I Measured 12 Korean Tokenizers. Two Were Borrowed.
Twelve tokenizers, one parallel corpus. The top three are recent Korean releases, while two others carry base vocabularies identical to Llama 3 and Mistral. Read more »
LLM
Open Weight Licenses Aren’t Really Open
I measured the licenses and gating of the 100 most downloaded open models, separating what you can freely ship from what sits behind an access request. Read more »
The Top Spot in AI Model Downloads Isn’t a Chatbot
A direct tally of HuggingFace’s top 100 downloads puts a 2021 embedding model at number one, not a chatbot. Likes point the opposite way from downloads. Read more »
Why AI Agents Fail: Context Engineering 101
Context engineering decides what enters an AI model’s context window — the choice that makes the same model succeed or fail. It underpins reliable AI agents. Read more »
Prompt Injection: How AI Agents Get Hacked
Prompt injection exploits an LLM’s inability to tell trusted instructions from untrusted data: direct vs indirect attacks, the trifecta, and real defenses. Read more »
AI Agent Benchmarks Are Breaking in 2026
Why soaring AI agent benchmark scores mislead — Berkeley’s exploit of 8 benchmarks, OpenAI’s SWE-bench retirement, and the pass@1 reliability gap. Read more »
Mixture of Experts (MoE), Explained
How do giant AI models stay fast despite huge parameter counts? A walk through Mixture of Experts: routers, active vs total parameters, and the VRAM trade-off. Read more »
AI Compute Costs: The Great Divergence
Frontier training costs head toward a $1B run by 2027, while running the same capability fell 1,000x in three years — the great AI compute divergence. Read more »
Reasoning Models Explained: How AI Thinks
A plain-English guide to reasoning models: how chain-of-thought and test-time compute let AI think before answering, from o1 to the real cost of overthinking. Read more »
Small Language Models: AI That Runs On Your Phone
Small language models run on 1-7B parameters and execute offline on phones, cutting cloud costs up to 90% while nearing the quality of far larger cloud models. Read more »