§02 — Wire
Latest developments from the AI world
AISI: Anthropic Mythos Preview and OpenAI GPT-5.5 completed the 32-step offensive cyber test at expert level; attack capability is doubling every four months.
Llama 4, ticari kullanım için ücretsiz. 405B parametre, 128K bağlam.
Four Chinese labs released open-weight coding models in 12 days; close to Western frontier on agentic engineering, at roughly one third the cost.
At Cloud Next '26 Google rolled out 260+ announcements including the Gemini Enterprise Agent platform, the open-weight Gemma 4, 8th-gen TPUs and free Google Vids video generation.
Fransız Mistral, GPT-4 seviyesinde performans ve Avrupa veri egemenliği vaadediyor.
A Zhejiang University study finds that the Centaur model's 160-task performance likely stems from overfitting; the "pick option A" test exposes pattern matching, not understanding.