Skip to content

Back to the catalog

jakewnuk.comarticle

The Digital Context Gap

First post in a series on building agentic infrastructure. It separates traditional machine learning (gradient-boosted trees, fraud detection) from black-box neural networks like LLMs. It argues LLMs add value in two cases: non-deterministic workflows that can't be hand-programmed, and as a "digitally native" interface that narrows the "digital context gap" between human and machine speed at handling context. It also challenges its own thesis, noting that models lack real comprehension and that even a 1% error rate is costly at volume. It closes with practical principles: start from the problem, design for failure, add validation layers and human-in-the-loop checkpoints, and treat LLMs as enhancement rather than replacement.

TL;DR Essay on when non-deterministic LLM workflows are worth using, framing the "digital context gap" and a "better together" human-in-the-loop approach.
Open original ↗
Document

What it claims

This is a conceptual, experience-based essay (first in a series on agentic infrastructure), not a method or release. Its central term differs from what the summary implies. The "digital context gap" here is not the gap between the context an agent needs and what it has — it is the gap between how fast humans can input/navigate digital context versus how fast a machine-native process can. Humans think faster than they speak and speak faster than they type; models parse code, plain language, and byte code at machine speed. LLMs offer a "digitally native" navigation tool to close that speed gap, alongside their older role in non-deterministic workflows (classification, generation, prediction) that resist traditional programming.

Key points the summary omits

  • The author scopes "predictive models" broadly (gradient-boosted trees, vision-language, multi-modal), not just frontier LLMs, and frames neural nets as black-box weighted layers that predict likely output without "understanding."
  • Stated production cost: a 1% error rate on high-volume tasks means millions of wrong outputs, and unlike traditional software there is no traceable fix — only retraining, fine-tuning, or prompt engineering.
  • The author explicitly hedges his own thesis: the context gap "may be more aspirational than real" — models offer speed and breadth of context processing, not depth of understanding; output is "statistically plausible, not semantically grounded."

How to use it

Adoption principles the post states directly:

  • Start with the problem, not the technology; use non-deterministic workflows only where deterministic alternatives are genuinely infeasible.
  • Design for failure: assume outputs will be wrong/biased/incomplete, and build validation layers and human-in-the-loop checkpoints into every LLM system.

Limits

Opinion piece with no benchmarks, code, or system design — despite the "agentic infrastructure" series framing, this installment stays at the conceptual level. It argues an "enhancement over replacement" ethos and even raises the counter-question of "whether there will be enough work left" for human operators, without resolving it.

Technologies named here

  • ChatGPT
  • LLM
  • Vision language model
  • gradient-boosted trees
  • neural networks

Techniques named here

  • Gradient-boosted trees
  • Benford's Law
  • Non-deterministic LLM workflows
  • Plain-language interfacing with machines
  • Retraining, fine-tuning, or prompt engineering
  • Digital context gap
  • Human-in-the-Loop Validation

Related in ai-policy