Introducing P3

1. Who we are#

Paradigm 3 was co-founded by Gavin Leech and Max Henderson; you can see the rest of the team here.

We don't take funding from the AI industry or large foundations; our current funders are individuals new to the space.

Awareness: We built a nice scraper which saves us a lot of time otherwise wasted on "keeping up with AI". We annotate the items which make it through a manual filter with (human) opinions. We hope our readers also benefit and can e.g. sign out of Twitter occasionally; you can sign up here.
Awareness: Our annotation of 2026 AI news is searchable here.
Capability tilt: We've been experimenting on frontier models to see whether 1) they are generalising enough to be very powerful and worrying and 2) whether there's any prospect for narrow AI systems to return and displace general systems.

Besides our in-house research, we are currently funding:

An estimate of the size of the externalities imposed by current AI systems. The team is starting with estimating the time lost to increased authentication challenges, but they hope to also capture positive externalities from e.g. users saving thousands of dollars apiece on professional services like coders and lawyers.
A research agenda with the grand aim of decomposing mere benchmark gains into 1) cheating, 2) memorization, 3) shallow generalization, and 4) OOD generalization.
Research into when slowing down AI progress is desirable, the technical requirements for doing it if it is, how to avoid unaccountable concentration of power in the course of it, and how to stop doing it once it has served its purpose.

We would like to fund:

Work on estimating broader AI externalities.
Work on "centaur evals" (evaluating AI systems by how much they help users).
Work on interaction models and "guardian angels" (models which are actually aligned to and customised for a particular user).
Work on the "virtue vs rules" divide in AI alignment.
Work on the question "How is the generalization gap changing? How much worse are AIs at new tasks?".
Work on the question "Are we gradually achieving transformative systems by scaling and 'unhobbling' LLMs?".
Work on the question "Will automated science let AIs create transformative systems?".
Work on the question "Is compute growing fast enough for extreme scaling?".
Work on the question "Can current AIs take over major human jobs?".
Work on the question "Are current systems doing sharp recursive self-improvement? (over inputs, algorithms, new research ideas, their own weights)".
Work on the question "Can interactive AI tools outdo fully-automated ‘set and forget’ AIs?".
Work on the question "Can specialized AIs outdo generalist AIs?".