BLEU, or Bilingual Evaluation Understudy, is an automatic metric originally developed for machine translation. It compares short sequences of words in generated text with one or more reference translations, combines precision across several n-gram lengths, and applies a penalty when output is too short. BLEU is inexpensive and useful for comparing systems on the same dataset, but it does not directly measure factuality, fluency, meaning, or human preference. Valid alternative wording can receive a low score, while awkward text can match many reference phrases. Results depend on tokenization and implementation, so reporting should include the exact evaluation settings and, when possible, human judgment.
AI video startup Higgsfield raised a $400 million Series B at a $5.4 billion valuation, more than quadrupling its worth in eight months as it pushes into enterprise marketing.
Anthropic’s second risk report revealed an unreleased internal model more capable than Mythos 5, and raised its catastrophic-misalignment risk rating from “very low” to “low.”
The US is preparing to warn dozens of countries they must choose between its Pax Silica AI coalition and China’s competing framework, or risk exclusion, according to a leaked draft.
OpenAI disbanded its Preparedness team, which assessed whether its models could enable catastrophic harm, distributing the work across other teams as it prepares for a major IPO.
Investors expect Anthropic to target a valuation of $2 trillion or more in an October IPO, which would be the largest listing in history, driven by surging revenue.
Alibaba released the open weights for Qwen3.8-Max, a 2.4-trillion-parameter model it says trails only Anthropic’s Fable 5, its first openly released flagship-class system.
Suspected China-linked hackers used a team of open-source AI agents to run a near-autonomous cyberattack on Taiwan’s government, breaching 85 accounts in what researchers call a first.
DeepSeek’s flagship V4 Pro left preview as build 0813, a 1.6-trillion-parameter model that nears frontier benchmarks at roughly a sixtieth of Anthropic’s Fable 5 pricing.
Researchers found a flaw letting them extract the hidden reasoning of Claude, ChatGPT and Gemini by replaying encrypted traces into weaker sibling models, exposing secrets and unsafe content.
SpaceXAI released Grok 4.6, built for long-running agents and visual work, claiming parity with GPT-5.6 Sol on a composite benchmark while keeping its low, efficiency-focused pricing.