Research, product & code
Gareth Paul Jones
Product Lead, AI at Poe / Quora
I like the part where research meets a real product.
01Working things out, in public.
Selected livestreams
-
Evaluation
Evals for AI Agents
A practical look at testing agents against the work we actually expect them to do.
Watch the session ↗1:58:35 -
Model behavior
Building a Multi-Signal Reward Function
Working through how several imperfect signals can produce a more useful training objective.
Watch the session ↗1:18:12 -
Agent systems
Managing Coding Agents with Swimlanes
An operating model for keeping parallel coding agents legible, scoped, and moving.
02From the archive
Maskito · December 2016
The feedback was there.
I kept building.
Coffee, paper mockups, and a face-filter app that failed. An honest account of what I missed when I was too eager to build.
Read the retrospective ↗Photograph & source
This photograph was published with my December 2016 Maskito retrospective. It
shows the original paper mockup.
Read the original entry
I still code because it is often the shortest route from a fuzzy idea to an honest conversation.
— Gareth