Sunday, August 23, 2026

—

1. Harvard's $699 Startup Bootcamp Debuts AI Avatars of Instructors for Pitch Practice
   T2TechCrunch：AI（RSS）

   Harvard Business School's HBS Foundry startup bootcamp, priced at $699, now features AI avatars of instructors that provide real-time feedback during practice pitches and mock board meetings. The initiative aims to lower barriers to entrepreneurial coaching, enabling participants to access personalized guidance anytime and improve their business communication and decision-making through repeated rehearsal. The avatars are built on instructors' voices and styles, but do not replace the depth of human mentorship.

2. DeepMind-Alumni Founded Inherent Releases AI Agent Faraday, Claiming to Outperform Anthropic and OpenAI at Research Replication
   T2TechCrunch：AI（RSS）

   Inherent, a British AI lab founded by alumni of DeepMind, has unveiled Faraday, an AI agent designed to replicate experiments from scientific papers. The company claims that Faraday outperforms models from Anthropic and OpenAI in research replication benchmarks. If the claims hold up, Faraday could become a crucial tool for scientific discovery, enabling researchers to rapidly validate published results and explore new hypotheses. It represents a shift from conversational AI to action-oriented agents that directly interact with scientific workflows. However, these results are currently based on the company's internal evaluations, and external validation is still pending. The broader research community will need to independently assess Faraday's capabilities before drawing firm conclusions. This announcement highlights the growing competition in the AI agent space, particularly among labs founded by former DeepMind researchers.

3. Claude Code v2.1.241 Released: Bug Fixes and Reliability Improvements
   T1.5Claude Code：GitHub Releases（RSS）

   Anthropic has released version v2.1.241 of Claude Code, its AI-powered coding assistant designed for terminal environments. This patch release focuses on bug fixes and reliability improvements, addressing issues that could affect daily development workflows. While the official release notes do not provide granular details about specific fixes, iterative maintenance like this is essential for ensuring the tool's long-term stability and user trust. For developers and engineering teams that rely on Claude Code for code generation, debugging, and automation tasks, upgrading to this version can reduce interruptions and deliver a smoother, more predictable experience. The update is available immediately through the official GitHub Releases page. Given the fast-moving nature of AI development tools, staying up to date with the latest patches is recommended to benefit from ongoing engineering refinements. This release also underscores Anthropic's commitment to continuously improving its developer-facing products based on real-world usage and feedback.

1. OpenAI says California should strengthen its AI safety bill
   T2TechCrunch：AI（RSS）

   OpenAI, one of the leading AI companies, has reversed its earlier position and is now calling on California lawmakers to strengthen SB 53, a proposed AI safety bill. In a letter, OpenAI argues that while the bill has improved, it still requires additional provisions to address risks from frontier AI systems. The company previously opposed the legislation but now sees a need for clear regulatory frameworks. This change signals a broader industry shift as AI developers increasingly accept oversight. California's decisions could shape national AI policy, making this move significant. The move comes amid ongoing debates over AI regulation and safety.

2. Frontier AI Labs Still Won't Say How They'd Contain a Rogue Model
   T2TechCrunch：AI（RSS）

   A new study finds that leading AI labs have few publicly documented plans for containing rogue models. As AI systems increasingly demonstrate unexpected and potentially dangerous behaviors, questions are mounting over whether developers are truly prepared to isolate or shut down a model that goes off course. The study urges frontier labs to disclose more of their containment protocols, warning that a lack of transparency could leave the industry scrambling in a crisis. It also points to the gap between public safety commitments and concrete operational measures, calling for clearer accountability and pre-defined playbooks for worst-case scenarios.

3. Is It Legal to Train AI on Copyrighted Books? The Complex Legal Landscape
   T2TechCrunch：AI（RSS）

   Most published authors have, without their knowledge or consent, contributed to the development of AI tools that may threaten their livelihoods. On the surface, this seems illegal, but the legal reality is far more complex. This article explores whether training AI models on copyrighted books constitutes fair use or infringement, drawing on recent court cases and expert opinions. It also discusses how the uncertainty could reshape data sourcing strategies for AI companies and the economic outlook for writers and publishers. With ongoing litigation and potential new regulations, the outcome of this debate will have far-reaching implications for the entire AI ecosystem.

—

—