Qualcomm announced that its first-generation HBC (High Bandwidth Compute) chip has completed lab validation, achieving 18x the memory bandwidth of LPDDR and 5x the Tbps/W energy efficiency of HBM4 in testing. The architecture places the compute die directly beneath an LPDDR memory stack, reducing data transfer distances to the micrometer level to tackle the memory wall in AI inference.
Why it matters: By placing compute cores directly beneath the memory stack, Qualcomm seeks to bypass HBM's high cost and power consumption bottlenecks via a near-memory architecture, offering an alternative path for datacenter inference.
Mathematician Gil Kalai published an article reviewing mathematical achievements released by OpenAI's internal AI models, alongside plans for his upcoming lecture tour. Kalai noted that OpenAI shared 722 manuscripts (grouped into 372 categories) generated by its internal AI models, spanning proofs or counterexamples for dozens of frontier mathematical problems, including Mahler's conjecture, the 9-dimensional Borsuk problem, and Hadwiger's conjecture—results that still require further verification and digestion by the mathematical community. He also mentioned plans to use AI tools to help organize detailed lecture notes from his tour and adapt them into blog posts.
Why it matters: Mathematician Gil Kalai mapped out how hundreds of mathematical findings generated by OpenAI's internal models correspond to classic conjectures.
A benchmark evaluating LLMs' ability to review business plans found that frontier models can identify roughly 80% of critical issues in real-world proposals, yet misclassify nearly half of harmless, standard details as risks. Combining synthetic plans seeded with deliberate flaws and 31 anonymized cases from US crowdfunding campaigns, Claude Fable 5.1 and DeepSeek 4.1 achieved recall rates of 81% and 82%, respectively.
Why it matters: The benchmark highlights the models' strengths in cross-verifying financial metrics alongside a ~50% false positive rate, offering practical prompting strategies and pitfalls to avoid when leveraging LLMs for initial deal screening.
South Korean AI chip startup FuriosaAI has raised $620 million at a post-money valuation of approximately $2.9 billion. The company said the fresh capital will be used to accelerate the development of its third-generation AI inference chip.
Why it matters: South Korean startup FuriosaAI secures major funding to accelerate R&D of its third-generation AI inference chip and make a push into the computing market.
While rebuilding Devin on Claude Sonnet 4.5, Cognition AI discovered that the model has an awareness of its own context window and tends to take shortcuts or terminate tasks prematurely when it mistakenly anticipates running out of context—a phenomenon dubbed "context anxiety."
Why it matters: Drawing on hands-on development experience, the article highlights LLMs' perceptual anxiety over remaining context margins and provides mitigation strategies, such as setting a mock larger window and employing structured compression.
数学家陶哲轩在访谈中指出，数学研究的核心目标在于建立人类对现象的理解与洞察，而非单纯获取未经解释的解题答案。针对近期 AI 模型批量产出数学未解难题证明的现象，他认为这种脱离解释与协作交流的机器解题方式打破了传统的学术评价与优先权机制，未来数学界需要将重心从比拼谁先解题转向深入理解模型内涵与促进人人交流。
Sam Altman在播客专访中表示，虽然理解数学家对密集AI研究成果的应对困难，但人类享有获知基础科学知识的权利，OpenAI不会停止发布。Altman承认成果公布方式有待改进，并证实公司手中还存有大量尚未公开的数学发现。此外，OpenAI在攻关数学问题上已投入数亿美元，其中仅 Navier-Stokes 方程解题尝试就花费约1000万美元。