`ai-engineering-from-scratch`: 435 lessons for end-to-end AI engineering

// related

• #0001

  #0001Otherr/MachineLearning2 days ago

  Vision LLMs vs. OCR for PDF Q&A: OCR Still Wins on Cost and Accuracy

  60radar

  For complex PDF Q&A, vision LLMs are pricier and less accurate than OCR pipelines. Stick with OCR for better cost-performance and reliability.

  • —The native vision LLM approach was the most expensive at $0.2552/query and ranked 5th in accuracy (52.0%) out of six methods tested.
  • —Vision models underperformed on chart- and table-heavy pages, the very area they were expected to excel in. Premium OCR handled these better.
  • —The vision LLM had a 7% intrinsic failure rate that persisted after retries, while OCR-based pipelines showed 0% failure, indicating higher reliability.

  Source: www.reddit.com/r/MachineLearning/comments/1tm0cqg/visionRead original →
• #0002

  #0002OtherGeekNews2 days ago

  `Electrobun 2.0` to Split from `Bun` After Rust Rewrite

  40radar

  ElectrobunDesktop app framework — packages native apps with web tech

  Desktop app runtime dependencies are changing, with less reliance on Bun. Useful only if you are evaluating Electron alternatives; otherwise low urgency.

  • —Electrobun 2.0 is moving toward a Rust rewrite and reduced Bun dependency, changing its runtime structure.
  • —The split was influenced by concerns that Anthropic lacks enough human review, staged rollout, and stabilization process.
  • —This is a niche signal for desktop app stacks: wait for migration notes before building production workflows on it.

  Source: news.hada.io/topic?id=29815Read original →
• #0003

  #0003OtherGeekNews2 days ago

  `mimalloc`: Microsoft's High-Performance Drop-In `malloc` Replacement

  40radar

  mimallocMemory allocator — drop-in `malloc` replacement for speed

  A no-code-change allocator swap can cut memory overhead and improve throughput in native workloads. Worth testing when hosting bills or latency are driven by allocation-heavy code.

  • —mimalloc works as a drop-in malloc replacement, so existing programs can adopt it without source changes.
  • —Claimed speedups: 13% faster than tcmalloc and over 2.5x faster than jemalloc in the cited comparison.
  • —The pitch is consistency across workloads, not winning one narrow benchmark. That matters for production services with mixed traffic.

  Source: news.hada.io/topic?id=29812Read original →