konakona
I'm Samuel
Machine learning research, Swift, and the occasional debugging rabbit hole.
Also available in your terminal:
10 posts
All 10 →2026
Two pages Claude built from my notes, and some thoughts on Opus 5.5
Two interactive pages Opus 5.5 one-shotted from my notes repo, plus my impressions of its taste, how it communicates, and Claude Code
Routing Tailscale exit node traffic through Mihomo
A rescue tunnel, narrowly scoped TUN interception, and a DNS poisoning investigation that made a Tailscale exit node work with Mihomo
Reproducing Qwen3.8-27B's Terminal-Bench 2.1 Results: Setup, Full Run, and Pitfalls
A complete reproduction of Qwen3.8-27B on Terminal-Bench 2.1 using SGLang, Harbor 0.21.0, and Terminus-2 across 16 A100 40GB GPUs, including setup, debugging, and result auditing.
NITP: Next Implicit Token Prediction for LLM Pre-training
Predicting the next token in both the output space and the hidden space.
From Linear Attention to Test-Time Training
Linear attention reduces the quadratic cost of standard attention to linear, which is great. But can we do better?
Debugging an OpenClaw web_fetch Failure
proxychains4 DNS hijacking collided with OpenClaw's SSRF protection, leaving web_fetch completely broken.
-
Paper walkthrough · 21 slides
Self-Distillation Enables Continual Learning (SDFT)
How a model can teach itself from demonstrations on-policy, why it forgets less than SFT, and where the theory and the training loss part ways.