Financial Journalist Iain Martin: AI Labs are Spending Hundreds of Thousands to Acquire Internal Communications of 'Dead Startups' as Reinforcement Learning Training Grounds
Financial and technology journalist Iain Martin noted in public comments that AI labs are paying hundreds of thousands of dollars to "failed startups" and their liquidation agencies to acquire their email, Slack, and Jira conversation records, which are used as training data for "reinforcement learning gyms"—training agent-based AI through simulated workplaces, decision processes, and real work scenarios.
English media and the startup community report that this type of "dead data" is seen as "operational residue" and "behavioral gold," as it contains a wealth of unfiltered meeting disagreements, internal gossip, cross-department conflicts, and real collaboration patterns. AI developers believe this data can train agents to understand workplace unspoken rules better than purely synthetic environments. Service agencies have emerged that specialize in packaging and selling the communication assets of old startups, selling them to AI labs categorized by "company worlds" (such as "Finance World," "Tax World," "Big Tech World").
This trend resonates with the backdrop of the rapid evolution of "AI agents in simulated work environments": Institutions like Anthropic are reported to plan to invest billions in "reinforcement learning gyms" to train AI to make autonomous decisions in emails, meetings, and task systems through highly realistic digital office scenarios, with "real dead company data" becoming a key bridge connecting the "real world" and the "algorithmic world."
Source: Public Information
ABAB AI Insight
This essentially represents the systematic recycling of "human organizational remnants" into AI capital. Dead companies are no longer just financially bankrupt; their internal communications have become a "behavioral data mine," re-monetized by AI labs in asset packages. This trading structure indicates that AI capital is no longer merely about "buying content" or "buying data," but about "buying organizational memory," transforming the failures and internal operating models of others into a "repository of experiences" for their own models.
From an institutional perspective, this highlights the fracture between "data ownership" and "privacy rights." Startup teams rarely pre-specify "who owns the communications after the company dies"; this data is often implicitly considered part of the "asset pool," disposed of by liquidators or creditors, while original employees seldom have substantial say in the "secondary sale of data." When AI labs purchase, they are more often transacting with the "asset holders" rather than the "actual data owners," which disperses and obscures the externalities and responsibilities of data use.
On a deeper level, this model is driving the standardization of "AI workplace personas." As AI agents are repeatedly trained in multiple "data dormitories" like "Big Tech World," "Finance World," and "Tax World," their behavioral norms will increasingly be determined by data from real work scenarios rather than manually coded rules by developers. While this increases the "realism" of the agents, it also means that the "unspoken rules of the workplace" across different industries will be compressed and distilled into purchasable "behavioral priors." In the future, a company's "AI culture" competition may no longer be about "rule documents," but rather about the "quality and diversity of data assets."