Tuesday 26 May 2026

Using AI coding assistants produces faster first drafts but requires significantly more review, verification, and correction time, resulting in slower overall development despite apparent productivity gains. The tradeoff between speed and code quality/understanding when using AI tools is a recurring friction point.

LLM training at scale requires massive flash storage infrastructure (2 petabytes in this case), highlighting the storage I/O bottlenecks and geopolitical supply-chain constraints that complicate AI training hardware procurement.

Developers are increasingly relying on AI tools and search over structured learning resources like programming books, raising concerns about shallow understanding, reduced debugging capability, and long-term skill degradation in the profession.

Ollama's gemma4 tool call parser fails with invalid JSON errors when model output contains backticks or single quotes in arguments, breaking tool calling workflows in coding agents even after a supposed fix in v0.20.1.

Ollama's model push fails with "max retries exceeded" and broken pipe errors on slow or unstable connections when uploading large files (7-8 GB), with no resume capability.

Qwen2.5-VL models in Ollama exhibit token repetition loops during OCR tasks that do not occur on the reference HuggingFace demo, and repeat_penalty parameter appears to have no effect on the issue across 7b, 32b, and 72b variants.

Running Nemotron-3-nano on Ollama causes a hard crash with an assertion failure in llama-sampling.cpp, killing the runner process and returning a 500 error to the API caller.

Custom Qwen3VLMoE models that are valid per llama.cpp tooling and run correctly in HuggingFace Transformers cause a nil pointer dereference panic in Ollama's vision model loader, making it impossible to run custom architectures derived from supported model families.

Using AI coding assistants produces faster first drafts but requires significantly more review, verification, and correction time, resulting in slower overall development despite apparent productivity gains. The tradeoff between speed and code quality/understanding when using AI tools is a recurring friction point.

C compiler extensions create portability friction across alternative compilers (non-GCC/Clang), forcing developers to work around non-standard behaviours that are silently accepted by dominant compilers but break on others.