Deterministic vs Probabilistic Systems — Bridging the Trust Gap

August 8, 2026

A few days ago I listened to an Andrej Karpathy video where he said computers could be thought of as either calculators or neural networks. That framing stuck with me. For decades, computing has been deterministic: binary logic, exact outputs, repeatable behavior. We also had fuzzy logic systems, which are conceptually close to neural networks, but they never became mainstream largely because they were hard to predict and verify.

Today, large language models like GPT and Claude are becoming the default interface to knowledge. They are probabilistic: the same prompt can produce slightly different answers, and even a correct-looking answer carries no certificate of correctness. That creates a tension, especially in fields like healthcare where accuracy, compliance, and legal liability matter.

Lossy vs. Lossless

The way I think about this is by analogy to image compression. In radiology we almost always use lossless compression, even though lossy compression would save storage and bandwidth. The reasons are well understood: legal liability, and the fact that lossy artifacts degrade downstream post-processing and analysis.

Neural network language models feel similar to lossy compression. They give us convenience, summarization, and natural-language access to information, but they discard exact provenance and reproducibility. The output is based on probability, so it will never be 100% accurate — and worse, even when it is accurate, you cannot know that from the output alone.

Why This Matters Now: APOLLO

This week I read about APOLLO, the Harvard Medical School/Mass General Brigham foundation model trained on 25.2 billion medical events across 33 years of patient records. It is a powerful example of what probabilistic models can do in clinical prediction — encoding longitudinal patient journeys and answering predictive queries.

APOLLO also makes the trust problem concrete. The convenience is real, but so is the need for accuracy and compliance. If a clinical decision support tool gives different answers to the same question, or if it cannot explain how confident it is, adoption becomes risky no matter how impressive the model is.

What Verifiability Could Look Like

I keep coming back to the Monte Carlo method. In physics and radiation therapy planning, Monte Carlo simulations run a probabilistic process millions of times to produce a stable, trustworthy distribution. We do not trust one random path; we trust the aggregate pattern.

A similar mindset may be needed for LLMs in healthcare:

• Consistency checks. Run the same clinical query multiple times. If answers vary widely, that itself is a signal to stop and escalate.
• Deterministic grounding. Use the LLM for reasoning and retrieval, then verify any factual claim against a structured, rule-based source such as a drug formulary, guideline database, or curated knowledge graph.
• Confidence scoring. Communicate uncertainty as clearly as the answer. In medicine, knowing that a model is uncertain can be as valuable as the answer itself.
• Specialized small models. Where possible, train narrower models for specific tasks rather than relying on one general model to recall everything from training data. A focused model is easier to validate and monitor.

The Real Gap

The point is not to undermine LLMs. They are genuinely useful. The point is to highlight the gap between deterministic and probabilistic systems, and to ask how we bridge it. Convenience is valuable, but in healthcare we also need compliance, accuracy, and confidence. Building that bridge — where probabilistic tools become verifiable enough to trust — is where the hard and important work lies.

Further reading:
APOLLO AI learns from 25 billion medical events to predict future disease — News-Medical, April 2026
Apollo: A Lightweight Multilingual Medical LLM — arXiv
Deterministic vs. Probabilistic AI — Cycode, 2026
Deterministic or Probabilistic? The Clinical AI Distinction — Longevitix, Jan 2026
Monte Carlo Committee Simulation with LLMs — medRxiv, 2026

Tags: AI, Healthcare, LLM, Deterministic, Probabilistic, Trust, Verifiability, APOLLO

[tag AI,healthcare,LLM,trust,verifiability,Apollo]