IDEAS, DECISIONS & LESSONS
Notes on AI, architecture and building useful products.
3 articles
· 8 min read
agentic-ai
Four public AI safety benchmarks measure four different things, and the closest one to "official" leaves out Claude, GPT, and Gemini.
Read more· 6 min read
ai-safety
Open-weight models and API-only models aren’t just different licenses — they’re two different threat models. Article 2 in a series on AI attack and defense, covering how fine-tuning and abliteration strip refusal from models whose weights you can download.
Read more· 6 min read
ai-safety
How Anthropic, OpenAI and Google DeepMind actually test frontier models before release: internal red teams, scored frameworks, public bug bounties.
Read moreLooking for an article, exploring a product, or starting an AI project?
