Digital representation is missing for billions of speakers - see the World Bank report on the "the digital divide" button.
If a LLM includes an Indigenous or scarce data language, the architecture penalises complex languages.
Infringement of cultural property and privacy is a feature, not a bug, of LLMs.
LLMs are highly polluting, with mega-scale data centres crowding out power for homes and schools.
Language revitalisation infrastructure that follows participatory design principles, is affordable and surveillance free.
Structural syntactic foundations, with architectures designed for morphological complexity and data sparsity.
Models that are massively less polluting, with zero water usage.
Typologically informed model architecture
Small model payloads that run on older PCs and cell phones
bezoku builds every component, resulting in auditability and low tech debt
Symbolic instructions that transfer to semantic solutions
"The best part of comedy is figuring out the puzzle pieces of a joke and getting the self-regard from having accomplished a difficult thing. Whatever your chosen profession is, please don’t let AI rob you of the fun part of it." Ronny Chieng, Comedian
Endangered languages have a different challenge to English or Mandarin - they are disappearing.
That does not mean they are immune to high resource language trends.
Cognitive Offloading is leading to a large scale cognitive crisis, as explained by the Dunning-Kruger effect, when AI output goes unquestioned.
Long form generative output in Indigenous and low resource languages, will not be supported.
Putting Indigenous people before profit.
Research can be found here - https://github.com/bezokurepo/papers
Leveraging Intel openVINO
Cloud | Device | USB
Dialogue and collaboration without exploitation or extraction
Source: Eugenia Urrere, Indigenius
Bezoku collaborates with Indigenous communities, ideating together
"The report advocates for a pivot away from resource-heavy AI models in favour of more compact models."
← Bezoku founder Ian Gilmour speaking at LT4ALL at UNESCO in February 2025
community.intel.com
Bezoku delivers secure, affordable SMLs with small energy footprints
About Bezoku AI start-up company Bezoku (https://bezoku.ai/) from South Florida, USA, works with communities in the design of generative AI applications. The Bezoku developer team consists of customer centric linguists, mathematicians, machine learning and software engineers who deliver commercia...
Our hackathon in May 2025 was a full foundation model build using LSTM architecture and CoNLL-U. It was awesome.
Read more on Intel's AI blog
community.intel.com
What Startups Built This Time at Intel® Liftoff Days
Intel® Liftoff Days returned from May 12–16 for our second big dev sprint of the year, bringing together another incredible group of early-stage AI startups. This is the highlight of each quarter in the program: five days to build, refine, test, and demo, all with full access to Intel® Tiber™ AI Clo...
Running on immersion cooled GPUs supplied by our compute partner DENVR in Canada.
Bezoku models are state of the art without costing the earth due to best in class power usage efficiency, cooled without water.
Flags are approximations and are illustrative - contact us to discuss any changes that avoid cultural mis-representation.
Contact us if you would like to participate or consult on our roadmap.
* Punjabi has two main scripts: Gurmukhi & Shahmukhi
By aligning ownership with our customer base, we aim for more than just the next quarter.
Indigenous and scarce data languages infrastructure