CIR
Searching for lower cost to capability architectures. CIR compares candidates against strong Transformer baselines on total cost.
Research
The technology beneath intelligence
DotrixAI does not compete by training the largest models. We develop the technology that models are built and run on.
Research scope
Seven areas, at different stages.
Each area carries an honest status. Directions are areas we intend to study, not active products.
- Active through CIR
Architectures
Neural architectures that reach a given capability with less computation, memory and energy.
- Active through CIR
Learning systems
Objectives, optimizers and data use that make learning more efficient per unit of cost.
- Research direction
Memory and retrieval
Ways for models to store and recall information without paying for it in every step.
- Partly studied in CIR
Training and inference systems
The systems layer that decides what training and serving actually cost.
- Research direction
Runtimes and compilers
Software that maps models onto real hardware with less wasted work.
- Long term direction
AI infrastructure
Compute, data center technology and distributed computation for efficient intelligence.
- Principle and direction
Safety
Reliability, controllability and secure deployment as capability increases.
Research programs
Future programs
DotrixAI may run several independent programs over time. New programs will be listed here when they begin.
Our approach
Our core asset is technology and research.
What we build
- Neural architectures and learning systems
- Memory and retrieval mechanisms
- Training and inference technology
- Runtimes, compilers and kernels
- AI infrastructure and distributed computing
- Safety technology
What we are not
- A chatbot or consumer AI company
- An API wrapper or automation agency
- A consulting firm
- A frontier model company
- An open source collective
- A business built on owning more GPUs
Technology stack
From research to licensable technology.
Our expected commercial model is technology licensing. Research results may become architectures, software or infrastructure others can license.
Possible forms include technology licensing, enterprise licensing and research partnerships.
None of these offerings are available today. There is no pricing, because there is nothing to sell yet.
- ResearchToday
- Architectures and learning systemsToday
- Software, runtimes and compilersDirection
- Licensable AI technologyGoal
- Compute infrastructureDirection
- AI developers and laboratoriesFuture users
- Models and intelligent systemsOutcome
Safety
Efficiency without safety is incomplete progress.
More intelligence means more capability. Greater capability can bring both benefit and risk. We treat safety as a principle and a research direction. We do not claim to have solved it.
Reliability
Systems should behave consistently within known limits.
Controllability
People should be able to direct and correct what systems do.
Misuse
Cheaper intelligence also lowers the cost of harmful use.
Alignment
Capable systems should pursue the goals people actually intend.
Systemic risk
Widely deployed technology can fail in correlated ways.
Secure deployment
Technology should be deployable without exposing its users.
Long term research directions
Infrastructure, later.
Over the longer term, DotrixAI may investigate physical AI infrastructure. Possible directions include compute systems, specialized AI hardware and data center technology.
We are also interested in distributed and geographically spread compute. Some ideas draw on decentralized systems design.
These are exploratory research directions. None of them is a current product or commitment.
- Research
- Technology
- Licensing
- More research
- More efficient intelligence
- Broader cognitive capacity