The AI revolution is optimizing for the wrong things
THE SO WHAT
If frontier models are getting better at code and worse at writing, that's a misalignment between benchmark incentives and real-world creative work. Leaders deploying LLMs should define their own task-specific evals—especially for language quality—rather than trusting generic "smartest model" rankings.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAnthropic strikes $45bn deal with Nscale to rent compute capacity
$45B of rented capacity locks in that frontier model economics are now dominated by long-term compute offtakes, not just cloud list prices. If you’re building on top of these models, assume your upstream vendors are capital-constrained by power and datacenter buildouts, not by demand.
Applied AIGemini Omni 1.1 Flash lets you build with more control
More control knobs on Gemini Omni 1.1 Flash mean the frontier models are moving from generic assistants toward tunable components in your stack. Teams building on Google should revisit where they still rely on brittle prompt hacks—some of that can likely be replaced with first-class controls and policies.
Applied AIGoogle’s AI Mode can now track flight prices, help book hotels, and more
Google turning AI Mode into a travel agent that tracks prices and books hotels is another step toward end-to-end transactional agents inside search. Travel, fintech, and marketplace operators should assume more high-intent flows will be intermediated by assistants and start designing for agent-to-API integrations, not just human UX.
Applied AINvidia Sees AI-Fueled Demand Boosting Sales 70% Next Year
A projected 70% revenue jump next fiscal year suggests AI capex is still in the build-out phase, not the digestion phase. If you’re planning GPU-dependent products, assume continued scarcity and pricing power—optimize for efficiency and multi-vendor strategies now.