Ask HN: What are you working on? (May 2026)

json2vec enables users to, for example, build tabular / transactional foundation models like TabBERT / PRAGMA dynamically... by just declaring their data schema. This is a space in which Netflix, Stripe, Revolut, Capital One, Nubank, J.P. Morgan, NVIDIA, etc. have been developing for several years.

json2vec goes a step further from just tabular data or structured transactional data. It enables arbitrary structured "json-like" observations with hierarchical BERT-like transformer encoder blocks. Financial transactions, chess positions, flight itineraries, raw tabular data, rideshare activity, ecommerce, behavioral sequence models... Any raw data able to be represented in `json` can be encoded into a tree of embeddings, and used for downstream finetuning for supervised machine learning... No feature engineering required.

https://github.com/granthamtaylor/json2vec

json2vec supports extensible plugin support for new data types (numbers, categories, raw text, datetimes, hashable objects [think: IP addresses and phone numbers], and raw embeddings), all of which may be pretrained via MLM-like self-supervised learning. If your needs are not met with the built-in datatypes, the framework is extensible in that you may build your own custom datatypes (think: geographical coordinates). Built in decision heads for a subset of datatypes enable predictive modeling multi-task and multi-array outputs (predicting fraud at a per-transaction level, or a per-account level).

json2vec also enables built in data pipelines for 100b+ training observations streaming from cloud storage. These pipelines integrate with layer of programmatic data querying and UDFs can consume the vast majority of upstream data processing so that developers don't waste time on massive batch data preprocessing jobs before model training.

Oh, and the best part: the model architectures instantiated by json2vec are mutable. Model developers can add and remove features and targets at their whim - allowing for truly reusable foundation models that can adapt for each individual use case.

My hope is that with a standardized hierarchical modeling framework, interested organizations can better collaborate with one by sharing reusable logic with one another instead of hardcoding use-case-specific architecture.

Each note gets a wholesome 'virtual desktop' "space":

- A markdown file (in a Git repo for all my notes)

- A working directory for project files

- Virtual desktop in hyprland that opens up as a terminal/browser split

- The terminal spawn a restorable tmux session where I edit the markdown file and open related terminals, where the note's workdir has its own `bin` directory in $PATH.

- Also, an ironbar widget showing the name of the current note at the bottom of the screen.

- Time tracking app to remember how much time I worked on each note.

And they are all bundled up together with save/restore capability, kinda like a VM, but on the application level. The idea is to support multitasking and never lose focus.

This Friday we pull GPT apart and rebuild bits by hand. A few weeks back we tackled ARC AGI 2026. Prevoiusly we did fine tuning and making GPT funny.

It is a sort of a guided hackathon (generally I plan goals for the day) and collaborative study group.

Much fun, no money, lots of smart folk asking good questions: https://playgroup.org.uk/