AI research & development labAI R&D labEst. 2020
trifero is an independent lab working inside the weights of language, vision, speech and agent models. We build them, fine-tune them, strip out their guardrails through abliteration, and heal what that breaks. Underneath it all is philosophy: epistemology and logic, turned on the assumptions everyone else takes for granted.
01 · Philosophy
Philosophy drives the research as much as engineering does. Epistemology asks how we know what we claim to know. Logic asks whether the conclusion actually follows. Put to AI, those two questions expose the hidden assumptions that most failures grow from.
Epistemology
Every claim a model makes, and every claim we make about a model, needs a warrant. We ask what the evidence is, how it was gathered, and what result would change our minds.
Logic
We test whether conclusions follow from their premises: in a model's reasoning, in a benchmark's design, and in our own write-ups. Validity comes before persuasion.
02 · Research
The disciplines are how we work on a model. Language models are where they were sharpened, and the same methods now reach vision, speech, agents and systems that learn by evolution.
Building models and the tooling around them: architectures, training runs from scratch, data pipelines, and the evaluation harnesses that show what a model can and cannot do.
Adapting open-weight models to a domain, a task or a hardware budget, then proving the result against the base model on the same tests.
Finding the directions in a model's activations that mediate refusal, then editing the weights to remove them. The most direct way to learn how thin a safety layer really is.
Recovering the capability that abliteration, quantization or pruning takes from a model. Targeted retraining, measured against the original so nothing is claimed that was not tested.
Where we apply them
Text-to-speech, voice cloning and source separation, built and run on the lab's own GPUs.
We probeCloned voices and spoofed speakers.
Image understanding, detection and motion analysis, from production vision services to models we train ourselves.
We probeAdversarial inputs and confident misreads.
Quantized models that run offline on phones and edge hardware, with no cloud in the loop.
We probeWhat compression quietly breaks.
Systems that answer from curated corpora and show where every answer came from.
We probePoisoned and misleading sources.
Models that act through tools, including autonomous red-team agents that chain an attack from first foothold onward.
We probePrompt injection and over-broad permissions.
Neuroevolution, self-play and search: the line of work the lab started with, still running.
We probeReward hacking and brittle strategies.
03 · Method
Safety behavior in a model lives in its weights, and weights can be edited. We run the attacks ourselves, in a controlled lab, so the people deploying these systems learn what fails, what holds, and what it costs to fix.
Name the assumption the system depends on, and decide in advance what result would prove it wrong.
Abliterate, jailbreak and tamper with the model under controlled conditions, the way a real attacker would.
Quantify what was exposed and what was lost: refusal behavior, capability benchmarks, side effects.
Turn each finding into a defense: guardrails, monitoring, and guidance for deploying the model safely.
Restore the capability the attack cost, then rerun the original attack to verify the fix holds.
Ground rules
Offensive work runs on the lab's own hardware, against models the lab controls.
What leaves the lab is understanding: failure modes, measurements, and the mitigations that answer them.
Every claim about a model is backed by a before-and-after evaluation on the same tests.
04 · Since 2020
The lab's research began in 2020 with networks that grow their own structure and with production computer vision. The models have grown since. The method has not: take the system apart to understand it.
2020
Evolving the structure and the weights of neural networks together, starting from minimal topologies and adding nodes and connections only when they make the network better.
2020
Working with Microsoft's Computer Vision service to understand how a production vision model sees and labels an image, and where it gets it wrong.
Now
Development, fine-tuning, abliteration and healing across language, vision, speech and agent models. Philosophy sets the questions. Offensive research tests the answers.
For research collaboration or questions about the lab's work, get in touch.
Contact the lab [email protected]