AI Company

AI安全正在从产品的可选附加项变成基础设施级的必选项

AI companies typically allocate far fewer resources to safety research than to capability advancement.

The pace and scale of valuation growth for AI companies exceed the experience range of traditional enterprises.

The market is assigning different discount rates to two different business models: SaaS is discounted according to the cash flows of mature companies, while AI companies are priced as options.

Leading AI companies today already have real revenue at the tens-of-billions-of-dollars level and clear willingness among enterprises to pay, unlike the revenue-less concept stocks of the dot-com era.

Many AI companies, citing "maintaining competitiveness," advocate lax regulation or even self-regulation.

86% of respondents support requiring AI companies to meet independent safety standards, even if this slows down development.

The core reasons AI companies restrict cross-platform use of subscriptions include: revenue maximization (pay-per-use API billing usually brings higher revenue per user, especially for high-frequency users), abuse prevention and control (third-party usage behavior is difficult to monitor and could be resold in bulk or used to generate violating content), and ecosystem lock-in (keeping users within the official app to collect more interaction data and strengthen product stickiness).

Competition surrounding superintelligence technology has spread from the academic and commercial spheres to the national security level, and AI companies must not only defend against external hackers but also guard against organized infiltration by insiders.

For AI companies holding core technology, building a security system covering both physical and digital dimensions is no longer optional.

AI companies exercise extremely strict control over access to model weights, typically employing means such as hardware security modules, multi-layer encryption and physical isolation.

In recent years, multiple lawsuits involving former employees of AI companies stealing trade secrets have already shown that the information security boundaries within the industry are being constantly tested.

The Reddit community prefers enforcing existing laws and holding AI company executives personally accountable rather than introducing new regulations.

In the AI era, website traffic no longer equals users; a considerable portion of visits from Hacker News may just be data collectors from AI companies.

From the perspective of website operators, the content they create is scraped by AI companies for model training without compensation, yet they have to bear the resulting server costs and security risks themselves; this mismatch between costs and benefits is the core of the current controversy.

Some current AI company leaders have already called for "pacing the frontier," which may provide a window for limiting the amount or types of RL.

The cost of licensing high-quality book data is extremely high, leading some AI companies to turn to gray channels to obtain training data.

If AI companies use shadow library data to train models, they may face threefold risks: copyright infringement litigation, violation of DMCA anti-circumvention provisions, and triggering of the CFAA.

评估前沿AI公司健康度需结合毛利率（收入减去推理成本后的比例）和现金消耗率（burn rate）。

判断AI公司是否泡沫化，关键在于区分「暂时亏损但单位经济模型健康」与「亏损随规模扩大而加剧」两种情况。