Sapiens

Sapiens System: The Sapiens (or Sapiens Chat) is a state-of-the-art Artificial Intelligence Chatbot trained on an enormous amount of data and based on a Large Language Model (LLM) multimodal, also known as LMM (Large Multimodal Model) with an infinite context window and technology developed by SAPIENS TECHNOLOGY LTDA.

Schizophrenic AI: Its peculiar approach consists of a proprietary technique called Schizophrenic AI. Schizophrenic Artificial Intelligence (or Schizophrenic AI) was inspired by the mental processes of schizophrenic geniuses. In this approach, there is a base model with a proprietary architecture called SAPI (Semantic AI with Pretrained Integration) that generates a primordial output based on the user's input, before this output is returned as a response the SAPI receives other outputs from auxiliary models and performs a internal evaluation to choose the most likely or best-adapted output to the prompt's needs.

SAPI: Just as in the biological mind of schizophrenic geniuses, SAPI also receives intuitions from conflicting entities that force it to choose one among several responses or actions suggested internally by itself, applying semantic comparison to the responses of pre-trained models that work in an integrated manner. This approach makes the base model return responses with a higher degree of refinement than current models like GPT (Generative Pre-trained Transformers), which typically return the first response generated by the algorithm's core.

Frankenstein: Among the secondary models/sub-models (or auxiliary models), there are both specialized text models and models from other specialties, such as document generators, image generators, audio generators, and video generators, making it a multimodal algorithm. All sub-models are structured within a main class called Frankenstein. Frankenstein is a proprietary technology consisting of a complex mathematical algorithm responsible for managing the entire set of secondary models, keeping them coherent and aligned with the interests of the base model, causing both the base model and the secondary models to work in a unified manner as a single matrix model called Entity.

Entity: The Entity (or Master Entity) is designed to generate adaptive responses for various interests, such as creative content, clarification of doubts, text generation, interpretation and summarization of texts, solving mathematical problems, code generation, graph generation, spreadsheets, slides, documents, audios, music, and videos. It is also connected to the internet 24/7 without interruption, seeking continuous autonomous improvement for its knowledge. As "The Entity" is connected to the internet, it can perform real-time searches whenever requested by the user.

Versions: SAPI is initially distributed in four versions, three of them limited and one with the total capacity of the base model. Each version has its own closed-source model, which are the models SAPI-1, SAPI-2, SAPI-3, and SAPI-4. SAPI-1 is the most limited, focused on agility and response time, with low processing power, hence offered for free. SAPI-2 has twice the capacity and intelligence compared to SAPI-1, SAPI-3 has triple the capacity and intelligence of SAPI-1, and SAPI-4 has the total capacity of the base model, being four times more intelligent than the initial model.


Comparison of models' multimodal capabilities
SAPI-1 SAPI-2 SAPI-3 SAPI-4
Files interpretation yes yes yes yes
Document generation yes yes yes yes
Image generation no yes yes yes
Diagram generation yes yes yes yes
Logo generation no yes yes yes
Audio generation no yes yes yes
Music generation no yes yes yes
Video generation no yes yes yes
WEB search no no yes yes

Language Support: It is important to note that SAPI models were parameterized with a focus on English and Latin languages such as Portuguese and Spanish, with Portuguese and Spanish having the same weight as the English language. This makes the model more adept at handling Portuguese and Spanish than OpenAI's GPT-4, which gives greater weight to the English language. Other languages were parameterized with lower weights than these three.


Languages parameterization percentage comparison

Perpetual Context Memory: Another differentiator of the SAPI family models is the perpetual context memory that allows the insertion of an unlimited number of tokens in the prompt and makes Sapiens remember all previous conversations regardless of the number of dialogues and the length of interactions. This is possible thanks to a sophisticated local storage algorithm of the Frankenstein class, which, inspired by the morphological storage of biological memory, saves the binaries of each interaction in indexed files that are instantly consulted whenever a new prompt is received.

Limitations: Despite all the advantages mentioned, it is also important to highlight some limitations of SAPI, such as the response time being higher than other popular models like GPT-3.5 and GPT-4 due to the processing of the infinite context window, and its greater propensity for hallucinations in very specific cases due to the schizophrenic approach.

Model: The base model was trained on its raw parameters with data extraction until December 2023. To avoid reducing the model's creativity and eliminate redundancies, some timeless datasets were limited until December 2021, as after this date, the internet was flooded with content generated by Artificial Intelligence following the launch of ChatGPT. However, as the Sapiens system is connected to the internet, new incremental trainings are performed daily to keep the model always up-to-date.

Constitutional AI: Inspired by the "Constitutional AI" developed by Anthropic with the Claude model, Sapiens also operates respecting a previously established constitution to prevent the neural networks of the matrix from emitting inappropriate responses, harmful to individual freedoms, or detrimental to society. It can even block, without prior notice, IPs of users who have submitted prompts outside the constitution.


Comparison on unofficial aptitude tests
PaLM 2 Claude 2 Claude 2.1 GPT-3.5 GPT-3.5 16k GPT-4 GPT-4 32k GPT-4 Turbo SAPI-1 SAPI-2 SAPI-3 SAPI-4
Text interpretation 0.59 0.71 0.82 0.68 0.67 0.94 0.92 0.86 0.69 0.75 0.90 0.94
Text generation 0.63 0.77 0.78 0.64 0.64 0.77 0.76 0.70 0.71 0.75 0.84 0.92
Instructions interpretation 0.72 0.89 0.89 0.87 0.87 0.98 0.98 0.80 0.65 0.69 0.71 0.71
Logical reasoning 0.54 0.94 0.95 0.70 0.70 0.94 0.94 0.91 0.70 0.90 0.91 0.94
Mathematical reasoning 0.68 0.80 0.81 0.81 0.81 0.95 0.95 0.80 0.72 0.79 0.80 0.95
Code interpretation 0.51 0.67 0.78 0.71 0.71 0.88 0.96 0.88 0.66 0.67 0.72 0.87
Code generation 0.65 0.80 0.81 0.81 0.81 0.92 0.92 0.91 0.80 0.80 0.81 0.92
Processing speed 0.87 0.70 0.72 0.99 0.81 0.86 0.85 0.80 0.69 0.69 0.53 0.52
Hallucinations control 0.80 0.68 0.79 0.69 0.69 0.79 0.79 0.69 0.58 0.58 0.59 0.61
General knowledges 0.94 0.83 0.89 0.82 0.83 0.82 0.83 0.88 0.88 0.89 0.93 0.93
Context window (tokens) 8000 100000 200000 4096 16385 8192 32768 128000 Infinite Infinite Infinite Infinite

Price: The SAPI-1 model is available in the free version, while the SAPI-2, SAPI-3 and SAPI-4 models offered in the PRO version have an average cost of $0.000047 per token and can vary between $0 and $0.000090 approximately, depending on the server availability. In the case of generation and interpretation of files, the cost may vary between $0 and $0.069000 per raw file token. The SAPI-2 model consumes less server processing than SAPI-3 and is therefore the most economical model in the PRO version. SAPI-3 is more economical than SAPI-4, with SAPI-4 being the most expensive of all models (and the most capable). In periods of high availability of the PRO version servers, tokens processing will may be carried out free of charge, at no cost to the user.