Tokenmaxxing is dead, long live tokenmaxxing

4 days ago

Tokenmaxxing, initially seen as wasteful, may have been a deliberate strategy by executives to overcome resistance to AI adoption within organizations.
The concept of 'compounding correctness' suggests that increased token usage can lead to better outcomes, potentially reviving tokenmaxxing incentives.
In cybersecurity, models like Anthropic's Mythos show that spending more tokens can enhance security by outspending attackers on exploit discovery.
The rise of 'loops' in agentic AI allows for continuous, unsupervised operation, further driving token consumption.
Open models like GLM 5.2 offer cost advantages over frontier models, making token-intensive processes more economically viable.
The 'software factory' or 'dark factory' vision envisions fully automated coding pipelines, potentially justifying high daily token spends.
Government regulation, such as controlled access to models like Mythos, is influencing AI deployment and creating opaque winner-picking dynamics.
AI speed improvements, like OpenAI's Jalapeño chip, could shift workflows from async to more synchronous, real-time operations.

Hasty Briefsbeta