@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
I just want to say I'm absolutely livid about Trump / Israel / AIPAC. That's it.
04.03.2026 06:02 — 👍 13 🔁 0 💬 0 📌 0
Im trying to build this thing for real again 1 year later: github.com/pboueri/inte...
If specs are truth then you should be able to wipe and rebuild your codebase reproducibly with “small” deviations. I think this is where it lands - like Makefiles
typeagent by Guido van Rossum (you know... python creator)
This is an in-progress project aiming at a Pythonic translation of TypeAgent KnowPro and a few related packages from TypeScript to Python.
github.com/microsoft/ty...
I just discovered a new number that did not existed before!
At Number Research Inc., they are attempting to find and document all* available numbers.
numberresearch.xyz
Google’s CEL-expr-python: the Common Expression Language in Python, now open source
opensource.googleblog.com/2026/03/anno...
How to Kill the Code Review by Ankit Jain
Human-written code died in 2025. Code reviews will die in 2026.
www.latent.space/p/reviews-dead
A strong guy
04.03.2026 02:02 — 👍 12 🔁 1 💬 1 📌 0
That loop matters because it gives the agent:
- Real feedback (errors, diffs, logs)
- Externalized state (repo, files, docs, worktrees, outputs)
- Steerability over time (you can course-correct based on outcomes)"
developers.openai.com/blog/run-lon...
Run long horizon tasks with Codex by Derrick Choi (OpenAI)
"In Codex, the loop is roughly:
1. Plan
2. Edit code
3. Run tools (tests/build/lint)
4. Observe results
5. Repair failures
6. Update docs/status
7. Repeat
This guide explains why and how to use it, and includes a faster Polar Express implementation alongside the official and original versions."
04.03.2026 01:32 — 👍 0 🔁 0 💬 0 📌 0
LLM Research Kit & Muon Optimizer Guide
"The Muon optimizer is replacing AdamW across the industry (DeepSeek, OpenAI, Meta, Moonshot AI). PyTorch 2.10 now includes an official implementation!
github.com/vukrosic/muo...
Two concrete improvements: the constant c goes from 0.03 to 0.13, and the solution partitions the entire vertex set into light components rather than just a subset.
04.03.2026 01:23 — 👍 1 🔁 0 💬 0 📌 0
Scaling long-running autonomous coding by Wilson Lin (Cursor)
They used the Marcus-Spielman-Srivastava interlacing polynomial method, a different approach from existing solutions.
cursor.com/blog/scaling...
Aadi Kulshrestha implemented tiled matrix multiplication on the GPU. This was an interesting exploration of machine learning and concurrency that directly connected to fundamental linear algebra operations.
github.com/r-chong/TSTo...
Databrick's FlashOptim
A better implementations of Adam, SGD, etc, that compute the same updates but save tons of memory.
Paper: arxiv.org/abs/2602.23349
Repo: github.com/databricks/f...
K-Search
It rethinks GPU kernel generation. Instead of treating LLMs as simple stochastic code generators, they leverage them as co-evolving Intrinsic World Models.
Repo: github.com/caoshiyi/K-S...
Paper: arxiv.org/abs/2602.191...
A detailed writeup on Google kCTF kernel exploits.
"A step-by-step guide to exploiting a 20-year-old bug in the Linux kernel to achieve full privilege escalation and container escape, plus a cool bug-hunting heuristic."
blog.calif.io/p/a-race-wit...
dLLM: Simple Diffusion Language Modeling
It packages the whole diffusion-LM workflow into a single modular framework. Standardized training, plug-and-play inference algorithms, reproducible evaluation.
Paper: www.arxiv.org/abs/2602.22661
Repo: github.com/ZHZisZZ/dllm
Model: huggingface.co/dllm-hub
Step 3.5 Flash 🔥New MoE model from StepFun
huggingface.co/stepfun-ai/S...
huggingface.co/stepfun-ai/S...
✨ Base & Base-Midtrain
✨ 196B total/11B active - Apache 2.0
✨ 256K context
✨ High-speed reasoning & agentic tasks
A great news for Intel.
newsroom.intel.com/corporate/in...
Why should you buy the new Macs with the M5 Pro or M5 Max? Because they have “super” cores. The base M5 doesn’t include these “super” cores, and even Intel’s newest CPU, Panther Lake, doesn’t have them either.
...and remember - they’re not just “performance” cores. They’re better. They’re “super”.
FYI: gpt-5.4-??? coming soon.
03.03.2026 19:59 — 👍 3 🔁 0 💬 0 📌 0You know - AI context.
03.03.2026 18:55 — 👍 0 🔁 0 💬 0 📌 0They're leaving...
03.03.2026 18:51 — 👍 1 🔁 0 💬 1 📌 0
Interesting... Alibaba Qwen team being Llama-fied. The old team is out and the new team is in.
New team: scholar.google.com/citations?hl...
IQuest-Coder-V1 Update! 7B & 14B series now on
@hf.co 🔥
huggingface.co/collections/...
✨ 7B/14B - Base, instruct, thinking
✨ Optimized for tool use & CLI agents
✨ 128k context length
FYI. All these foundational concepts are about 50 years old.
03.03.2026 06:44 — 👍 4 🔁 0 💬 1 📌 0No AI content today. I’m so happy reading about MAGA imploding in real-time that my social feed has become a bit polluted. I’ll need to start cleaning it up soon, but it’s so entertaining.
03.03.2026 04:50 — 👍 20 🔁 0 💬 0 📌 0This is a solid "starter pack" for anyone looking to build real-world applications with an AI. If you’re wondering how to begin, just start by asking an AI.
03.03.2026 02:35 — 👍 11 🔁 0 💬 1 📌 0
If you're learning to code with AI, prioritize learning these foundational concepts FIRST:
- Relational Databases and CRUD operations for data management.
- The MVC architecture for a clean separation.
- Version Control with Git, including the pull request.
