A practical comparison of activation functions for ML engineers: ReLU and the dying neuron problem, GELU and why transformers switched from ReLU, SiLU and Swish for efficiency, SwiGLU and GeGLU gated variants used in Llama and PaLM, the 2/3 expansion factor explained, and a decision guide for choosing activations in CNNs versus transformer architectures.
A complete guide to running Ollama as a persistent Linux systemd service: checking if the installer already created a service, writing a service unit file from scratch with a dedicated ollama user, adding NVIDIA GPU support with group permissions and environment variables, all key environment variables including OLLAMA_HOST, OLLAMA_MODELS, OLLAMA_KEEP_ALIVE, and OLLAMA_NUM_PARALLEL, managing the service with systemctl, viewing logs with journalctl, a separate oneshot service to pull models at boot, storing models on a separate data drive, and troubleshooting common startup failures.
A practical guide to weight initialization for ML engineers: why poor initialization causes vanishing and exploding gradients, Xavier initialization for tanh and linear activations, Kaiming initialization for ReLU networks, GPT-2 style scaled residual initialization for LLMs, embedding initialization, and a concrete checklist for initializing custom architectures correctly.
A complete guide to using Ollama from Deno and Bun: importing the ollama npm package in Deno with the npm: specifier and –allow-net, streaming responses, using the OpenAI SDK in Deno, a deno.json configuration with import maps and task shortcuts, installing and using the ollama package in Bun, writing fast CLI scripts that compile to standalone executables with bun build –compile, testing Ollama integrations with bun:test, and when to choose Node.js vs Deno vs Bun for local LLM projects.
A practical comparison of normalization layers for ML engineers: what batch norm, layer norm, group norm, and RMSNorm each compute and why it matters, batch norm train/eval discrepancy and the hidden bugs it causes, why layer norm is the transformer default, RMSNorm as used in Llama and Mistral, group norm for small-batch detection tasks, and a decision guide for choosing the right normalization for your architecture.
A complete migration guide from OpenAI to Ollama: the two-line Python and JavaScript SDK change, streaming that works identically, batch embeddings with nomic-embed-text, an environment variable pattern for switching between cloud and local without code changes, LangChain migration via both the OpenAI-compatible endpoint and the native ChatOllama integration, LlamaIndex migration with OpenAILike, and a clear breakdown of what the compatibility layer supports versus where gaps exist including function calling, vision, and logprobs.
A practical guide to reward model training for ML engineers: preference data structure and quality requirements, Bradley-Terry pairwise loss and training objective, LoRA-based reward model fine-tuning, pairwise accuracy as a training metric, reward hacking failure modes and mitigations, best-of-N sampling as a simpler alternative to PPO, and when to use rejection sampling fine-tuning versus full RLHF.
A practical guide to using local LLMs for research paper workflows: fetching papers from arXiv with the Python library, extracting full text from PDFs with PyMuPDF, structured seven-point paper summarisation prompts, quick abstract-only relevance triage for reading list management, batch processing a list of arXiv IDs with Markdown output, an interactive Q&A session tool for deep reading, choosing between Llama 3.2 and Mistral Nemo for different paper lengths, and the privacy advantages of local processing for embargoed or proprietary research.
A technical deep-dive into diffusion models for ML engineers: the forward noising process and variance schedule, the reparameterisation trick for efficient training, ε-prediction U-Net training loop, DDIM deterministic sampling and fewer inference steps, latent diffusion and how Stable Diffusion works, and the score matching theory connecting it all.
A clear explanation of GGUF quantization formats in Ollama: what quantization does to model weights and memory, the full spectrum from Q2_K to F16 with sizes and VRAM requirements for a 7B model, why K-quantization is better than older Q4_0, a reference table of quality vs size trade-offs, when to use Q4_K_M vs Q5_K_M vs Q8_0, how to pull specific quantization tags, what Ollama pulls by default, and why lower quantization can actually be faster on memory-bandwidth-constrained consumer GPUs.
