Friday — August 07, 2026

The race for training data has hit a new low: the physical destruction of rare books. According to multiple sources, AI labs are purchasing entire collections of rare, out-of-print, and physically fragile books, only to slice off the spines and run them through high-speed scanners that effectively destroy the binding. The rationale is simple—these texts contain unique information not indexed on the open internet, and they represent a 'gold mine' for training high-quality reasoning models.

The era of astronomical AI pricing is officially over. OpenAI's decision to slash GPT-5.6 prices comes after months of pressure from CFOs and CTOs who realized that integrating AI agents into their workflows was costing more than hiring a junior developer. The cuts apply to both input and output tokens, making it significantly cheaper to run high-volume tasks like customer support or code generation. This is likely a direct response to the rise of cheaper open-source models and the 'efficiency' trend in the industry. For startups, this is a double-edged sword: it lowers the barrier to entry, but it also means the incumbents are willing to fight a price war to maintain market share.

Mozilla's report attempts to define what 'open source AI' actually means, highlighting that many models claiming to be open are merely 'weight accessible' without providing the full training data or code. Simultaneously, a viral Hacker News post argues that the safety arguments against open-source AI are 'bad' because they rely on hypothetical risks rather than concrete harms. Yet, the biggest irony is that AI's top startups are increasingly silent. They are not publishing research papers, not releasing model weights, and not sharing benchmarks. This creates a 'transparency paradox' where the loudest advocates for open source are the ones with nothing to hide, while the biggest players are building walls.

In a move that has sent shockwaves through the AI community, the former CTO of OpenAI has launched a new frontier model that reportedly matches or exceeds the capabilities of GPT-5.6 in several benchmarks. The release skips the typical red-team and staged rollout that has become standard practice at major labs. This is a significant departure from the norm, suggesting that the 'responsible AI' framework is fracturing. Insiders suggest this is a response to the perceived stagnation at OpenAI, where internal politics and safety committees are slowing down innovation. The model is already available via API, and early adopters report exceptional reasoning capabilities, though safety researchers are raising alarms about the lack of guardrails.