Breaking

Trillium Labs Launches Nonprofit to Open Research on Self-Improving AI

October 2, 2026 at 12:03 EDT

Foundation Models
AI Agents
Open Source

Trillium Labs, a new nonprofit led by Nathan Lambert and Tom Zick, launched on October 2, 2026 with a plan to publish experiments in frontier AI research that many large labs keep internal. Its agenda includes post-training recipes, model behavior, agents, and recursive self-improvement, putting the organization in the middle of a live debate over whether risky AI research is safer when restricted or reproducible.

Oct. 2, 2026 · Trillium Labs

Open science enters risky AI research

A new nonprofit is putting post-training, model behavior, agents, and self-improvement research into public view, arguing that reproducibility can improve scrutiny.

$30M

planned training spend

months of training runway

$40M-$100M

fundraising target

public models listed yet

Budget scale

Each block represents $10M. The training plan is drawn against the upper end of the fundraising goal.

$30Mtraining plan

$100Mupper target

The open-research loop

Run experiments

Post-training, agents, and behavior tests.

→

Publish details

Recipes, logs, evaluations, and failures.

→

External scrutiny

Replication, risk checks, and debate.

Upside

Open records can make model behavior easier to audit and help outsiders reproduce safety-relevant claims.

Concern

Publishing work on agents and self-improvement could spread sensitive capability know-how before guardrails mature.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.