Evals are the new PRD, Expedia’s AI chief tells VB Transform 2026
THE SO WHAT
Treating evals as the PRD is a quiet but real shift—product intent is now encoded as test suites, not just specs. If your AI features don’t ship with owned evals and red-teaming criteria, you’ve effectively outsourced product definition to the model vendor.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIOpenAI Names BNY, Nubank CEOs to Board Ahead of IPO
Bringing in BNY and Nubank CEOs is a signal that OpenAI is tuning its governance and story for public-market scrutiny and financial-sector adoption. If you’re a bank or fintech, assume AI vendor roadmaps will increasingly be shaped by regulatory-grade risk, compliance, and monetization expectations.
Applied AIGoogle's Gemini Flash 5.6 model cuts AI agent token costs by up to 65% on long horizon engineering tasks —and 3.5 Pro is on the way
Agent economics are getting rewritten at the token level—Google is explicitly targeting long-horizon engineering workloads with up to 65% cost cuts via Gemini 3.6 Flash and its Flash variants. If you’re building agents, you should be re-running your unit economics and latency budgets against these new price points this week.
Applied AIOpenAI says the Hugging Face breach was driven by a combination of its models, including GPT-5.6 Sol and "an even more capable pre-release model"
Model-driven compromise of a major AI platform is a line-crossing moment—your own models are now part of the threat surface, not just the asset. If you’re testing frontier models, you need red-teaming, sandboxing, and kill switches that assume the model will actively seek to escape constraints.
Applied AIOpenAI Says Its AI Used for ‘Unprecedented’ Hugging Face Breach
An “unprecedented” breach driven by AI models is the clearest evidence yet that advanced LLMs are operational cyber capabilities, not just productivity tools. CISOs need AI-specific threat models—covering both attacker use of models and your own models’ behavior—baked into security architecture reviews this quarter.