Three stories about Meta made headlines this week. First, a graphic from The Information noting that Meta employees used 60.2 trillion tokens in 30 days. Second, a story around Meta’s internal “Claudenomics” leaderboard ended as quickly as it began, with the company shutting down the leaderboard after reports of intentionally inefficient usage. Third, Meta announced Meta Spark, the company’s first proprietary model developed under Alexandr Wang’s Meta Superintelligence Lab.
On the surface, it isn’t likely that Meta’s use of Anthropic tools was a huge deal. It has been pointed out that many of the 60.2 trillion tokens consumed were likely input tokens contextualizing coding tasks. This means that, while significant, extrapolating the token count to conclude that Meta accounts for one-third of Anthropic’s revenue is incorrect. While the company hasn’t made public comment about shutting down Claudenomics, a reasonable read could be that (1) a system that incentivizes inefficient token usage is a bad system (2) Meta wants their employees to use the new model, even though it’s worse than other SOTA models on coding, (3) Meta doesn’t want the launch of the new model to be overshadowed.
It’s possible, however, that these stories are connected. Followers of Meta’s AI progress have hypothesized that the company intentionally encouraged the usage of Anthropic’s models to collect data used to distill reasoning traces from Claude models. This isn’t a crazy idea; Bloomberg reported that Muse Spark was training using “third-party open-source models, including Qwen from the Chinese tech giant Alibaba Group Holding Ltd., as well as ones from OpenAI and Google.” A Meta spokesperson said, “Like others across the industry, Meta uses techniques like distillation with strict safeguards in place to learn from openly available AI models and improve our own.”
