中国电信领投，觅蜂科技再获数亿元融资，聚焦物理AI数据服务平台

TL;DR - An AI-assisted, Lean-verified proof reportedly resolves the 70-year-old Sendov conjecture. Terence Tao then simplified the formalization and found that it also proves the stronger Phelps–Rodriguez conjecture.

• Lech Mazur developed the proof with GPT-5.6 Pro assistance and approximately 90,000 lines of Lean 4 code.
• Tao reorganized and re-formalized it in roughly 15,000 Lean lines, exposing a surprisingly elementary argument based on algebraic identities and inequalities.
• The proof combines analytic arguments for broad degree ranges with Lean-checked Bernstein polynomial certificates for degrees 5–100.
• Its boundary-case analysis characterizes equality, yielding the stronger strict-distance result except for the known extremal polynomial family.

TL;DR - Apple is reportedly pursuing a jointly trained China-focused LLM with Alibaba while retaining third-party models. The strategy aims to preserve control over system-level agents, privacy, user experience, and the commercial value of Apple’s platform.

• Apple could use its own model for intent understanding, task planning, permissions, and cross-app orchestration, while outsourcing general knowledge and generation.
• Mature training methods, open-source ecosystems, and efficient MoE architectures have made usable foundation models faster and less costly to build.
• A proprietary agent would keep sensitive data and privileged device operations within Apple-controlled boundaries.
• Controlling the primary AI interface is strategically important as agents increasingly bypass apps, search, advertising, and other traditional platform entry points.

TL;DR - Physical-AI data platform Maniformer raised several hundred million yuan to scale its MEgo data-collection hardware and end-to-end collection, governance, and evaluation infrastructure. The funding targets tens of millions of hours of real-world interaction data for robotics and embodied-AI models.

• MEgo devices collect synchronized environmental and hand-operation data with millimeter-level trajectory accuracy across factories, homes, logistics, and other real-world settings.
• MEgo Engine automates preprocessing, spatial reconstruction, multimodal annotation, and quality evaluation, reportedly improving processing efficiency by more than 10×.
• Maniformer says each data-feedback cycle improved task success rates by 5–10% in some scenarios.
• China Telecom will support integrated cloud-network and embodied-data services, while Zhangjiang Group will provide industrial and commercial deployment environments.

TL;DR - Fei-Fei Li argues that AI should augment rather than replace people, with human oversight, broad public participation, and domain-specific safeguards guiding adoption. The interview matters because it connects AI’s technical limitations—statistical learning, data dependence, and lack of subjective experience—to practical governance and deployment choices.

• AI still learns largely from statistical patterns in massive datasets, unlike humans’ data-efficient, contextual learning.
• Healthcare illustrates the best near-term model: AI can support diagnosis, research, and robotic surgery, but sparse data and patient variability limit autonomous operation.
• AI cannot genuinely experience empathy, emotion, or private subjective memories; human creativity and judgment therefore remain central.
• Li advocates participatory governance, ethics education, and strong safety oversight for high-risk applications rather than leaving decisions solely to technology companies.

TL;DR - Fields Medalist Timothy Gowers argues that LLMs currently excel at mathematical breakthroughs framed as finding counterexamples or constructing rare objects. Their broad knowledge and low-cost search enable cross-domain experimentation, but they still lack mathematicians’ intuition for prioritizing promising paths.

• Recent advances cited include counterexamples related to the Jacobian and Erdős unit-distance conjectures, plus constructions for non-sofic groups and multicolor Ramsey bounds.
• LLMs can cheaply explore large search spaces, recombine techniques, and import tools from distant fields such as algebraic number theory.
• Current systems often pursue plausible but unproductive approaches and repeatedly reformulate problems without materially nearing solutions.
• Gowers views genuinely novel, reusable mathematical methods—not merely successful constructions—as a stronger test of top-tier AI creativity.

TL;DR - An AI-assisted, Lean-verified proof reportedly resolves the 70-year-old Sendov conjecture. Terence Tao simplified the formalization and found that it also proves the stronger Phelps–Rodriguez conjecture.

• Lech Mazur’s proof used GPT-5.6 Pro and roughly 90,000 lines of Lean 4 code.
• Tao reorganized the argument into about 15,000 Lean lines and identified stronger implications.
• The proof reduces the problem to elementary identities and inequalities involving roots and critical points.
• Degrees 5–100 use exact Bernstein-polynomial certificates checked by Lean; higher degrees are handled analytically.