Another Dev's Two Cents

Another Dev's Two Cents

Sign in Subscribe

Latest — 23 Oct 2025

GMKTec Evo-X2 Ryzen AI Max 395+ Benchmarks

I recently got my hands on a GMKTec Evo X2 for local model inference. Here's my hardware details nish@gmktec-evo-x2:~$ sudo lshw -short H/W path Device Class Description ========================================================= system NucBox_EVO-X2 (EVO-X2-001) /0 bus GMKtec /0/0 memory 64KiB BIOS /0/b memory 1280KiB L1 cach

More issues

On the DORA 2025 AI Report - AI Adoption and use

The DORA State of AI Assisted Software Development report came out recently. It's a massive 142 page report that details analysis done by the DORA team. It captures trends and observations covering a 5000 participant study that focuses on AI Adoption and tool use in the Software industry.

Notes on OpenAI's AppSDK

OpenAI's dev day was today. While I wrote up a short summary of what was announced on bluesky, one of the major announcements was the AppSDK for ChatGPT. It looks like OpenAI plans to position ChatGPT as a platform for the future not unlike the Google Play and

Vector Norms

A Norm of a vector \(\vec{v}\) describes the magnitude or size of a vector. It is usually denoted as \(||\vec{v}||\). There are a few common norms worth discussing. Eucledian Norm (L2 Norm) Let's consider a vector v = [3, 4] we can calculate the Eucledean norm as,

Notes on - Why do LLMs freak out over the seahorse emoji?

A fantastic deep dive into the seahorse emoji phenomena[1] was recently published by Theia[2]. It's engaging, well presented and worth reading. The post presents a case using meta-llama/Llama-3.3-70B-Instruct. However I wanted to verify this behavior with smaller models which unsurprisingly fail the challenge as

How LLM Structured Decoding works

Last week I happened to be in a discussion that involved getting an LLM to generate JSON reliably. A major frustration expressed was that no matter how much they tried the model would often fail to follow instructions during generation. I pointed out that most major vendors support some variant

Notes on the phi-4-reasoning Technical Paper

Phi-4 Reasoning Microsoft recently released the phi-4 reasoning model as well as its technical report. They explore using supervised finetuning as well as synthetic dataset curation to train phi-4 a 14B parameter model to compete with and often outperform significanly larger models such as DeepSeek-R1-Distill-Llama-70B. We present Phi-4-reasoning, a 14-billion

Writing a backup service for Ubuntu in Rust

While performing some routine maintenance of my Ubuntu server, $ docker compose pull $ docker compose up -d I ran into an interesting situation where nginx-proxy-manager failed to serve my blog properly, I wasn't sure what the issue was but these types of issues are usually solved with a restart.

Adding Gaussian Noise to a signal in Python

Gaussian noise is data that is added to a signal in order to introduce a distortion. The data follows a Gaussian/Normal distribution. It's a well understood distribution often used to introduce noise to training data as an augmentation technique. Generating noise to add to a signal is

Building an Ubuntu service for my Antec Flux Pro

I recently put together a new PC build. It's primarily aimed at machine learning so I wanted to get a case that had really great airflow. I asked around and was directed to a list put together by GamersNexus[1] highlighting their top cases of the year. The