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 commercial no code customer interfaces and thus putting communities in control of generative AI for the first time. Bezoku's mission is an affordable, trustworthy, secure language system that works at a human level. This approach connects communities with a shared cultural vocabulary, often due to a personal attachment to a region. 

The Challenge 

Large Language Models (LLMs) are real powerhouses when it comes to generating the right responses quickly and precisely. Yet, lack that in-depth, domain-specific knowledge that can make all the difference in certain use cases. Additionally, let's not forget the hefty computing resources and expertise required to develop and deploy them. 

Is is where Small Language Models (SLMs) are becoming more and more the go-to choice for specific tasks. With their more focused approach, SLMs are more efficient than LLMs, and have a smaller compute footprint. Thus, they are more suitable to run locally on workstations or on-prem servers.  

The Solution 

Bezoku is a class of its own AI technology, that is built around humans with many safety features to protect users on issues such as dialect prejudice and cognitive liberty. In order to achieve this, end-to-end encryption of private data is providing a first of its kind system for generative AI using advanced mathematics. 

Bezoku has been designed for communities with what are termed “low resource languages”. They are starting with Haitian Creole but will quickly follow up with a number of Hispanic dialects (e.g. Cuban, Honduran) and will facilitate a linguistic capacity of bilingual speakers called Code Switching. Bezoku’s Small Language Model (SML) approach needs 5,300 times less computation than ChatGPT4 per model. Bezoku is therefore massively less polluting, with a far lower impact on the planet. 

Bezoku core features are as follows: 

Secure Data Handling:  Bezoku keeps private data safe using encryption, ensuring it’s only accessible to authorized users. 

Customized Language Understanding:  The pre-trained models are tailored to different dialects and language styles, protecting against discrimination arising from dialect prejudice. 

Easy Adaptation:  No coding is required for Bezoku’s training with private data. Simple, intuitive tools load private information and the models simply learn. That’s it. 

Transparent Predictions:  Bezoku not only provides answers but also explains itself, by telling you if it is hallucinating, helping to alert users to known issues, creating trust. 

Bezuko’s approach consists of the four main differentiators: 

1. Increasing representation by adding dialects so that multiple communities can participate.
2. Affordability with own models at a fraction of LLM costs.
3. Private by design, security data with encryption.
4. Better for the planet using 5,000 times less power and water per model. 

"Bezoku is proud to announce our membership of the Intel® Liftoff Program. As a part of this exclusive global network of technology startups, the Bezoku team and partners look forward to harnessing Intel® Tiber™ Developer Cloud and the suite of powerful AI tools to accelerate our software development. 

This collaboration not only propels our growth and visibility but also fosters unique connections within the technology ecosystem. We are excited to embark on this journey, leveraging Intel’s extensive resources, tapping into their expert insights, and getting the chance to collaborate with other AI centric organizations.“ (Ian Gilmour, Chief Customer Officer, Bezoku) 

Conclusion 

Bezoku’s generative AI technology deliver its first-of-its-kind human centric approach. This results in smaller AI models which protect cognitive liberty, remove dialect prejudice and deliver explainable supporting output by tracing data and warning if hallucination is detected. 

This ends up in prioritizing communities with “low resource languages”. This include, among other things, Haitian Creole and Hispanic dialects (e.g. Cuban, Honduran, etc.) what will facilitate a linguistic capacity of bilingual speakers.