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
I saw a video stating that a nice 2 bedroom apartment in HCMC (Saigon), Vietnam is $2,000 per month.
Is that true?
Is that true?
February 6, 2026 at 4:22 AM
I saw a video stating that a nice 2 bedroom apartment in HCMC (Saigon), Vietnam is $2,000 per month.
Is that true?
Is that true?
Pydantic's Monty
A minimal, secure Python interpreter written in Rust for use by AI.
github.com/pydantic/monty
A minimal, secure Python interpreter written in Rust for use by AI.
github.com/pydantic/monty
GitHub - pydantic/monty: A minimal, secure Python interpreter written in Rust for use by AI
A minimal, secure Python interpreter written in Rust for use by AI - pydantic/monty
github.com
February 6, 2026 at 3:52 AM
Pydantic's Monty
A minimal, secure Python interpreter written in Rust for use by AI.
github.com/pydantic/monty
A minimal, secure Python interpreter written in Rust for use by AI.
github.com/pydantic/monty
Microsoft's LiteBox
LiteBox is a sandboxing library OS. Example use cases include:
- Running unmodified Linux programs on Windows
- Sandboxing Linux applications on Linux
github.com/microsoft/li...
LiteBox is a sandboxing library OS. Example use cases include:
- Running unmodified Linux programs on Windows
- Sandboxing Linux applications on Linux
github.com/microsoft/li...
February 6, 2026 at 3:50 AM
Microsoft's LiteBox
LiteBox is a sandboxing library OS. Example use cases include:
- Running unmodified Linux programs on Windows
- Sandboxing Linux applications on Linux
github.com/microsoft/li...
LiteBox is a sandboxing library OS. Example use cases include:
- Running unmodified Linux programs on Windows
- Sandboxing Linux applications on Linux
github.com/microsoft/li...
I want to live in Tokyo for a few months.
Somehow, decades of watching anime have done absolutely nothing for my Japanese language skills.
Somehow, decades of watching anime have done absolutely nothing for my Japanese language skills.
February 6, 2026 at 3:23 AM
I want to live in Tokyo for a few months.
Somehow, decades of watching anime have done absolutely nothing for my Japanese language skills.
Somehow, decades of watching anime have done absolutely nothing for my Japanese language skills.
Kind of genius move from Microsoft GitHub Copilot CLI folks.
Instead of "claude --dangerously-skip-permissions", you can do same thing in Copilot CLI by "copilot --yolo".
Instead of "claude --dangerously-skip-permissions", you can do same thing in Copilot CLI by "copilot --yolo".
February 6, 2026 at 2:53 AM
Kind of genius move from Microsoft GitHub Copilot CLI folks.
Instead of "claude --dangerously-skip-permissions", you can do same thing in Copilot CLI by "copilot --yolo".
Instead of "claude --dangerously-skip-permissions", you can do same thing in Copilot CLI by "copilot --yolo".
Maximum Likelihood Reinforcement Learning
They observe that reinforcement learning does not maximize this likelihood, and instead optimizes only a lower-order approximation. Inspired by this observation, they introduce Maximum Likelihood Reinforcement Learning (MaxRL),
They observe that reinforcement learning does not maximize this likelihood, and instead optimizes only a lower-order approximation. Inspired by this observation, they introduce Maximum Likelihood Reinforcement Learning (MaxRL),
February 6, 2026 at 2:46 AM
Maximum Likelihood Reinforcement Learning
They observe that reinforcement learning does not maximize this likelihood, and instead optimizes only a lower-order approximation. Inspired by this observation, they introduce Maximum Likelihood Reinforcement Learning (MaxRL),
They observe that reinforcement learning does not maximize this likelihood, and instead optimizes only a lower-order approximation. Inspired by this observation, they introduce Maximum Likelihood Reinforcement Learning (MaxRL),
Goodfire AI's Intentionally designing the future of AI by Thomas McGrath
This essay describes how he think we can do backprop, why interp is key, the relevance to alignment, and how we should do it right.
Blog: www.goodfire.ai/blog/intenti...
This essay describes how he think we can do backprop, why interp is key, the relevance to alignment, and how we should do it right.
Blog: www.goodfire.ai/blog/intenti...
February 6, 2026 at 2:42 AM
Goodfire AI's Intentionally designing the future of AI by Thomas McGrath
This essay describes how he think we can do backprop, why interp is key, the relevance to alignment, and how we should do it right.
Blog: www.goodfire.ai/blog/intenti...
This essay describes how he think we can do backprop, why interp is key, the relevance to alignment, and how we should do it right.
Blog: www.goodfire.ai/blog/intenti...
Reposted by Sung Kim
Our OpenScholar paper is now in @nature.com 🎉
OpenScholar is an open-source model for synthesizing scientific research—with citations as accurate as human experts. 🧵
OpenScholar is an open-source model for synthesizing scientific research—with citations as accurate as human experts. 🧵
February 4, 2026 at 4:21 PM
Our OpenScholar paper is now in @nature.com 🎉
OpenScholar is an open-source model for synthesizing scientific research—with citations as accurate as human experts. 🧵
OpenScholar is an open-source model for synthesizing scientific research—with citations as accurate as human experts. 🧵
They are stating that PPO/GRPO in RL setting is fundamentally flawed in LLM regime due to the large, long-tailed vocabulary.
They propose Divergence Proximal Policy Optimization (DPPO) with the benefits, such as:
- super stable and smooth training
They propose Divergence Proximal Policy Optimization (DPPO) with the benefits, such as:
- super stable and smooth training
February 6, 2026 at 2:36 AM
They are stating that PPO/GRPO in RL setting is fundamentally flawed in LLM regime due to the large, long-tailed vocabulary.
They propose Divergence Proximal Policy Optimization (DPPO) with the benefits, such as:
- super stable and smooth training
They propose Divergence Proximal Policy Optimization (DPPO) with the benefits, such as:
- super stable and smooth training
They propose a new paradigm called Drifting Models, which evolve the pushforward distribution during training and naturally admit one-step inference. We introduce a drifting field that governs the sample movement and achieves equilibrium when the distributions match.
February 6, 2026 at 2:30 AM
They propose a new paradigm called Drifting Models, which evolve the pushforward distribution during training and naturally admit one-step inference. We introduce a drifting field that governs the sample movement and achieves equilibrium when the distributions match.
Crypto bros are hilarious.
On a serious note, Bitcoin hits highs then cycles into crypto winter. It is expected to go down to $30K or so. It should stay there for a while. Accumulate Bitcoin then for a new high.
On a serious note, Bitcoin hits highs then cycles into crypto winter. It is expected to go down to $30K or so. It should stay there for a while. Accumulate Bitcoin then for a new high.
February 6, 2026 at 12:55 AM
Crypto bros are hilarious.
On a serious note, Bitcoin hits highs then cycles into crypto winter. It is expected to go down to $30K or so. It should stay there for a while. Accumulate Bitcoin then for a new high.
On a serious note, Bitcoin hits highs then cycles into crypto winter. It is expected to go down to $30K or so. It should stay there for a while. Accumulate Bitcoin then for a new high.
Now they’re just one-upping themselves. Amazon says it will spend $200B in CapEx in 2026.
Google will be spending $180B on CapEx in 2026!
February 5, 2026 at 10:49 PM
Now they’re just one-upping themselves. Amazon says it will spend $200B in CapEx in 2026.
I couldn't stop laughing.
February 5, 2026 at 10:01 PM
I couldn't stop laughing.
You’d think a large model for a tabular data is revolutionary, except we’ve been doing it for DECADES.
techcrunch.com/2026/02/05/f...
techcrunch.com/2026/02/05/f...
Fundamental raises $255M Series A with a new take on big data analysis | TechCrunch
Fundamental has built a new foundation model to solve an old problem: how to draw insights from the huge quantities of structured data produced by enterprises.
techcrunch.com
February 5, 2026 at 9:51 PM
You’d think a large model for a tabular data is revolutionary, except we’ve been doing it for DECADES.
techcrunch.com/2026/02/05/f...
techcrunch.com/2026/02/05/f...
It’s kind of funny that both Anthropic and OpenAI released updated models on the same day - only an hour apart.
You’d think they know exactly what the other is doing at any given time.
You’d think they know exactly what the other is doing at any given time.
February 5, 2026 at 9:17 PM
It’s kind of funny that both Anthropic and OpenAI released updated models on the same day - only an hour apart.
You’d think they know exactly what the other is doing at any given time.
You’d think they know exactly what the other is doing at any given time.
These high-performance supercars are painful to drive when you’re going slow, too.
February 5, 2026 at 2:35 PM
These high-performance supercars are painful to drive when you’re going slow, too.
OpenAI is expanding their consulting group.
"We pair OpenAI Forward Deployed Engineers (FDEs) with your teams, working side by side to help you develop the best practices to build and run agents in production."
openai.com/index/introd...
"We pair OpenAI Forward Deployed Engineers (FDEs) with your teams, working side by side to help you develop the best practices to build and run agents in production."
openai.com/index/introd...
Introducing OpenAI Frontier
OpenAI Frontier is an enterprise platform for building, deploying, and managing AI agents with shared context, onboarding, permissions, and governance.
openai.com
February 5, 2026 at 2:32 PM
OpenAI is expanding their consulting group.
"We pair OpenAI Forward Deployed Engineers (FDEs) with your teams, working side by side to help you develop the best practices to build and run agents in production."
openai.com/index/introd...
"We pair OpenAI Forward Deployed Engineers (FDEs) with your teams, working side by side to help you develop the best practices to build and run agents in production."
openai.com/index/introd...
I guess it’s funny the first time, but it gets lame pretty quickly.
February 5, 2026 at 2:16 PM
I guess it’s funny the first time, but it gets lame pretty quickly.
Interesting! Voyage AI has release an open-weight embedding model.
voyage-4-nano is ideal for local development and prototyping while providing an easy path to production.
- Shared Embedding Space: voyage-4-nano shares an embedding space as its larger siblings
voyage-4-nano is ideal for local development and prototyping while providing an easy path to production.
- Shared Embedding Space: voyage-4-nano shares an embedding space as its larger siblings
February 5, 2026 at 1:44 PM
Interesting! Voyage AI has release an open-weight embedding model.
voyage-4-nano is ideal for local development and prototyping while providing an easy path to production.
- Shared Embedding Space: voyage-4-nano shares an embedding space as its larger siblings
voyage-4-nano is ideal for local development and prototyping while providing an easy path to production.
- Shared Embedding Space: voyage-4-nano shares an embedding space as its larger siblings
Sage advice on crypto from Masood Boomgaard.
February 5, 2026 at 1:40 PM
Sage advice on crypto from Masood Boomgaard.
Love how the market works:
“AI will be everywhere” → sell SaaS stocks
“The AI bubble is imploding” → sell AI hardware stocks
Both can’t be true at the same time, yet U.S. hyperscalers are guiding toward $700B in CapEx for 2026.
“AI will be everywhere” → sell SaaS stocks
“The AI bubble is imploding” → sell AI hardware stocks
Both can’t be true at the same time, yet U.S. hyperscalers are guiding toward $700B in CapEx for 2026.
February 5, 2026 at 1:19 PM
Love how the market works:
“AI will be everywhere” → sell SaaS stocks
“The AI bubble is imploding” → sell AI hardware stocks
Both can’t be true at the same time, yet U.S. hyperscalers are guiding toward $700B in CapEx for 2026.
“AI will be everywhere” → sell SaaS stocks
“The AI bubble is imploding” → sell AI hardware stocks
Both can’t be true at the same time, yet U.S. hyperscalers are guiding toward $700B in CapEx for 2026.
They develop TinyLoRA, a new ft method. with TinyLoRA + RL, models learn well with dozens or hundreds of params. For example, they use only 13 parameters to train 7B Qwen model from 76 to 91% on GSM8K.
"Learning to Reason in 13 Parameters"
Paper: arxiv.org/abs/2602.04118
"Learning to Reason in 13 Parameters"
Paper: arxiv.org/abs/2602.04118
February 5, 2026 at 1:15 PM
They develop TinyLoRA, a new ft method. with TinyLoRA + RL, models learn well with dozens or hundreds of params. For example, they use only 13 parameters to train 7B Qwen model from 76 to 91% on GSM8K.
"Learning to Reason in 13 Parameters"
Paper: arxiv.org/abs/2602.04118
"Learning to Reason in 13 Parameters"
Paper: arxiv.org/abs/2602.04118
Why do tech companies even advertise during the Super Bowl? It feels like it doesn’t accomplish anything meaningful.
February 5, 2026 at 5:58 AM
Why do tech companies even advertise during the Super Bowl? It feels like it doesn’t accomplish anything meaningful.
I’m using Codex more than Claude Code these days. I used to rely on Claude Code more before.
February 5, 2026 at 1:47 AM
I’m using Codex more than Claude Code these days. I used to rely on Claude Code more before.