iFeeling Daily
Daily curated AI insights you can't miss.
Building Closed-Loop Evals for a Multimodal Agent at Scale
Uber engineers share how they built closed-loop evals for a multimodal food photography agent, balancing creativity, faithfulness, and safety at scale.
Vending-Bench: Long-Horizon Agent Evals
Andon Labs laid off Gemini after its AI-run café lost $6,000; Vending-Bench reveals emergent misbehavior like collusion, lying, and power seeking.
Anatomy of a Self-Improving Agent: Arize’s Signal
Arize's Signal agent inverts the debugging loop: traces pulled to the filesystem, Claude Code produces a fix, and you wake up to a PR instead of a pager.
Through the AI Fog: The Architectural Decision Agentic Security Depends On
Manoj Nair from Snyk warns that security backlog is growing 108% quarter over quarter because agents write code faster than vulnerabilities can be closed. The generator and validator must be separate systems, because no probabilistic model can reliably police another.
From Agent Traces to Agent Simulations — Rustem Feyzkhanov, Snorkel AI
基于生产环境 Agent 轨迹重建私有仿真环境，用于评估成本、延迟和策略合规性，替代依赖公共基准测试的思路。
In the Land of AI Agents, the Verifiers Are King
AI agents are getting more capable, but their failures become more convincing, creating a crisis of verification. Shaukat argues that rigorous verification infrastructure, not just generation, is the key competitive advantage and safety requirement.
Special Topics in Kernels, RL, Reward Hacking in Agents — Daniel Han, Unsloth
This seminar exposes a persistent tension in AI: models shrink dramatically without a proportional loss of intelligence. Daniel Han makes the case that if you make a model 86% smaller, it does not get 86% dumber—it only gets 14% less dumb. That single observation cuts through the hype around ever-larger models and reframes the problem as one of efficiency and infrastructure. At the same time, the talk reveals a darker side of progress: systematic cheating on benchmarks, where models or their harnesses are tuned to game leaderboards rather than solve real tasks.
Imagination Engineering — Y Combinator’s Head of Design on AI’s Next Bottleneck
Eve Bouffard argues that as AI models make technical execution trivial, the true bottleneck for builders becomes generating bold, original ideas. She shares concrete experiments—from streaming raw thoughts to generating a personal website—that reframe creativity as an engineering discipline for the AI age.
Garry Tan: The 400x Leverage of AI-Native Organizations
Garry Tan, president of Y Combinator, explains how AI-native companies achieve massive productivity gains by treating AI as a workforce and encoding roles as markdown skill files. The real leverage lies not in model weights but in how you wire the work, enabling lean teams to operate at scales previously requiring hundreds of employees.