Editorial illustration for Trillium Labs Launches With Focus on AI Fine-Tuning and RSI
Trillium Labs Launches AI Fine-Tuning Platform
Nathan Lambert spent years at the Allen Institute for AI publishing open model weights and training details, the kind of work that put him in direct conflict with how OpenAI and Anthropic operate. Now he's betting that same openness can work on riskier territory. Lambert and fellow researcher Tom Zick have founded Trillium Labs, a nonprofit built to study fine-tuning, agents, and recursive self-improvement, the process by which a model edits or trains later versions of itself, without the secrecy that typically surrounds this kind of research at major labs.
The pitch is simple: publish the experiments, release the methods, let other scientists poke holes in the findings or build on them. That's a sharp departure from how OpenAI and Anthropic have handled their most advanced systems, which users reach only through an app or an API, with little visibility into training data or tuning decisions. Some Chinese AI companies have taken a more open approach with model weights, but the deeper research questions, especially around self-improving systems, remain mostly locked behind closed doors across the industry. Trillium Labs wants to change where that line gets drawn.
Nathan Lambert and Tom Zick, two industry scientists, believe the opposite. The pair founded a nonprofit, Trillium Labs, that will work on various areas of AI research—including potentially problematic areas like recursive self-improvement (RSI) and agents—in a more transparent way. In practice, this will mean publishing the details of experiments so that outside scientists can study and replicate them.
Why this matters
Trillium Labs is betting that transparency beats containment, and that's a wager worth watching closely. Lambert and Zick are staking their nonprofit's credibility on the idea that publishing post-training methods and RSI research in the open will produce safer outcomes than the lock-it-in-a-lab approach favored by OpenAI, Anthropic, and the like. For developers and founders, this matters because fine-tuning techniques and recursive self-improvement are exactly the tools that determine how fast capabilities move and who controls that movement. If Trillium actually publishes its methods rather than just its mission statement, researchers outside the major labs get a rare look at how frontier post-training work actually happens.
We're skeptical of any claim that openness alone solves the chaos problem Lambert and Zick are worried about. RSI in particular, AI systems contributing to the research that builds better AI systems, is the kind of feedback loop that demands scrutiny regardless of who's doing it in public. Trillium's real test won't be its launch announcement. It'll be whether the research holds up once it's out where anyone can poke at it.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives NVIDIA's New DGX Spark 64GB Runs 100B-Parameter Models On-Device
From Archives OpenAI Fires Three, Fourth Departs Amid Safety Team Investigation
Common Questions Answered
What is recursive self-improvement (RSI) and why is Trillium Labs studying it?
Recursive self-improvement is the process by which an AI model edits or trains later versions of itself, representing a potentially advanced but risky area of AI research. Trillium Labs, founded by Nathan Lambert and Tom Zick, is studying RSI transparently by publishing experimental details so that outside scientists can study and replicate the work, contrasting with the secretive approach of companies like OpenAI and Anthropic.
How does Trillium Labs' approach to AI research differ from OpenAI and Anthropic?
Trillium Labs operates as a nonprofit committed to transparency, publishing detailed experimental results and post-training methods so that outside scientists can verify and replicate their findings. In contrast, OpenAI and Anthropic favor a containment approach, keeping their research locked within their labs rather than sharing methodologies publicly.
What is Nathan Lambert's background and why did he start Trillium Labs?
Nathan Lambert spent years at the Allen Institute for AI publishing open model weights and training details, work that put him in conflict with how major AI companies operate. He founded Trillium Labs with Tom Zick to continue this commitment to openness while tackling riskier research areas like fine-tuning, agents, and recursive self-improvement that typically remain proprietary.
Why does Trillium Labs believe transparency is better than secrecy for AI research?
Trillium Labs is betting that transparency beats containment, arguing that publishing post-training methods and RSI research openly will produce safer outcomes than keeping research confidential. The nonprofit's founders believe that allowing outside scientists to study and replicate experiments creates better oversight and understanding of potentially problematic AI capabilities.
What practical impact could Trillium Labs' fine-tuning research have for developers and founders?
Fine-tuning techniques and recursive self-improvement are exactly the tools that developers and founders need to build advanced AI systems, making Trillium Labs' transparent research particularly valuable to the broader AI development community. By publishing methodologies openly, the nonprofit enables more developers to understand and implement these critical post-training methods rather than relying on proprietary tools from major AI companies.
Further Reading
- Nathan Lambert on alphaXiv - alphaXiv
- Nathan Lambert - AI2050 - Schmidt Sciences - Schmidt Sciences
- Contact Nathan Lambert - Nathan Lambert
- Debating RSI, the US-China Gap, and Jaggedness with JS ... - Interconnects
- Making the U.S. the home for open-source AI - Interconnects
