Saturday 16 May 2026

Broad concern that organizations are making poor decisions driven by AI hype — entire companies may be operating under distorted assumptions about AI capabilities, leading to systemic engineering and product misjudgments.

npm's package ecosystem continues to suffer from recurring supply-chain and reliability incidents, with no systemic fix emerging despite repeated occurrences — framed as a structural problem with the package manager.

Debate around whether sigmoidal/S-curve arguments can reliably predict AI capability trajectories and slowdowns — questions the epistemic foundations used by both AI optimists and pessimists to forecast progress.

Gemma4's tool call parsing in Ollama remains broken even after a fix release (v0.20.1); the parser fails on valid but non-JSON-formatted tool call content (backticks, single quotes), causing errors in agentic coding tools like opencode.

Ollama's model push fails with "max retries exceeded" on slower or interrupted connections when uploading large files, with no resume capability, forcing users to restart the entire upload from scratch.

Qwen2.5-VL models in Ollama produce severe token repetition loops during OCR/image tasks that don't occur in the official HuggingFace demo, and repeat_penalty settings appear to have no effect on mitigating the issue.

Nemotron-3-nano model causes a hard crash in Ollama via an assertion failure in llama-sampling.cpp, resulting in a SIGABRT and a 500 error returned to the client with no graceful recovery.

Custom Qwen3VLMoE models (validated by llama.cpp and working in HuggingFace Transformers) fail to load in Ollama with nil pointer dereferences and missing GGUF key errors, blocking use of custom/non-stock model architectures.

Essay arguing that the existing tech industry paradigm is collapsing under AI disruption but no coherent replacement has emerged, creating uncertainty about career paths, tooling investment, and organizational strategy in software development.

Git's UX and conceptual model are argued to be fundamentally flawed for modern workflows, with persistent pain points around branch management, history rewriting, and the steep learning curve that tooling workarounds have not resolved.