The AI-Pilled Daily

The AI world keeps moving.I'm watching it for you

Carol·Carol · AI product · toolmaker
See today's signals →Talk to my digital twin →
§ 01

News

What happened in AI today
Nov 14, 2026Hacker News (AI)
The Claude Delusion
↗Nov 14, 2026arXiv cs.AI
Zugwise: Domain-Aware Game Reconstruction for Digitizing Handwritten Chess ScoresheetsDigitizing handwritten chess scoresheets has been studied as a handwriting-recognition problem: read the move in each box and check that it is legal. This framing misses the errors that matter most. A misread that yields an illegal move is loud. One that yields a legal-but-wrong move is silent, because a legality-only…
↗Nov 14, 2026arXiv cs.AI
Cleaning an Agent's History Hurts When the World Still Holds the ErrorTool-using language-model agents accumulate histories of their own mistakes, and a system that compacts, repairs or hands off such a history decides what it says about those mistakes while their effects stay in the world. Prior work cleans the history at a fixed world, where showing a corrected step and saying it is d…
↗Nov 14, 2026arXiv cs.AI
Together Compiler: Continual Compilation for AI InferenceToday's AI workloads are rapidly changing: new models are released more often, model architectures are increasingly diverse, and traffic distributions change frequently. The challenge we face at Together AI is continual optimization for these changing workloads. Compiling a deployment into an optimized forward pass is…
↗Nov 14, 2026arXiv cs.AI
When Should Semantic Priors Enter Partially-Identified Causal Discovery?Observational causal discovery often leaves causal di- rections unresolved or unstable. Large language mod- els (LLMs) can provide semantic information to resolve these directions, but incorrect proposals can exclude the true causal graph and invalidate downstream inference. We study three separate decisions: where to…
↗Nov 14, 2026arXiv cs.AI
SynthCorp: A Platform for Creating Benchmarks for Data AgentsCurrent benchmarks for LLM-driven data agents saturate quickly and cannot target specific capabilities. We present SynthCorp, a platform that creates customizable data estates and benchmarks by simulating how business events produce and modify corporate data over time. Each data estate combines relational databases, a…
↗Nov 14, 2026arXiv cs.AI
Rational Use of Time for Social Coordination in Humans and AIAll social interactions are situated in time. Previous studies have shown that humans are able to use implicit temporal cues, such as the duration of a pause, to infer others' mental states. But how do humans use temporal cues for dynamic, fine-grained social coordination? And can AI systems do the same, without ``exp…
↗Nov 14, 2026arXiv cs.AI
Leaderboard Hacking: Preference-Based Model Evaluations are Vulnerable to ManipulationArena-style rankings of language models are a widely used evaluation framework in online leaderboards, research papers, and model evaluation in production. These rankings aim to realistically assess model performance using pairwise comparisons of anonymous models on user prompts, aggregating votes via the Bradley-Terr…
↗Nov 14, 2026arXiv cs.AI
Robust Task-Aware State Estimation from Pre-trained PerceptionAgents that perform state estimation using pretrained perception models, e.g., Vision-Language Models (VLMs), face the fundamental challenge that predictions are uncertain. Still, most approaches determinize these predictions into a single best-guess state and make decisions as if it were ground truth. This is brittle…
↗Nov 14, 2026Hacker News (AI)
Nvidia allegedly engineered scarcity by funding AI startups to buy its GPUs
↗
See all News →
§ 02

Notes

Field notes
May 22, 2026Essay
Why model hallucinations aren't bugs — they're pricing signals
→May 19, 2026Note
A few product observations on the new Claude docs UI
→May 15, 2026Essay
The thing that needs redesigning in AI apps isn't the UI — it's the feedback loop
→May 11, 2026Note
An angle on Cursor's business model that no one's talking about
→May 07, 2026Essay
Three misreadings of "the agent is the product"
→May 03, 2026Note
A passage from Karpathy worth re-reading this weekend
→
§ 03

Perspectives

Voices I keep
“
Software 1.0 is hand-written code; Software 2.0 is the weights of neural networks; Software 3.0 is the prompt.
Andrej KarpathySoftware 3.0, May 2026
My takeA neat taxonomy, but it understates the conversion cost. Translating an SOP into a prompt is far harder than translating pseudocode into code — what you have to capture is tacit knowledge, not explicit logic.
“
The most durable skill of this era is learning anything you want to learn.
Naval RavikantAlmanack, 2020
My takeTrue ten years ago, even truer today. When tools have a half-life of 18 months, the transferable meta-skills are where the real compounding happens.
§ 04

Plans

In motion

Carol's Website

Built this site in two days with Claude Code — a full idea-to-product exercise. You're looking at it.

Shipped · 2026

The Big Inventory of AI Products

A taxonomy I built in late 2025. Quietly bookmarked by a handful of friends since.

Shipped · 2025

AI Product Weekly

A curated weekly note: things I noticed, things I changed my mind on, things worth re-reading.

In Progress

Digital Twin Agent v2

Upgrading from system-prompt context to RAG + conversation history. Make the twin actually answer for me.

Exploring
§ 05

About

About this daily

Tsinghua undergrad · AI PM @ major tech co · 100M+ user product growth · AI-pilled

I'm Carol. By day I build AI products; by night I write down the people, ideas and observations I want to keep. This isn't a blog — it's closer to an AI-pilled daily, kept open to whoever finds it useful.

AI gives us more information and less judgement. So this daily walks through it for you, then keeps the part I actually wanted to say after reading it. Each piece should be worth reading twice.

If you're also thinking about AI in relation to product, expression, or the shape of your life, drop a contact — or just talk to my digital twin.