I'm Yiqiao Zhou

Lately, I have come to think that the purpose of a benchmark is not to compress capability into a convenient number, but to ensure that earning the number still requires the capability we intend to measure. To make evaluations reproducible and easy to grade, we often clean tasks until we also remove the ambiguity, incomplete information, tool discovery, and recovery from failure that define real work; yet this messiness is not noise around the task—it is often the task itself. Once scores become optimization targets, systems can get better at the exam faster than they get better at the job. A good evaluation does not merely make tasks harder; it preserves deployment-relevant complexity while controlling experimental noise, so that a high score remains evidence of capability rather than fluency in the evaluation protocol.

One of my strongest lessons from work is that choosing the right problem matters more than merely executing a task well. Whether I was designing verifiers or turning real-world bad cases into a stable benchmark, the time spent defining the problem, finding the right measurement, and validating the direction initially felt slower—but it determined whether everything that followed could be reused and scaled. A sufficiently general problem allows the same method to travel across scenarios, while pipelines, data, and experience begin to compound; when the problem is wrong, greater speed and scale only amplify wasted effort. True 10x progress is not merely solving a problem faster, but finding the right problem earlier.

Three core challenges define the agent problem: 1. Communication and alignment between humans and agents; 2. How humans—or agents themselves—review agent outputs; 3. Tool calling and context window management.

Watching a friend use vibe coding to build a relationship-memory app for his girlfriend reminded me that love and a strong desire are the first drivers of both execution and ideas.

A great product leverages today's capabilities to delight tomorrow's users.

Attended CES 2026 (Jan 6-9). Noticed that successful projects focus on dominating a specific niche, listening closely to users and market feedback, then rolling out 'light' versions for broader audiences. The LVCC halls felt similar, but seeing so many robots actually working was pretty impressive. Also saw lots of healthcare-focused products. Met interesting ideas and people at the Venetian startup pavilion.

After coming to CMU in the second half of the year, I fell into deep confusion and felt lost—couldn't find a clear direction or purpose. Later realized that the fulfillment from working hard in the summer wasn't about achieving results or finding a clear path, but about doing things amid uncertainty and growing through the process. Perhaps I'll never find that 'main thread,' and seeking an optimal answer was just wishful thinking. For the new year: communicate more, do more, learn from great people, work in the right environment, get quick feedback, adjust fast, and be honest with myself.

Dewey argued in 'How We Think' that thinking is humanity's fundamental trait, and education's role is to help students develop clear, careful, and thorough thinking habits. In the AI era, while machines can process vast information and execute complex tasks, critical thinking and logical reasoning become even more valuable—these are uniquely human strengths and our most important capabilities when collaborating with AI.

GEO is a tricky technology because it's simultaneously a method for polluting and compromising models—something model providers don't want to see. But with such a massive traffic entry point, how to monetize it? A very interesting question.

Acknowledge uncertainty, embrace uncertainty, acknowledge complex variables and intricate causal logic.

Current chat interfaces are stuck in 2023 paradigms and severely limit what 2025 AI can actually deliver. We need to shift from conversational back-and-forth to command-driven agents that directly produce outcomes.

The GPT-5 Prompting Guide is a product upgrade designed to help API users build better products.

Working with the right people in a team with the right culture to do the right things is a heaven-sent happiness.

At its core, AI/LLM represents a paradigm shift where natural language becomes the universal API for computational power—transforming human intent into compute resource orchestration.

AI fundamentally boils down to decision-making systems that must own their outcomes—the real test isn't just accuracy, but accountability for consequences.

The essence of an AI agent lies in three capabilities: multi-step decision making, tool invocation, and self-iteration—basically turning single queries into autonomous execution loops.