transformers that operate on raw pixels, no tokenizer, pre-training, or extra losses needed. JiT excels in high-dimensional spaces where traditional noise-predicting models can fail.
JiTs are simple large-patch Transformers that operate on raw pixels, no tokenizer, pre-training, or extra losses needed. JiT excels in high-dimensional spaces where traditional noise-predicting models can fail.
November 19, 2025 at 5:42 AM
Just image Transformers (JiT)
JiTs are simple large-patch Transformers that operate on raw pixels, no tokenizer, pre-training, or extra losses needed. JiT excels in high-dimensional spaces where traditional noise-predicting models can fail.
JiTs are simple large-patch Transformers that operate on raw pixels, no tokenizer, pre-training, or extra losses needed. JiT excels in high-dimensional spaces where traditional noise-predicting models can fail.
When I visit Korea, I love eating these Korean dishes — they’re incredibly tasty, healthy, and dirt-cheap, even by Korean standards, less than USD $5:
- Siraegi Soup (시락국)
- Beansprout Soup (콩나물국밥)
- Noodles in Anchovy Broth (멸치국수)
- Siraegi Soup (시락국)
- Beansprout Soup (콩나물국밥)
- Noodles in Anchovy Broth (멸치국수)
November 19, 2025 at 3:54 AM
When I visit Korea, I love eating these Korean dishes — they’re incredibly tasty, healthy, and dirt-cheap, even by Korean standards, less than USD $5:
- Siraegi Soup (시락국)
- Beansprout Soup (콩나물국밥)
- Noodles in Anchovy Broth (멸치국수)
- Siraegi Soup (시락국)
- Beansprout Soup (콩나물국밥)
- Noodles in Anchovy Broth (멸치국수)
Cloudflare’s bot-detection ML system expected a limited number of features, but a database permissions change caused the feature file to suddenly contain far more features than allowed. This overflow caused the bot-detection module to crash, which in turn disrupted Cloudflare’s core traffic.
On November 18 Cloudflare experienced a service outage. Here's a detailed breakdown of what happened. https://cfl.re/43Bw8AI
Cloudflare outage on November 18, 2025
Cloudflare suffered a service outage on November 18, 2025. The outage was triggered by a bug in generation logic for a Bot Management feature file causing many Cloudflare services to be affected.
blog.cloudflare.com
November 19, 2025 at 3:02 AM
Cloudflare’s bot-detection ML system expected a limited number of features, but a database permissions change caused the feature file to suddenly contain far more features than allowed. This overflow caused the bot-detection module to crash, which in turn disrupted Cloudflare’s core traffic.
Reposted by Sung Kim
Thinking Machines Lab, led by Mira Murati, hires Meta AI veteran Soumith Chintala as it rapidly expands its team and prepares for major funding.
A top Meta AI researcher has joined Mira Murati's hot new startup, Thinking Machines Lab
Thinking Machines Lab, led by Mira Murati, hires Meta AI veteran Soumith Chintala as it rapidly expands its team and prepares for major funding.
www.businessinsider.com
November 18, 2025 at 11:02 PM
Thinking Machines Lab, led by Mira Murati, hires Meta AI veteran Soumith Chintala as it rapidly expands its team and prepares for major funding.
Google Antigravity was developed by the former Windsurf development team, so it’s not surprising that they reused the Windsurf codebase.
"Cascade" is WindSurf's old agent.
"Cascade" is WindSurf's old agent.
November 19, 2025 at 12:16 AM
Google Antigravity was developed by the former Windsurf development team, so it’s not surprising that they reused the Windsurf codebase.
"Cascade" is WindSurf's old agent.
"Cascade" is WindSurf's old agent.
How many AI IDE does Google have? Did I miss any?
- Firebase Studio: firebase.studio
- Google Antigravity: antigravity.google
- Google AI Studio: aistudio.google.com
- Gemini Code Assist: codeassist.google
- Google Jules: jules.google
- Gemini CLI: geminicli.com
- Firebase Studio: firebase.studio
- Google Antigravity: antigravity.google
- Google AI Studio: aistudio.google.com
- Gemini Code Assist: codeassist.google
- Google Jules: jules.google
- Gemini CLI: geminicli.com
Google AI Studio
The fastest path from prompt to production with Gemini
aistudio.google.com
November 18, 2025 at 11:53 PM
How many AI IDE does Google have? Did I miss any?
- Firebase Studio: firebase.studio
- Google Antigravity: antigravity.google
- Google AI Studio: aistudio.google.com
- Gemini Code Assist: codeassist.google
- Google Jules: jules.google
- Gemini CLI: geminicli.com
- Firebase Studio: firebase.studio
- Google Antigravity: antigravity.google
- Google AI Studio: aistudio.google.com
- Gemini Code Assist: codeassist.google
- Google Jules: jules.google
- Gemini CLI: geminicli.com
So… Wei-Ren Luo has retired from TSMC at age 75 and may now re-join Intel as an Executive Vice President (EVP) overseeing the wafer manufacturing group and R&D.
He played a key role in TSMC’s large-scale EUV adoption after spending roughly 20 years at Intel earlier in his career.
He played a key role in TSMC’s large-scale EUV adoption after spending roughly 20 years at Intel earlier in his career.
November 18, 2025 at 1:39 AM
So… Wei-Ren Luo has retired from TSMC at age 75 and may now re-join Intel as an Executive Vice President (EVP) overseeing the wafer manufacturing group and R&D.
He played a key role in TSMC’s large-scale EUV adoption after spending roughly 20 years at Intel earlier in his career.
He played a key role in TSMC’s large-scale EUV adoption after spending roughly 20 years at Intel earlier in his career.
FlashMoBA
FlashMoBA is a memory-efficient sparse attention mechanism designed to accelerate the training and inference of long-sequence models where it achieves up to 14.7x speedup over FlashAttention-2 for small blocks.
FlashMoBA is a memory-efficient sparse attention mechanism designed to accelerate the training and inference of long-sequence models where it achieves up to 14.7x speedup over FlashAttention-2 for small blocks.
November 17, 2025 at 6:55 AM
FlashMoBA
FlashMoBA is a memory-efficient sparse attention mechanism designed to accelerate the training and inference of long-sequence models where it achieves up to 14.7x speedup over FlashAttention-2 for small blocks.
FlashMoBA is a memory-efficient sparse attention mechanism designed to accelerate the training and inference of long-sequence models where it achieves up to 14.7x speedup over FlashAttention-2 for small blocks.
ByteDance Seed's Virtual Width Networks (VWN)
(Comment — based on the number of authors on this paper, it must be a big deal.)
All the benefits of wider representations without incurring the quadratic cost of increasing the hidden size.
(Comment — based on the number of authors on this paper, it must be a big deal.)
All the benefits of wider representations without incurring the quadratic cost of increasing the hidden size.
November 17, 2025 at 6:49 AM
ByteDance Seed's Virtual Width Networks (VWN)
(Comment — based on the number of authors on this paper, it must be a big deal.)
All the benefits of wider representations without incurring the quadratic cost of increasing the hidden size.
(Comment — based on the number of authors on this paper, it must be a big deal.)
All the benefits of wider representations without incurring the quadratic cost of increasing the hidden size.
I tried Gemini 3 in my dream last night, and I’m convinced AGI has arrived.
November 17, 2025 at 1:54 AM
I tried Gemini 3 in my dream last night, and I’m convinced AGI has arrived.
If cryptos continue to decline, we’ll soon find out whether they’ve become a systemic risk.
November 16, 2025 at 11:53 PM
If cryptos continue to decline, we’ll soon find out whether they’ve become a systemic risk.
Reposted by Sung Kim
Reading group tomorrow: "How to build a consistency model: Learning flow maps via self-distillation" with Nicholas Boffi! arxiv.org/abs/2505.18825
Join us on zoom at 9am PT, 12pm ET, 6pm CET: portal.valencelabs.com/starklyspeak...
Join us on zoom at 9am PT, 12pm ET, 6pm CET: portal.valencelabs.com/starklyspeak...
November 16, 2025 at 5:06 PM
Reading group tomorrow: "How to build a consistency model: Learning flow maps via self-distillation" with Nicholas Boffi! arxiv.org/abs/2505.18825
Join us on zoom at 9am PT, 12pm ET, 6pm CET: portal.valencelabs.com/starklyspeak...
Join us on zoom at 9am PT, 12pm ET, 6pm CET: portal.valencelabs.com/starklyspeak...
gtr - Git Worktree Runner
"Normally, you can only work on one git branch at a time in a folder. Want to fix a bug while working on a feature? You have to stash changes, switch branches, then switch back. Git worktrees let you have multiple branches checked out at once in different folders
"Normally, you can only work on one git branch at a time in a folder. Want to fix a bug while working on a feature? You have to stash changes, switch branches, then switch back. Git worktrees let you have multiple branches checked out at once in different folders
November 16, 2025 at 11:24 PM
gtr - Git Worktree Runner
"Normally, you can only work on one git branch at a time in a folder. Want to fix a bug while working on a feature? You have to stash changes, switch branches, then switch back. Git worktrees let you have multiple branches checked out at once in different folders
"Normally, you can only work on one git branch at a time in a folder. Want to fix a bug while working on a feature? You have to stash changes, switch branches, then switch back. Git worktrees let you have multiple branches checked out at once in different folders
Hooray for an introvert!
November 16, 2025 at 10:42 PM
Hooray for an introvert!
Apple may also be Intel Foundry Services’ newest customer.
"Solve complex electrical and mechanical integration challenges of memory packages incorporated into advanced packaging technologies such as CoWoS, EMIB, SoIC, and PoP."
jobs.apple.com/en-il/detail...
"Solve complex electrical and mechanical integration challenges of memory packages incorporated into advanced packaging technologies such as CoWoS, EMIB, SoIC, and PoP."
jobs.apple.com/en-il/detail...
November 16, 2025 at 10:34 PM
Apple may also be Intel Foundry Services’ newest customer.
"Solve complex electrical and mechanical integration challenges of memory packages incorporated into advanced packaging technologies such as CoWoS, EMIB, SoIC, and PoP."
jobs.apple.com/en-il/detail...
"Solve complex electrical and mechanical integration challenges of memory packages incorporated into advanced packaging technologies such as CoWoS, EMIB, SoIC, and PoP."
jobs.apple.com/en-il/detail...
Anthropic, a company that is actively campaigning for regulatory capture, reported that it disrupted the first known AI-orchestrated cyber-espionage campaign ( www.anthropic.com/news/disrupt... ).
November 16, 2025 at 2:14 AM
Anthropic, a company that is actively campaigning for regulatory capture, reported that it disrupted the first known AI-orchestrated cyber-espionage campaign ( www.anthropic.com/news/disrupt... ).
I really like the new Steam Machine’s form factor. I’m looking forward to buying one of the many expected variants soon.
November 15, 2025 at 1:39 AM
I really like the new Steam Machine’s form factor. I’m looking forward to buying one of the many expected variants soon.