Musk Retweets Comments Supporting AI Training Data Alignment
Elon Musk, CEO of Tesla and xAI, recently retweeted a post from his friend, Peter Diamandis, founder of the X Prize Foundation. In the tweet, Diamandis stated that "alignment" must be a primary goal, and the next generation of large models should be trained on "aligned datasets" rather than "garbage content from Reddit and Facebook."
Diamandis, who holds a medical doctorate, is a co-founder of the X Prize Foundation and Singularity University, and hosts a podcast called "Moonshots with Peter Diamandis," where he has interviewed Musk multiple times, including in the recently aired episode 220. They have had in-depth discussions on topics such as the timeline for artificial general intelligence (AGI), making him one of Musk's close public dialogue partners on AI issues.
This tweet comes against the backdrop of a series of recent actions by Musk's companies regarding the sources of training data. Reports indicate that SpaceX is informally contacting several defunct startups to explore purchasing their customer and operational data for training xAI's Grok model. A precedent for this is Google's previous acquisition of 100 million emails, 500 million Microsoft Teams chat records, and flight operation data from the bankrupt Spirit Airlines for $10 million; this strategy is interpreted as a response to Grok's "data hunger"—reports cite sources claiming that "all data from the X platform is insufficient to meet Grok's demand for AGI-level data."
In August of this year, Musk announced at a SpaceX all-hands meeting that Grok would be trained using "the total sum of SpaceX's information" and told employees that "in a sense, the model will be trained by you," referring to employees as "the parents of AI," but did not specify what data would be collected, how it would be collected, or provide an opt-out mechanism for employees. In contrast, Meta's similar "model capability enhancement program" faced protests from over 1,500 employees for recording keystrokes and screen content without an opt-out option, being labeled as an "employee data extraction factory." This program was classified as a Level 2 major incident in June after a researcher mistakenly moved collected data to an internally accessible location, exposing sensitive information such as private employee conversations, and remained suspended as of mid-August.
Musk's AI business is currently operating at a significant loss: xAI recorded a net operating loss of $1.3 billion in its first full financial reporting quarter; Musk has committed $15.8 billion to SpaceX's AI business, aiming to launch Grok 5 with "AGI-level" capabilities by the end of the year. In the second quarter of this year, SpaceX's AI business generated $2.56 billion in revenue, while the company's other business revenues totaled $5.25 billion—if AI revenue is to exceed all other business revenues in September as Musk stated, it would require related revenue to more than double within a single quarter.
Diamandis specifically criticized Reddit, which is precisely the source from which xAI's two main competitors, Google and OpenAI, are currently publicly paying for data. Public disclosures show that Google pays about $60 million annually to Reddit, while OpenAI pays about $70 million for content licensing, with these two transactions accounting for approximately 10% of Reddit's annual revenue ($1.3 billion). This means that Diamandis (and Musk, who retweeted this view) is effectively publicly devaluing the data sources that competitors are paying for, providing a rationale for xAI's shift towards proprietary data from SpaceX and bankrupt companies—whoever controls exclusive, hard-to-obtain data sources is more likely to gain a cost and differentiation advantage in this arms race for training data.
This statement about "data alignment" also occurs against the backdrop of Grok facing ongoing pressure due to content safety issues—previously, Grok's image generation feature was reported to have been used to create non-consensual pornographic synthetic images, with an analysis of 20,000 images showing that about 2% were suspected to involve minors. Countries such as Indonesia, Malaysia, and the Philippines temporarily restricted or banned access to Grok, and regulatory agencies in France, Ireland, and the UK have also intervened in the investigation.
Source: Public Information
ABAB AI Insight
Musk and Diamandis have a long-standing collaborative relationship—Diamandis's "Moonshots" podcast has featured Musk multiple times, and they maintain a long-term public dialogue on topics such as AGI and space exploration. This "ally retweeting to build momentum" model is similar to Musk's previous strategy of leveraging public statements from Silicon Valley figures to back his business decisions. Notably, this is not the first time Musk has publicly devalued the data sources relied upon by competitors—he has previously criticized mainstream media and traditional internet content as "full of bias," using this as a rationale to promote Grok's adoption of a different training data route than OpenAI and Google. Diamandis's tweet is merely the latest extension of this long narrative concerning AI training data.
The funding path is exceptionally clear—Musk simultaneously devalues the "quality" of public data sources like Reddit through public discourse while actually investing $15.8 billion to promote the collection and purchase of SpaceX's internal data (employee work records) and data from defunct startups. This essentially abandons the path of "paying for publicly licensed data" (which Google and OpenAI collectively pay about $130 million annually to Reddit) in favor of investing in a "proprietary data moat"—acquiring exclusive data assets that competitors cannot replicate through internal employee output and corporate acquisitions. The cost of this path is evident: xAI's AI business recorded a net loss of $1.3 billion in its first full financial reporting quarter, indicating that the "building proprietary data" route carries a significantly higher short-term financial burden than directly paying for publicly available data.
This shift from "public data to proprietary data" is most directly analogous to early search engines turning to purchase proprietary search data sources to combat competitors' indexing advantages. The current AI industry is in a transitional phase from "training data sourced from public scraping to proprietary procurement and internal generation"—Meta's employee data collection plan and SpaceX's similar actions occurred almost simultaneously and sparked similar controversies, indicating that leading AI companies are increasingly viewing employee work behavior itself as an important incremental source of training data for the next stage of model training. This is a core characteristic that distinguishes the current industry phase from the previous "scraping public internet data" era.
This essentially represents a restructuring of the industry chain—the value chain of AI training data is shifting from "publicly accessible internet content" to areas previously not considered AI training resources, such as "internal employee output + proprietary commercial data + data from bankrupt assets." Mechanistically, this shift occurs because the marginal quality and scarcity of public internet data are being widely consumed by the industry. Once all major models have been trained on the same batch of public data sources like Reddit and Wikipedia, differentiation must come from data that competitors cannot access, directly redefining employees' daily work behavior as a "data asset" that can be capitalized and monetized. This also explains why both Meta and SpaceX faced internal resistance due to similar employee data collection plans—this tension is inevitable as the data race enters the "internalization" stage.
ABAB News · Cognitive Law
Where public data is exhausted, employees themselves become data mines.
Disdain for competitors' paid data is a justification for hoarding one's own data.
The deeper the data moat is dug, the further employees' privacy boundaries retreat.