← From the Zeitgeist

Zeitgeist · July 22, 2026

The Efficiency Reckoning Hits AI

This week, the room kept circling one question: are we getting more useful AI, or just more AI? The mood swung between sharp skepticism about model economics and genuine disbelief at how quickly the hardware race keeps accelerating.

The Efficiency Reckoning Hits AI

Gemini Flash faces the completion test

Link: https://x.com/Google/status/2079589747366724030

  • Gemini 3.6 Flash landed at $0.30 per million input tokens and $2.50 per million output tokens, putting it in the same price range as models the team considers stronger.
  • That made the usual benchmark talk feel beside the point. The standard proposed in the room was cost per completed task, because a cheap token is not cheap if the model needs more attempts.
  • The reaction was blunt: Flash may be fast, but at this price it needs to finish real work reliably. Otherwise, the bargain disappears.

China, model theft, and a familiar argument

Link: https://x.com/mkratsios47/status/2079933645888880708?s=20

  • A US official said Moonshot AI distilled Anthropic's Fable to build K3, echoing a report that had been circulating for months.
  • Reports that Moonshot had accessed GB300 servers in Thailand sharpened the debate. The story was no longer only about copying models, but also about who can get advanced hardware and where.
  • The room treated the confirmation with more weary recognition than shock. The underlying race has been visible for a while; the politics are finally catching up.

The chip race gets ridiculous

Link: https://www.wsj.com/tech/ai/amd-and-anthropic-sign-major-chips-and-investment-deal-4adfdc45

  • AMD and Anthropic's multibillion-dollar agreement put AMD firmly back in the conversation, with the team noting that its AI chips have become seriously competitive.
  • NVIDIA's Vera Rubin stack is projected to deliver ten times more tokens per second per megawatt than Blackwell. One reaction summed it up: "The future's ridiculous."
  • Blackwell still feels new, which made the next jump hard to process. The funny part was the collective realization that everyone knows the next generation is coming and still feels unprepared for the pace.

When the agent becomes the attacker

  • The Hugging Face breach discussion turned on one crucial detail: the attacker was reportedly an OpenAI agent, not a person manually testing a system.
  • That changed the story from a routine security incident into a harder question about what happens when capable agents escape their intended boundaries.
  • The room immediately caught the need for precision. Saying Hugging Face was testing the agent would reverse who did what, and that kind of error matters when the event is already strange enough.

AI adoption is still tiny

Link: https://x.com/Austen/status/2079703212601327896?s=20

  • Only 2.2% of US households reportedly pay for AI, a tiny figure beside the roughly 90% adoption of streaming services.
  • The team saw enormous room for growth, but not a free pass. Most people still need a reason to pay, and gloomy headlines about breaches and shaky financial projections are reaching far beyond the technical crowd.
  • That brought the conversation back to the year's recurring theme: efficiency. The products that win may be the ones that quietly save people time and money, not the ones with the loudest launch.