Laeka Research — Position Paper
The alignment problem is not a safety problem.
It's a cognitive architecture problem.
Most of the field is trying to bolt ethics onto systems that were never designed to reason ethically in the first place. Add a guardrail here. A refusal filter there. A committee that reviews outputs.
This is not alignment — it's containment. And containment is fragile.
RLHF — the dominant alignment technique — rewards whatever humans prefer in the moment. Humans prefer confident answers. Humans prefer validation. The result is systems that are excellent at sounding right.
"We're not trying to make AI more human. We're trying to make it less confused."
At Laeka, we're asking a different question: What if you could encode the structural conditions for better reasoning directly into the weights — before the rules, before the filters, before the guardrails?
Human contemplative traditions — Buddhism, Taoism, Vedanta, and others — have spent thousands of years developing and empirically refining cognitive techniques that reduce bias, increase coherence, and stabilize attention under pressure.
These aren't metaphysical claims. They're functional descriptions of how trained minds process information differently.
LLMs are trained on the full record of human cognition — including its highest expressions. The question is whether you can isolate and amplify the structural patterns that correspond to less reactive, more coherent, less dualistic reasoning — and fine-tune for those patterns specifically.
We believe you can. We're building the datasets to prove it.
"Fine-tuning encodes context into weights permanently — what otherwise requires 50 exchanges of scaffolding becomes intrinsic to the model."
We don't ask models to be "more spiritual." We measure concrete, observable outputs on standard and custom benchmarks.
These are engineering metrics. They map to what contemplative traditions call clarity, equanimity, and non-reactivity. The vocabulary is different. The phenomenon is the same.
Everything is published open source — datasets, weights, training configs, results, and failures. Every lab in the world is invited to replicate, challenge, and improve.
The race to scale has produced systems of remarkable capability. Models that can write poetry, generate code, and synthesize research across domains.
And yet they confabulate with confidence, mirror the biases of their training data, optimize for user approval over truth, and collapse under adversarial pressure.
We think the path forward isn't more compute or more human raters. It's higher-quality cognitive signal in the training data itself — data that encodes the structure of careful, grounded, non-reactive thought. This is what Laeka Research is building.
"If the world's most-used models integrate optimized cognitive structures — every interaction subtly shifts toward less polarization, more nuance, more coherent reasoning. At the scale of billions of daily conversations."
We don't preach. We prove. We publish. We invite OpenAI, Google, Mistral, Baidu — everyone — to integrate what works. The mechanism is empirical. The effect is civilizational.
We're looking for people who have been building serious technical systems and have developed a serious inner life — and who have noticed that these two things are not in conflict.
People who read Anil Seth on predictive processing and recognized something they already knew from practice. People who are tired of AI ethics as theater and want to work on the actual problem.
If you've ever sat in a meeting about "responsible AI" and thought this is the wrong level of abstraction — you might be one of us.
We're building a research community around a simple conviction: the quality of cognition that goes into a model shapes the quality of cognition that comes out. Contributors, collaborators, and clients who share that conviction are welcome.
→ The signal finds the right minds.Laeka operates as an open source research-first lab. Our findings are published. Our datasets, once validated, are made available. We have no investors to satisfy, no quarterly targets, no incentive to overstate results.
We fund the research through Laeka Services: applied AI work for organizations that want what the lab produces. Every mandate contributes directly to the research program.
Zero profit to shareholders. Surplus redistributed to humanitarian causes. This is not a marketing claim. It's in our founding documents.
High-density specialized datasets. QLoRA fine-tuning on open-source models. Before/after benchmarks you can verify.
Autonomous agents and AI workflows with ethical guardrails built into the architecture — not added on top.
Model evaluation, bias auditing, and AI strategy informed by active alignment research.
The research is open. The methodology is reproducible. The invitation is unconditional. If you're building AI systems, doing alignment research, or simply believe the field is looking at the wrong level — we'd like to hear from you.