Reading

Zhidx reported on November 5 that Shanghai-based video generation foundation model startup Sand.ai has completed a $200 million (approx. 1.334 billion yuan) A+ funding round. This is the company's third financing event in 2026, bringing its cumulative fundraising to $300 million (approx. 2.001 billion yuan) over 10 months. While Sand.ai did not disclose its post-money valuation, the company stated it has officially joined the ranks of unicorns following this round.

The round was co-led by Cornerstone Capital and Wangsu Science & Technology. Other participants include Era Capital, Panlin Capital, Dongzheng Capital, Tianchuang Capital (along with Shandong Railway Industrial Investment and Shandong Railway Fund), as well as Shanghai state-backed entities like Pudong Venture Capital and Shanghai Guofang. Existing investors such as Matrix China, Huaye Tiancheng, and CAS Star continued to add capital. Sand.ai noted that the funds will primarily be used for video foundation model R&D, training and inference infrastructure, product iteration, and global market expansion.

Founded in January 2024 and headquartered in Shanghai, Sand.ai focuses on video generation foundation models and application products. Its founder and CEO, Cao Yue, holds a Ph.D. in Software Engineering from Tsinghua University and previously served as a principal researcher at Microsoft Research Asia and head of the Visual Model Research Center at Beijing Academy of Artificial Intelligence (BAAI). He was also a co-founder of Light Year, an AI large model company acquired by Meituan. Cao contributed to visual model research including Swin Transformer, which won the ICCV 2021 Marr Prize.

Technically, Sand.ai has evolved its architecture from autoregressive generation and joint audio-video modeling to hundred-billion-parameter MoE models. In August 2026, the company released and open-sourced MAGI-2 Preview, becoming the first team globally to scale a video generation model to 100 billion parameters while releasing model weights. The model uses an ultra-fine-grained MoE architecture with approximately 114B total parameters, activating only about 6B per forward pass. Additionally, Sand.ai's video generation product VidMuse surpassed $10 million in Annual Recurring Revenue (ARR) two months after launch.

Following this funding, Sand.ai plans to continue advancing next-generation large-scale video models, accelerating the deployment of open-source foundation models and applications, and expanding its global developer ecosystem. Amid intensifying competition in video generation, the company's ability to convert model capabilities into stable product revenue and overseas market share remains a key focus for the next phase.

In its latest quarterly earnings call, HubSpot CEO Yamini Rangan articulated a clear strategic trade-off: eliminating pricing friction, packaging friction, and AI adoption friction for customers, even at the cost of extended buying cycles and near-term revenue pressure. Her core argument is that enterprise customers are evaluating AI differently—they demand to see it operate effectively with their own data and in their specific environments, seeking proof before purchase. HubSpot has chosen to align with this shift by introducing free trials and pricing tied to value and outcomes, rather than attempting to lock customers into traditional subscription models.

This assessment is grounded in observations of changing enterprise AI procurement behaviors. Rangan noted that AI decisions are no longer purely technical choices but strategic ones, resulting in larger procurement committees involving CEOs, CTOs, and board members. HubSpot’s response involves engaging senior decision-makers earlier in the sales cycle, bringing in multiple stakeholders, and providing technical expertise. A key metric reported was that in deals with five or more buying committee members, the win rate nearly tripled when HubSpot proactively engaged multiple stakeholders; deals involving C-suite executives also demonstrated higher close rates.

However, trial growth alone does not equate to revenue conversion. Rangan acknowledged that while agentic trials grew 64% over the past three months, customers require hands-on operational support to translate trials into tangible results. Consequently, HubSpot’s strategy is bifurcated across two market segments: the upmarket (enterprise), which relies on partners and solution consultants for manual guidance—checking in a week after trial initiation to confirm clear outcomes and follow up—and the downmarket (SMBs), where significant resources are invested in simplifying the product experience. For SMBs, agentic onboarding allows customers to simply provide an email address to access their calendar and see results quickly, without relying on human intervention.

Regarding pricing models, HubSpot is evolving from Phase One to Phase Two. In 2024, Phase One lowered seat pricing, removed minimum seat requirements for Sales Hub and Service Hub, and introduced core seats. While initial average selling prices (ASPs) declined, this attracted more customers and improved retention. The current Phase Two aims for a hybrid model of seats and credits: seats provide platform AI and data access, while credits scale with AI value and customer outcomes. Specifically, pilots in the Nordics and Benelux have eliminated separate Hub subscriptions, reduced the price jump from Starter to Pro tiers, and unified all usage-based features (email sends, ticket resolution, data enrichment, API calls) into credits. Results show declining ASPs but increased Pro+ customer volume and credit consumption reaching two to three times pre-pilot levels.

This pilot will expand to all new customers across EMEA in Q4, with results determining the pace of global rollout. Rangan summarized that HubSpot’s goal is to make the platform easier to buy and AI easier to adopt, believing now is the time to remove friction, accelerate AI adoption, and capture market share early in the AI platform shift. She explicitly stated this is a deliberate trade-off: accepting longer near-term buying cycles in exchange for stronger long-term growth and profitability.

This paper, 'Towards Deep Learning Models Resistant to Adversarial Attacks,' published in 2018 by Aleksander Madry and colleagues at the Massachusetts Institute of Technology (MIT), delves into enhancing neural network security through robust optimization. The study highlights that deep neural networks have inherent weaknesses, making them highly susceptible to adversarial attacks—inputs that are nearly indistinguishable from original data but cause the model to misclassify. This vulnerability is particularly dangerous in safety-critical systems such as autonomous driving and facial recognition.

To address this issue, the authors propose a unified framework based on saddle-point (min-max) problems. This framework formalizes adversarial training as an optimization problem: the inner layer maximizes to find the adversarial perturbation that maximizes the loss function, while the outer layer minimizes by adjusting model parameters to reduce risk under this worst-case scenario. This approach not only encompasses previous attack and defense mechanisms but also provides a clear objective for safety guarantees, specifically resisting all attackers bounded by certain norms (such as the ℓ∞ norm).

In terms of implementation, the paper proves that Projected Gradient Descent (PGD) serves as a reliable 'first-order adversary' capable of effectively solving the aforementioned optimization problem. Experimental results show that models trained with PGD achieve over 89% accuracy against the strongest white-box attacks on the MNIST dataset and 46% on CIFAR-10. This indicates that by increasing model capacity and adopting correct optimization strategies, the adversarial robustness of deep learning models can be significantly improved, bringing them closer to the security standards required for practical deployment.

For researchers and engineers interested in AI safety, machine learning theory, or those needing to build high-robustness visual systems, this classic 2018 paper, 'Towards Deep Learning Models Resistant to Adversarial Attacks,' provides the foundational framework for understanding the nature of adversarial attacks and designing effective defense mechanisms. It is recommended to start reading from Section 2, 'An Optimization View on Adversarial Robustness,' to grasp the core methodology.

In the 2024 shareholder letter released by Berkshire Hathaway Inc. (United States), Warren Buffett reiterated his core communication philosophy: treating shareholders as partners. He wrote, "Our goal is to communicate with you in a manner that we would like you to use if roles were reversed—that is, if you were CEO of Berkshire and my family and I were passive investors, we would be entrusting our savings to you." This "owner mindset" is reflected not only in performance disclosures but also in the public acknowledgment of mistakes. Buffett noted that he has made errors in assessing a company's future economic prospects or management capabilities, and cited Charlie Munger to emphasize that the greatest sin is delaying the correction of errors, so-called "thumb-sucking." He criticized the boardroom culture of large listed companies that prohibit mentioning "mistakes" or "errors," arguing that this implies a false managerial perfectionism.

The letter illustrates Berkshire's unique acquisition and management philosophy through a specific case: Pete Liegl and his company, Forest River. When Buffett considered acquiring this Indiana-based recreational vehicle manufacturer in 2005, Liegl proposed an unconventional compensation plan. Liegl stated, "I looked at Berkshire's proxy statement; I don't want to earn more than my boss, so please pay me $100,000 a year." Additionally, he requested an annual bonus equal to 10% of earnings exceeding current levels. Buffett accepted this proposal, even though the parties did not explicitly define vague terms such as "appropriate" or "material." Buffett recalled, "These vague terms never caused any problems." This case demonstrates that simple incentive mechanisms are often more effective than complex contractual clauses based on high trust, and it reflects Buffett's insight into the nature of talent—he explicitly stated that he never looks at candidates' alma maters because business acumen is often innate rather than cultivated solely through elite education.

For modern readers, understanding this governance structure based on trust and long-termism is crucial. Suppose the founder of a startup faces acquisition by a large conglomerate; if the acquirer insists on cumbersome valuation adjustment mechanisms (VAMs) and strict KPI assessments, it may stifle the founder's innovative drive. Conversely, if the acquirer grants full autonomy and simple profit-sharing incentives, as Berkshire did with Liegl, it may better stimulate managerial loyalty and potential. This contrast reveals that the "human" factor in capital allocation is often more decisive than pure financial models.

The Berkshire Hathaway Shareholder Letter, written by Warren Buffett, was published in 2024. It is suitable for readers interested in corporate governance, M&A integration, and long-term investment philosophy. It is recommended to start with the chapter reflecting on "mistakes" and then delve into the case of Pete Liegl to appreciate Berkshire's unique corporate culture.