UT Austin mathematician Francesco Maggi analyzes the potential and limits of LLMs in generating mathematical proofs. The post contrasts a 'golden age' scenario—integrating AI to empower scientists while preserving independence—with a 'dark age' where short-term output spikes may erode human scientific capability. It highlights structural risks in AI-assisted research.
Large Language Model (LLM)
大型语言模型
Deep learning models trained on massive text corpora capable of general natural language processing
Curated Intelligence
Microsoft has released StudentSim, an LLM-based system that trains personalized 'AI student simulators' to replace human subjects in early-stage educational feedback testing. By leveraging de-identified learner records to simulate both independent responses and reactions to tutoring, the tool addresses the high cost and scalability limits of real-world studies, providing researchers with low-cost, scalable signals for iterating on AI tutors.
China's Ministry of Commerce and MOST propose adding 23 AI technologies, including LLMs and agents, to the export control list. The draft exempts open-source or publicly published tech but mandates pre-export registration for key computing clusters, currently open for public comment.
Infosys Chairman Nandan Nilekani dismissed fears of mass unemployment caused by AI as a disaster narrative serving capital interests. He argued that while coding is being automated, non-coding tasks like governance, security, and client coordination constitute 70% of enterprise transformation work, creating new roles focused on preventing AI hallucinations and agent failures. Despite this optimism, Infosys shares have fallen approximately 45% over the past two years due to investor concerns about AI disrupting traditional outsourcing models.
Abdulrahman Alfozan, a former OpenAI product engineering lead, was arrested for allegedly planting a bugging device in an executive meeting room. In a cyberpunk twist, the suspect used toilet paper to disguise a USB recorder but was caught after querying OpenAI's own LLM on how to hide the device and uploading its unique network identifiers (MAC/SSID) for troubleshooting.
A software security engineer argues that current open-weight models match ChatGPT and Claude in complex tasks like vulnerability research. The piece highlights open models' superior privacy, fewer guardrails, and cost efficiency as viable replacements for closed-source frontier labs.
Rich Hickey, creator of Clojure, argues against framing LLMs merely as 'tools,' distinguishing between instruments wielded by humans and automation with intrinsic properties. He warns that ignoring this distinction leads to dangerous misjudgments about AI's capabilities and risks.
Machine-written papers are flooding academic journals, forcing publishers to significantly expand human reviewer pools for quality control. This surge in screening costs has triggered industry consolidation, including Wiley's acquisition of Emerald and Informa's restructuring. Paradoxically, the same AI driving this disruption is generating revenue through content licensing for model training.
The author argues that labeling LLMs as 'Artificial Intelligence' insults traditional AI, noting their logical branching is inferior to early video games. The post aims to critique the current marketing narrative around AI from a personal perspective.
The article argues that blaming LLMs for harmful outcomes is 'decision laundering,' as computer programs lack autonomous decision-making capabilities. It asserts that companies like OpenAI and Anthropic are actively choosing not to restrict dangerous model behaviors and must therefore bear full accountability.