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
Only available in China? Doubao-Seed-Code model is not one of the available model for Trae ( docs.trae.ai/ide/models?_... )
ByteDance unveils China’s most affordable AI coding agent at just US$1.30 a month
www.scmp.com/tech/big-tec...
ByteDance unveils China’s most affordable AI coding agent at just US$1.30 a month
www.scmp.com/tech/big-tec...
ByteDance unveils China’s most affordable AI coding agent at just US$1.30 a month
The TikTok owner’s cloud unit has launched a new AI coding assistant, stoking China’s AI coding price war.
www.scmp.com
November 13, 2025 at 6:48 AM
Only available in China? Doubao-Seed-Code model is not one of the available model for Trae ( docs.trae.ai/ide/models?_... )
ByteDance unveils China’s most affordable AI coding agent at just US$1.30 a month
www.scmp.com/tech/big-tec...
ByteDance unveils China’s most affordable AI coding agent at just US$1.30 a month
www.scmp.com/tech/big-tec...
I like Docker and I hate Docker, but let’s be honest, nobody actually loves Docker.
November 13, 2025 at 6:35 AM
I like Docker and I hate Docker, but let’s be honest, nobody actually loves Docker.
It looks like Qualcomm may be Intel Foundry Services’ newest customer. There’s no press release or formal confirmation, but multiple job openings reference project responsibilities that align closely with Intel's advanced packaging - EMIB.
November 13, 2025 at 6:32 AM
It looks like Qualcomm may be Intel Foundry Services’ newest customer. There’s no press release or formal confirmation, but multiple job openings reference project responsibilities that align closely with Intel's advanced packaging - EMIB.
Nice! Tesla is also Intel Foundry Services' customer.
"Dojo 3 chip production is now distributed between TSMC and Samsung Electronics, with packaging operations handled at Intel's Arizona facility."
www.digitimes.com/news/a202511...
"Dojo 3 chip production is now distributed between TSMC and Samsung Electronics, with packaging operations handled at Intel's Arizona facility."
www.digitimes.com/news/a202511...
Elon Musk’s secret fab plan: new U.S. chip plant targets 2026 ramp
Elon Musk, owner of SpaceX and Tesla Inc., is making significant strides to develop a comprehensive semiconductor manufacturing supply chain in the US. Sources indicate that the fan-out panel-level pa...
www.digitimes.com
November 13, 2025 at 6:30 AM
Nice! Tesla is also Intel Foundry Services' customer.
"Dojo 3 chip production is now distributed between TSMC and Samsung Electronics, with packaging operations handled at Intel's Arizona facility."
www.digitimes.com/news/a202511...
"Dojo 3 chip production is now distributed between TSMC and Samsung Electronics, with packaging operations handled at Intel's Arizona facility."
www.digitimes.com/news/a202511...
Weibo, China's Twitter, released VibeThinker-1.5B — SOTA reasoning in a tiny model.
🚀 Performance: Highly competitive on AIME24/25 & HMMT25 — surpasses DeepSeek R1-0120 on math, and outperforms same-size models in competitive coding.
🚀 Performance: Highly competitive on AIME24/25 & HMMT25 — surpasses DeepSeek R1-0120 on math, and outperforms same-size models in competitive coding.
November 13, 2025 at 6:23 AM
Weibo, China's Twitter, released VibeThinker-1.5B — SOTA reasoning in a tiny model.
🚀 Performance: Highly competitive on AIME24/25 & HMMT25 — surpasses DeepSeek R1-0120 on math, and outperforms same-size models in competitive coding.
🚀 Performance: Highly competitive on AIME24/25 & HMMT25 — surpasses DeepSeek R1-0120 on math, and outperforms same-size models in competitive coding.
No More Train–Inference Mismatch!
vLLM demonstrates bitwise consistent on-policy RL with TorchTitan (training) + vLLM (inference) — the first open-source run where training and inference numerics match exactly.
vLLM demonstrates bitwise consistent on-policy RL with TorchTitan (training) + vLLM (inference) — the first open-source run where training and inference numerics match exactly.
November 13, 2025 at 6:21 AM
No More Train–Inference Mismatch!
vLLM demonstrates bitwise consistent on-policy RL with TorchTitan (training) + vLLM (inference) — the first open-source run where training and inference numerics match exactly.
vLLM demonstrates bitwise consistent on-policy RL with TorchTitan (training) + vLLM (inference) — the first open-source run where training and inference numerics match exactly.
LeJEPA: a novel pretraining paradigm free of the (many) heuristics we relied on (stop-grad, teacher, ...)
- 60+ arch., up to 2B params
- 10+ datasets
- in-domain training (>DINOv3)
- corr(train loss, test perf)=95%
- 60+ arch., up to 2B params
- 10+ datasets
- in-domain training (>DINOv3)
- corr(train loss, test perf)=95%
November 13, 2025 at 6:19 AM
LeJEPA: a novel pretraining paradigm free of the (many) heuristics we relied on (stop-grad, teacher, ...)
- 60+ arch., up to 2B params
- 10+ datasets
- in-domain training (>DINOv3)
- corr(train loss, test perf)=95%
- 60+ arch., up to 2B params
- 10+ datasets
- in-domain training (>DINOv3)
- corr(train loss, test perf)=95%
Baidu, China's Google, releases ERNIE 5.0 — their latest natively omni-modal foundational model.
ernie.baidu.com
ernie.baidu.com
November 13, 2025 at 6:16 AM
Baidu, China's Google, releases ERNIE 5.0 — their latest natively omni-modal foundational model.
ernie.baidu.com
ernie.baidu.com
November 13, 2025 at 6:15 AM
I like this narrative by Baidu founder Robin Li
Currently, we have an unhealthy ‘upright pyramid’ AI industry structure
- Application Layer
- Model Layer
- Chip Layer
They are shifting to a healthy AI industry structure, which is an ‘inverted pyramid'
- Application Layer
- Model Layer
- Chip Layer
Currently, we have an unhealthy ‘upright pyramid’ AI industry structure
- Application Layer
- Model Layer
- Chip Layer
They are shifting to a healthy AI industry structure, which is an ‘inverted pyramid'
- Application Layer
- Model Layer
- Chip Layer
November 13, 2025 at 3:09 AM
I like this narrative by Baidu founder Robin Li
Currently, we have an unhealthy ‘upright pyramid’ AI industry structure
- Application Layer
- Model Layer
- Chip Layer
They are shifting to a healthy AI industry structure, which is an ‘inverted pyramid'
- Application Layer
- Model Layer
- Chip Layer
Currently, we have an unhealthy ‘upright pyramid’ AI industry structure
- Application Layer
- Model Layer
- Chip Layer
They are shifting to a healthy AI industry structure, which is an ‘inverted pyramid'
- Application Layer
- Model Layer
- Chip Layer
If Microsoft has access to all of OpenAI’s IP, why are their AI models so lackluster?
November 13, 2025 at 2:28 AM
If Microsoft has access to all of OpenAI’s IP, why are their AI models so lackluster?
Went to a fancy buffet alone and was in and out in 30 minutes. Pretty sure I didn’t get my money’s worth.
November 12, 2025 at 6:00 AM
Went to a fancy buffet alone and was in and out in 30 minutes. Pretty sure I didn’t get my money’s worth.
Are prompting and activation steering just two sides of the same coin?
The paper formalizes a Bayesian framework for model control: altering a model's "beliefs" over which persona or data source it's emulating. Context (prompting) and internal representations (steering)
The paper formalizes a Bayesian framework for model control: altering a model's "beliefs" over which persona or data source it's emulating. Context (prompting) and internal representations (steering)
November 12, 2025 at 5:42 AM
Are prompting and activation steering just two sides of the same coin?
The paper formalizes a Bayesian framework for model control: altering a model's "beliefs" over which persona or data source it's emulating. Context (prompting) and internal representations (steering)
The paper formalizes a Bayesian framework for model control: altering a model's "beliefs" over which persona or data source it's emulating. Context (prompting) and internal representations (steering)
How character.ai trained their proprietary model Kaiju (13B, 34B, 110B), before switching to OSS model.
Here are a few optimizations that they did
- MuP-like scaling
- MQA + SWA
- Clamping everywhere to control activation
- KV Cache sharing
Here are a few optimizations that they did
- MuP-like scaling
- MQA + SWA
- Clamping everywhere to control activation
- KV Cache sharing
November 12, 2025 at 5:39 AM
How character.ai trained their proprietary model Kaiju (13B, 34B, 110B), before switching to OSS model.
Here are a few optimizations that they did
- MuP-like scaling
- MQA + SWA
- Clamping everywhere to control activation
- KV Cache sharing
Here are a few optimizations that they did
- MuP-like scaling
- MQA + SWA
- Clamping everywhere to control activation
- KV Cache sharing
MUVERA
Multi-Vector Retrieval via Fixed Dimensional Encodings is an interesting approach by Google Research. It transforms multi-vector representations into single fixed-size vectors (fixed dimensional encodings).
Multi-Vector Retrieval via Fixed Dimensional Encodings is an interesting approach by Google Research. It transforms multi-vector representations into single fixed-size vectors (fixed dimensional encodings).
November 12, 2025 at 5:36 AM
MUVERA
Multi-Vector Retrieval via Fixed Dimensional Encodings is an interesting approach by Google Research. It transforms multi-vector representations into single fixed-size vectors (fixed dimensional encodings).
Multi-Vector Retrieval via Fixed Dimensional Encodings is an interesting approach by Google Research. It transforms multi-vector representations into single fixed-size vectors (fixed dimensional encodings).
AELLA is an open-science initiative to make scientific research accessible via structured summaries created by LLMs
Available now:
- Dataset of 100K summaries
- 2 fine-tuned LLMs
- 3d visualizer
Available now:
- Dataset of 100K summaries
- 2 fine-tuned LLMs
- 3d visualizer
November 12, 2025 at 5:35 AM
AELLA is an open-science initiative to make scientific research accessible via structured summaries created by LLMs
Available now:
- Dataset of 100K summaries
- 2 fine-tuned LLMs
- 3d visualizer
Available now:
- Dataset of 100K summaries
- 2 fine-tuned LLMs
- 3d visualizer
HipKittens: fast AMD kernels and a collection of opinionated programming primitives to make AMD kernel dev easier! by @simranarora.bsky.social
November 12, 2025 at 5:32 AM
HipKittens: fast AMD kernels and a collection of opinionated programming primitives to make AMD kernel dev easier! by @simranarora.bsky.social
LOL. I hope it does not work out for him, except Palantir.
November 12, 2025 at 3:40 AM
LOL. I hope it does not work out for him, except Palantir.
Why does this test remind me of Idiocracy? Especially question number 2 - 21% of test takers got it wrong:
“If you have one bucket that holds 2 gallons and another bucket that holds 5 gallons, how many buckets do you have?”
The red indicates the percentage of people who got it right. See page 49.
“If you have one bucket that holds 2 gallons and another bucket that holds 5 gallons, how many buckets do you have?”
The red indicates the percentage of people who got it right. See page 49.
November 12, 2025 at 1:41 AM
Why does this test remind me of Idiocracy? Especially question number 2 - 21% of test takers got it wrong:
“If you have one bucket that holds 2 gallons and another bucket that holds 5 gallons, how many buckets do you have?”
The red indicates the percentage of people who got it right. See page 49.
“If you have one bucket that holds 2 gallons and another bucket that holds 5 gallons, how many buckets do you have?”
The red indicates the percentage of people who got it right. See page 49.
I understand that UCSD stated that one out of eight incoming freshmen needs remedial math in 2025, compared to one out of a hundred in 2019; but why are five of them math majors!
Read more about it here: senate.ucsd.edu/media/740347...
Read more about it here: senate.ucsd.edu/media/740347...
November 12, 2025 at 1:06 AM
I understand that UCSD stated that one out of eight incoming freshmen needs remedial math in 2025, compared to one out of a hundred in 2019; but why are five of them math majors!
Read more about it here: senate.ucsd.edu/media/740347...
Read more about it here: senate.ucsd.edu/media/740347...
vLLM is now optimized for Arc Pro B-Series GPUs
You can buy 8 of these GPUs and cluster them for 192GB of VRAM for under $5,000.
blog.vllm.ai/2025/11/11/i...
You can buy 8 of these GPUs and cluster them for 192GB of VRAM for under $5,000.
blog.vllm.ai/2025/11/11/i...
Fast and Affordable LLMs serving on Intel Arc Pro B-Series GPUs with vLLM
Intel® Arc™ Pro B-Series GPU Family GPUs deliver powerful AI capabilities with a focus on accessibility and exceptional price-to-performance ratios. Their large memory capacity and scalability with mu...
blog.vllm.ai
November 11, 2025 at 11:21 PM
vLLM is now optimized for Arc Pro B-Series GPUs
You can buy 8 of these GPUs and cluster them for 192GB of VRAM for under $5,000.
blog.vllm.ai/2025/11/11/i...
You can buy 8 of these GPUs and cluster them for 192GB of VRAM for under $5,000.
blog.vllm.ai/2025/11/11/i...
ByteDance released the Doubao Seed Code (closed-weight).
It is only available in China and it is supposed to be pretty good.
In Chinese: exp.volcengine.com/ark?model=do...
It is only available in China and it is supposed to be pretty good.
In Chinese: exp.volcengine.com/ark?model=do...
November 11, 2025 at 11:18 PM
ByteDance released the Doubao Seed Code (closed-weight).
It is only available in China and it is supposed to be pretty good.
In Chinese: exp.volcengine.com/ark?model=do...
It is only available in China and it is supposed to be pretty good.
In Chinese: exp.volcengine.com/ark?model=do...
Inside High-Frequency Trading Systems: The Race to Zero Latency
These systems aren’t engineered for milliseconds — they’re tuned for microseconds, even nanoseconds. Every component, from the network card to the FPGA bitstream, is obsessed with one goal: shaving latency down to the bare minimum.
These systems aren’t engineered for milliseconds — they’re tuned for microseconds, even nanoseconds. Every component, from the network card to the FPGA bitstream, is obsessed with one goal: shaving latency down to the bare minimum.
Inside High-Frequency Trading Systems: The Race to Zero Latency
When people talk about high-frequency trading (HFT), they often imagine black-box algorithms making millions in milliseconds. The reality…
levelup.gitconnected.com
November 11, 2025 at 10:42 PM
Inside High-Frequency Trading Systems: The Race to Zero Latency
These systems aren’t engineered for milliseconds — they’re tuned for microseconds, even nanoseconds. Every component, from the network card to the FPGA bitstream, is obsessed with one goal: shaving latency down to the bare minimum.
These systems aren’t engineered for milliseconds — they’re tuned for microseconds, even nanoseconds. Every component, from the network card to the FPGA bitstream, is obsessed with one goal: shaving latency down to the bare minimum.
Baidu released ERNIE-4.5-VL-28B-A3B-Thinking, a lightweight multimodal reasoning model.
- 3B active parameters with enhanced semantic alignment between visual and language modalities
- "Thinking with Images" feature that enables zooming in and out to capture finer details
- Apache License 2.0
- 3B active parameters with enhanced semantic alignment between visual and language modalities
- "Thinking with Images" feature that enables zooming in and out to capture finer details
- Apache License 2.0
November 11, 2025 at 12:56 PM
Baidu released ERNIE-4.5-VL-28B-A3B-Thinking, a lightweight multimodal reasoning model.
- 3B active parameters with enhanced semantic alignment between visual and language modalities
- "Thinking with Images" feature that enables zooming in and out to capture finer details
- Apache License 2.0
- 3B active parameters with enhanced semantic alignment between visual and language modalities
- "Thinking with Images" feature that enables zooming in and out to capture finer details
- Apache License 2.0
If I had to report to this clown, I’d leave too. What’s with the trucker hat and that loud t-shirt?
Meta’s chief artificial intelligence scientist, Yann LeCun, has reportedly told associates he plans to leave the Silicon Valley company in the coming months.
www.ft.com/content/c586...
Meta’s chief artificial intelligence scientist, Yann LeCun, has reportedly told associates he plans to leave the Silicon Valley company in the coming months.
www.ft.com/content/c586...
November 11, 2025 at 12:52 PM
If I had to report to this clown, I’d leave too. What’s with the trucker hat and that loud t-shirt?
Meta’s chief artificial intelligence scientist, Yann LeCun, has reportedly told associates he plans to leave the Silicon Valley company in the coming months.
www.ft.com/content/c586...
Meta’s chief artificial intelligence scientist, Yann LeCun, has reportedly told associates he plans to leave the Silicon Valley company in the coming months.
www.ft.com/content/c586...