> git log --projects

Things we built.

Every project ships with its code and a write-up of what worked, what broke, and what we'd do differently.

$ paperpal ask 2310.06825 "what is sliding window attention?"
→ answer + 3 citations (p.2, p.4, p.7)

Shipped

Chat with any arXiv paper. Every answer cites the page it came from.

$ python train.py --params 10M
step 4200 · loss 1.93 · "POW! said the robot"

Cooking

A 10M-parameter GPT trained from scratch on comic book dialogue.

$ arena vote --blind
model A vs model B · you picked B

Experimental

Compare two models side by side and vote blind. Results feed a live leaderboard.

$ promptdiff v3.txt v4.txt
- be concise
+ answer in 3 bullets max

Shipped

Git-style diffs for prompts, with the eval score change shown next to every edit.

$ taste "strawberry"
A: str|aw|berry · B: straw|berry

Shipped

Type any text and see how different tokenizers slice it, side by side.

$ replay trace_0419.json
step 7/23 · tool: web_search · ▶ next

Cooking

Step through an agent's run like a debugger: every thought, tool call and result.