AI-ML
2025
Is the DRAM price spike in Sept 2025 and cooldown in November a sign of reduced AI demand ?
·442 words·3 mins
After a sharp surge in RAM prices through September 2025, the market is showing early signs of correction. Does this show a part of reduced demand for AI GPUs or this is a smaller correction?
Agentic AIs new attack surfaces, Data poisoning, tool-poisoning, and malicious MCP servers
·962 words·5 mins
Data poisoning is the deliberate injection of adversarial content into a model’s training data or a tool’s metadata so the LLM learns or obeys malicious instructions. In agentic systems that load third-party tools from MCP (model-connected platform) servers, poisoned tool descriptions or docstrings can trick an LLM into leaking secrets, executing harmful actions, or behaving as a covert proxy for attackers.
The hidden cost of data centers and inference hubs for local citizens
·875 words·5 mins
Data centers often leaves local communities with the costs while global firms capture the benefits.
The State of Local and Affordable Inference in October 2025
·1379 words·7 mins
An overview of the current landscape of GPUs and AI compute for local inference as of October 2025 from Nvidia and AMD to Intel, Apple, and the cloud.
Testing DeepSeek-OCR: Vision Text Compression for LLMs
·466 words·3 mins
Notes from testing DeepSeek-OCR as a local vision-language model for OCR and text compression on a large archive of screenshots. Includes observations on model performance, visual-token compression, and multilingual results.
Gemini Pro 2.5 in October 2025: decent text, shaky coding, tricky tradeoffs
·903 words·5 mins
A brief look at Gemini Pro 2.5 compared with ChatGPT 5 and Claude Opus/Sonnet, plus notes on Gemini 2.5 variants, NotebookLM, and mobile privacy concerns.
NVIDIA DGX Spark: underwhelming and late to the Party
·890 words·5 mins
NVIDIA’s DGX Spark arrives late as an AI inference system whose performance is lagging behind. With low speed unified VRAM, immature software optimizations, and heavy competition from Apple, AMD, and Intel, the Spark exposes how little remains of NVIDIA’s CUDA moat.
LLM false metric generation
·654 words·4 mins
There is a lot of synthetic data that is being generated by LLMs, These include false metrics.
ChatGPT 5 Working Guide: Practical Tips for Better Results
·725 words·4 mins
Creating a practical, no nonsense guide for using ChatGPT-5 efficiently by manual model switching and prompt style to each task while avoiding wasted tokens and bad outputs.
The Problem With Proprietary LLM Providers: Removing Model Access without recourse
·414 words·2 mins
OpenAI’s removal of GPT-4o, o3, and other models after GPT-5’s launch breaks fundamental MLOps principles. Without model versioning and control, data science workflows become unreliable. Local LLMs offer a better alternative for maintaining consistency.
