Posts
Most posts here also have a Korean version, linked at the top of each post.
- Notes
[Notes] Why Data Scientists and Data Engineers Need Sales Skills
Thoughts after reading how the CEO of Terminal X built a sales team. Engineers and data scientists need the kind of selling that starts from the other person's point of view, and unlike in sales, experience at real scale matters a lot in these jobs. If you don't have it yet, build as much end to end as you can.
- Financial AI
[FDS] Finding the 0.015%: Transaction Sequence Models, Anomaly Detection and Graphs
In 2024, card fraud in Europe was 0.015% of transactions by count, and fraud labels can take a week or more to arrive. How WeChat Pay, Stripe and Visa have added transaction sequence foundation models alongside their existing models, why XGBoost is still there, label-free anomaly detection and the foundation models from Leman Akoglu's group at CMU, and graph methods for finding fraud rings.
- Financial AI
Where Transformers Fit in Credit Models: Affirm, Revolut, Parafin and Ant
Affirm has put a transformer in front of its underwriting model. It reads each credit-report account as one token and feeds the resulting embeddings into XGBoost. From Revolut's PRAGMA and Parafin's ParaFormer to Ant's Falcon TST and BYB, where financial companies have placed transformers next to tree models, and how to read the numbers they publish.
- Payments
Will AI Companies Become Payments Companies? Who Makes Money When an Agent Buys
Meta plans to keep its personal agent Muse free and take a small fee from the transactions it completes. Where AI companies, payment firms and clouds are each coming at payments from, the four kinds of money on one agent purchase and the liability attached to each, and whose side a fee-earning agent is on, seen from fintech.
- Financial AI
Sales Forecasting (2) Inside Time Attention and Group Attention
A time series transformer lays a seller's sales and covariates out as a grid, then alternates attention across time and down the series in a group. Following the Chronos-2 architecture in PyTorch from tokenisation to the group mask, and what to watch when using it on sales, including why daily quantiles don't add up to a monthly one.
- Financial AI
Sales Forecasting (1) From Tree Models to Chronos
For a fintech that lends against sales, forecasting sits at the centre of limits, repayment and risk. First decide whether to predict the amount, the growth rate or the distribution, because that choice carries all the way into model selection. From LightGBM quantile regression to the Chronos family of time series transformers.
- LLM Systems
Nobody Touched a Line of Code and the System Broke: The Paths by Which Data Quietly Rots
The types check out, the values sit in range, and only the model goes strange. A price column that started recording dollars instead of won, a RAG system answering confidently from a two-year-old research report, and a feedback loop where the model's own misreading hardens into ground truth.
- AI Coding
Taking Apart Andrew Ng's AI Engineering Skills Map: Four Items Saying One Thing
Four AI engineering competencies, drawn from an analysis of ten thousand job postings. Rather than copying the list across, here is the structure underneath it. It is not a flat list but a stack, the same thing sits inside all four items, and the ability to build a model is not on the list at all.
- LLM Systems
Thinking About the FDE: Ten Names, One Job
Forward Deployed Engineer, solutions architect, resident SA, customer success engineer, deployment strategist, applied AI engineer. The names keep multiplying, and laid out together there are only two axes. Why AI companies are now racing to hire the role Palantir invented, and what the actual core of it is.
- World Models
How the Two Lineages of World Models Got Here: Seventy Years, Made Simple
Fine, there are neural and symbolic lineages of world models, but where did each start, and how did it arrive here? The one that tried to write the world down in logic and the one that tried to learn it from data, the walls each hit, and where each turned. Written to be readable with no background.
- World Models
Is Palantir's Ontology Really an Ontology?
Palantir puts the word Ontology at the centre of its product. It uses none of the W3C standards and does no logical inference, which has drawn a sharp charge: that this is database modelling in philosophical packaging. Taking the objects, links and actions apart, then measuring them against the classical definition.
- World Models
Norms Are Not in the Data: The Two Lineages of World Models
World models come in two lineages, and only one of them gets talked about. Reading the four axes of a symbolic world model as families of formalism turns up three layers of a different kind, and the last one asks a question the other three never touch. On why a model that learns from observation cannot, in principle, learn what is permitted.
- AI Coding
[Notes] These Days My Manager Isn't the First to Read My Report
You file a report and your manager feeds it to GPT for a summary and a critique. If almost everything now passes through a model somewhere, you can simply reverse the order and run it yourself first. How to cross-check across models, the four questions that get you a real review, and open-source tools like promptfoo, LibreChat and llm-council.
- LLM Systems
The Model Is Smart, So Why Is the System Stupid? Most AI Failures Start Outside the Model
Why an AI system that drew applause in a demo falls apart quietly in production. Why a prompt is not code, the CACE principle where one line changes everything, silent failures that return 200 OK, and the fact that four out of five incidents begin outside the model.
- LLM Systems
How Do You Grade an LLM? Eleven Evaluation Methods, Sorted into Four Families
LLM evaluation starts diverging the moment a correct answer scores zero. From BLEU, ROUGE and BERTScore through LLM judges and juries to agent trajectory scoring and safety gates: eleven methods grouped into four families, with what each is for and where each falls over.
- World Models
World Models (3) Why a Game Is the Harshest Test Bed
Real-time response, long-run consistency, many users, controllability. The four gates a world model has to pass to go from spectacle to something useful. A game demands all four at once at extreme settings, which is why it became the frontier, and those gates translate almost directly into the robot's problems. Part three of three.
- World Models
World Models (2) Why Not Hand the Whole World to the Model
One enormous model that learns everything, or a model that does only what it is good at while a system takes the rest? Dissecting the hybrid architecture Roblox chose at that fork: the game engine holding truth while a world model paints realism, and the design principle of separating state from appearance underneath it.
- World Models
World Models (1) What Comes After the LLM: AI That Predicts the Next Frame
If predicting the next token conquered language, can predicting the next frame conquer the physical world? What world models are, why they are called the key to unlocking the robotics data bottleneck, and what they still cannot do. Part one of three.
- Payments
Ah Bao Arrives: Alipay's Turn Toward an AI-Native Super App for a Billion Users
On 16 June 2026 Alipay formally shipped a version carrying the AI assistant Ah Bao (阿宝). Swipe right on the home screen and the icon grid gives way to a conversational interface that calls more than ten thousand services by voice or plain language. The largest interface change in Alipay's twenty-year history, and the next round of its agent contest with WeChat.
- AI Coding
Productivity in AI Coding Comes from Decomposition and Feedback Loops, Not Better Prompts
Notes from an AI-coding workshop. The point is not that a one-line spec produces working code. It is that you accept the context limits of an LLM, split human judgement from the AI's repetitive execution, and run the work as vertical slices on top of a fast feedback loop.
- AI Coding
Where Does AI Coding Tool Usage Get Recorded? The Four Tokens ccusage Shows You
Coding agents like Claude Code and Codex leave a usage log on your machine every time they work. ccusage reads those logs and reports four numbers: Input, Output, Reasoning and Cache Read. What each one means, why cache reads are counted separately from input, and whether that cache lives on your laptop or on a server.
- Payments
How Chinese Fintech Actually Works: The Triangle of Super Apps, Mini Programs and Payment Rails
The essence of Chinese fintech is not a good finance app but a three-layer structure: mini programs riding on super apps, with NetsUnion, POS and IoT holding it up from below. The market is around US$51 billion; NetsUnion clears more than a trillion transactions a year.
- Payments
The Evolution of the Digital Yuan: In 2026, e-CNY Becomes Interest-Bearing Digital Deposit Money
From January 2026, balances in named e-CNY wallets earn interest and sit inside bank liabilities, deposit insurance and reserve requirements. The next stage of China's institutionalisation of a CBDC, moving from digital cash toward digital deposit money.
- Payments
WeChat's Different Road: A 1.4 Billion MAU Super App Absorbs the Agent
While Alipay pushes payment infrastructure with AI付 and AI收, Tencent is putting an OpenClaw-based ClawBot inside WeChat and releasing Hunyuan resources into the mini program ecosystem. An agent strategy that starts from conversation and daily-life touchpoints rather than from payments.
- Payments
The Age of AI Paying: Alipay's AI付, AI收 and the Start of the Agent Economy
Alipay announced 300 million AI agent payments alongside Token Pay and AI Wallet. Not a payment feature but an early large-scale commercialisation of agentic commerce infrastructure, with an AI acting within delegated authority to pay, collect and settle.
- Financial AI
[FDS] How Does Stripe Judge Payment Fraud in 100ms?
Where do machine learning and deep learning actually get used in financial AI? Fraud detection in practice, through the Stripe Radar case.
- Financial AI
Agentic Engineering in Financial AI
Not automating because you trust the model, but automating on the assumption that it will be wrong. What agentic engineering means when the agent sits next to money, risk and regulation.
- LLM Systems
Context Compaction: Don't Give the Agent Everything, Keep Only What It Needs
You cannot hand an agent the whole conversation, every log and every tool result on each turn. Working through the compaction techniques that keep the task state a model needs for its next decision and drop the rest.
- LLM Systems
Better Agents Come From a Better Harness, Not a Better Prompt
How to turn a failing agent into a working one without touching the prompt once. What harness engineering is, and why it matters more than prompt engineering the moment an agent leaves the demo.
- Notes
Starting This Blog
A new blog, opened with AstroPaper.