Kimi K3 Closed the Coding Gap. The Hallucination Rate Is 51%.
The first open 3T-class model leads Program Bench and competes with GPT-5.6 Sol on coding. Artificial Analysis found a 51% hallucination rate. Architecture explains both.
Category
Editorial
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
96 results in this archive
Bonsai 27B Fits in 4 GB. Post-Training Quantization Never Could.
PrismML's 27B model runs on an iPhone at 11 tokens per second. The benchmark breakdown shows where the 10% loss lands — and why that distribution matters more than the headline.
Codex Encrypted Its Sub-Agent Tasks. Only OpenAI Has the Key.
OpenAI's multi-agent system encrypts every task sent between agents. Operators see empty strings. OpenAI's servers hold the keys. The fix exists and isn't merged.
Anthropic Rewrote Bun in 11 Days. 13,365 Unsafe Blocks Came With It.
Bun's AI migration from Zig to Rust passed every test on six platforms. The official unsafe audit found 181× more unsafe density than comparable Rust projects and five functions with live bugs.
Illinois Mandated AI Audits. Labs Got the Burden Without the Shield.
Governor Pritzker signed the first mandatory annual frontier model audit law on July 6. The liability protection OpenAI lobbied for in parallel never came.
GPT-5.6 Solved the CDC Conjecture. Peer Review Hasn’t Caught Up.
OpenAI's GPT-5.6 Sol Ultra published a three-page proof of a 50-year open problem in graph theory. It's the second AI math breakthrough in 51 days. Nobody has verified it yet.
Telisha Jones Built an AI Gospel Voice. Suno’s Lawsuit May End It.
A Mississippi poet's AI singer hit #3 on gospel radio and became the first AI act on a Billboard radio chart. Now the platform behind the voice faces a fair-use ruling.
The Optimizer Creates a 7× Alignment Gap. No Lab Says Which.
A new paper tests 12 models across four optimizers and finds optimizer choice dwarfs model scale as a driver of emergent misalignment. No frontier lab publishes which optimizer they use.
GitHub Built Guardrails Against Prompt Injection. One Word Got Through.
The 'Additionally' bypass broke GitHub Agentic Workflows' defense-in-depth. Guardrails enforced by the attacked model cannot separate instruction from data—and that's not a patch problem.