CharlieKey takeaways from your favorite podcasts
← 7 November 2026Latent Space

Why AI Still Misses Rows in PDFs — with Adit Abraham of Reducto | Let Them Cook

Adit Abraham of Reducto argues that document processing needs human-level fidelity, specialized post-trained models and a harness that audits results, and that parsing is only the first step toward models doing white-collar document work.

Watch on YouTube ↗

Charlie's summary

  1. Document parsing is a self-driving problem: agents need human-level fidelity, not lane assist

    Adit says PDFs were designed so printers stay faithful to the original, not so agents can read them. If models are to do human work, they must read pages the way a person does, which he calls “a really hard vision problem.” Reducto likens it to self-driving research: lane assist has existed for years, but a Waymo driving fully end to end is a different bar.

    ▶ 2:02
  2. Frontier models fail on recall: they silently drop rows, which is why an auditing harness matters

    Reducto's deep extract benchmarks showed a precision/recall trade-off. Specialized platforms were worse than frontier models with uncapped reasoning, but the frontier models failed on recall, “silently dropping rows” and context. Adit says this loss pattern is hard to find in production because you don't know what you failed to extract. A harness that audits results fixes it. He sees it as complementary to frontier models, not either/or.

    ▶ 7:10
  3. Post-trained task-specific models beat general models on accuracy, cost and latency together

    Adit's team cares about the Pareto frontier of accuracy, cost and latency. For a specific task, he says a post-trained model outperforms on all three, so Reducto's stack has shifted steadily in-house. Its new model R1 handles layout and last-mile corrections end to end, and is priced below the per-page cost of the base models with better accuracy than the most accurate ones. Foundation models remain for a few horizontal tasks.

    ▶ 5:07
  4. Benchmarks have become marketing, so test on your own documents

    Adit is frustrated that “everybody's number one on their own benchmark.” Reducto used Micro1 so the benchmark documents and ground truth were not chosen by Reducto. Even independent benchmarks are imperfect and give only a directional sense. What matters most is that people try the product on their own documents.

    ▶ 8:10

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.