# Blog — John Leonardo

> Technical posts and AI/ML experiments by John Leonardo.

- [Thoughts on AI now in August 2026](https://jdleo.me/blog/thoughts-on-ai-aug-2026.md) (2026-08-28): My opinions on where we are, and where we go from here.
- [My Current AI Opinions](https://jdleo.me/blog/current-ai-opinions.md) (2026-03-15): My current thoughts on AI and how to be good at it.
- [TinySafe v3: How I Built a Near-SOTA Safety LLM for Under $100](https://jdleo.me/blog/tinysafe-v3.md) (2026-03-07): A 4B parameter model fine-tuned with QLoRA gets within 0.008 F1 of SOTA on ToxicChat, then hits a wall. Five versions, $100, and a lot of lessons about why 0.83 is so hard to break.
- [I Built TinySafe, a Safety Model that Beats 8B Guard Models with 71M Parameters for $37](https://jdleo.me/blog/tinysafe.md) (2026-02-28): Building a sub-2ms safety classifier that outperforms LlamaGuard, ShieldGemma, and every encoder-based model in its weight class.
- [Which AI Models Lean in Which Political Direction?](https://jdleo.me/blog/ai-political-compass.md) (2026-01-25): I asked the top LLMs 20 different political compass questions to figure out how left/right/libertarian/authoritarian they were.
- [Are Computer Science Majors "Cooked" because of AI?](https://jdleo.me/blog/are-computer-science-majors-cooked.md) (2025-09-22): A speculative deep dive on whether AI will augment or replace new software engineers.
- [I Trained AI on 70k Clash Royale Battles to Settle the Ultimate Debate: Does Your Deck Actually Matter?](https://jdleo.me/blog/clash-royale-ai-deck-prediction.md) (2025-09-14): Using machine learning on 70,000+ real battles from the Clash Royale API to definitively answer whether deck composition actually predicts victory, or if it's just skill and luck.
- [The Nutrition Prediction Benchmark: Testing LLMs on Google Cafeteria Menus](https://jdleo.me/blog/nutrition-benchmark.md) (2025-09-07): Testing large language models on their ability to predict accurate nutritional information from Google cafeteria dishes, revealing surprising insights about AI's understanding of food science.
- [The Semantic Diversity Benchmark: A New Way to Test AI Language Models](https://jdleo.me/blog/semantic-diversity-benchmark.md) (2025-08-31): A simple but powerful benchmark for testing AI language models by asking them to generate maximally semantically unrelated words, revealing surprising insights about model capabilities.
- [New Grad Job Search for Software Engineers](https://jdleo.me/blog/new-grad-swe-search.md) (2025-08-02): A quick guide for new grads looking to break into the software engineering industry, including networking tips, technical preparation advice, and insights into why the current job market is more challenging.
- [Transformers are Limited](https://jdleo.me/blog/transformers-are-limited.md) (2025-08-01): The transformer architecture is fundamentally limited for true reasoning.
- [Welcome to My Blog](https://jdleo.me/blog/first-post.md) (2025-06-09): Welcome to my blog!