Most PDF parsers fail on real-world documents like machine manuals with complex tables and layouts. Docling emerged from IBM Research to solve this specific pain point, trading speed for accuracy. Red Hat integrated it into RHEL AI and InstructLab, while IBM embedded it across watsonx—enterprise adoption that signals this tool solves problems companies will pay to fix.

Anthropic's API restrictions turned OpenCode from another AI coding tool into a developer rebellion. 650,000 monthly users, 500+ contributors, and serious technical trade-offs—CPU abuse, memory leaks, and an architecture that actually works with any LLM provider. The metrics show momentum, but the controversies reveal what it costs to escape vendor lock-in.

Perplexica went from zero to 28,000 GitHub stars since April 2024 by offering what Perplexity AI won't: a fully self-hosted, privacy-first AI search engine. The project demands Docker expertise and homelab patience, yet thousands of developers choose configuration headaches over sending their queries to the cloud. One developer's MIT-licensed answer to proprietary AI search proves data sovereignty trumps polish.

Document conversion for RAG pipelines means wrestling with textract, losing table structure, and duct-taping libraries together. MarkItDown promises one pip install to turn PDFs, PowerPoints, and emails into clean Markdown—but PDF performance issues and async gaps show this problem isn't fully solved yet.

DeepSeek-V3 proves frontier AI training can cost $5.5M instead of $100M through MLA architecture and stability techniques. The catch: local deployment still requires 10× H100 GPUs and 768GB VRAM, pricing out individual developers. We examine whether efficiency innovations expand access or just change which organizations can compete.

Training video AI has been locked behind OpenAI's closed doors and million-dollar compute budgets. Open-Sora provides a fully documented, system-optimized path to train an 11B video model for $200K—with inference on consumer GPUs and 3-day training cycles that make competitive video generation accessible to ML teams.