Why prompt injection is still possible in LLM applications
How a chat conversation actually reaches the model, and why it can still confuse untrusted content with trusted instructions.
Illia Pantsyr | Blog
Occasional writing on AI engineering, LLMs, and things I learn building software.
Latest posts
Smart, Safe, or Fast: Every conversation AI assistant picks two
Why conversational AI systems trade off between capability, control, and latency, and why the latency corner is effectively fixed.
How to lead your organization's AI transformation
Notes from a YOW! talk on why blocking AI tools, cloning ChatGPT internally, and appointing a Chief AI Officer all tend to fail.
12 essential client management rules
Notes from Jim Donovan's talk on client service: drop the jargon, use pauses, and offer advice that runs counter to your own interest.
Anthropic demonstrated how AI can analyze itself
How Clio surfaces usage patterns across millions of Claude conversations while keeping any individual conversation private.
Could GPU hardware failures limit AI training scaling?
Notes on an Epoch AI report asking whether GPU failure rates and checkpointing overhead put a ceiling on how large training clusters can grow.
Lost in the Middle: How Language Models Use Long Contexts
A summary of the paper showing that models retrieve far more reliably from the start and end of their context than from the middle.
Deployment Strategies
All-at-once, in-place, blue/green, canary and shadow deployments: how each works, what it costs, and when the complexity pays off.
Ways to reduce JVM Docker image size
Taking a Spring Boot image from 465MB down to 184MB with multi-stage builds, jlink, jdeps and a leaner base image.
Datafaker 2.0
What changed in Datafaker 2.0: schemas and transformers, Java Records support, and generating large volumes of test data.
Class Loaders in JVM: An Overview
How the JVM finds, loads and links a class, and what each of the bootstrap, platform and application class loaders is actually responsible for.