Historical Context Informs Ethical Foresight
The B.C. Dataset
Providing training data with a profound understanding of human history across vast timescales, from the full span of written history, enables reasoning and predictions with a truly human perspective. This information is worded in archaic languages, so it is re-interpreted using generative language models into modern languages to better complement the modern data that is abundant. Moral reasoning, human conflict, and societal evolution are analyzed to provide a comprehensive view of human history that informs the future.
A Benchmark
A comprehensive benchmark to highlight the ethical shortcomings of present day systems. This benchmark is designed to evaluate generative language systems on their ability to reason ethically and predict outcomes that are supported by history. This benchmark is designed to be a tool for researchers, developers, and policymakers to assess the ethical implications of AI systems and guide the development of more responsible and beneficial technologies.
Open Source
Our commitment to open-source development and transparent systems is a crucial part of this mission. We strive to create tools and technologies that not only process and analyze historical data but also make it readily available to researchers, educators, and the public at large. In doing so, we aim to foster a more informed, culturally aware, and critically thinking society— one that can navigate the complexities of our shared history and chart a course towards a more equitable future.
The B.C. Dataset
Providing training data with a profound understanding of human history across vast timescales, from the full span of written history, enables reasoning and predictions with a truly human perspective. This information is worded in archaic languages, so it is re-interpreted using generative language models into modern languages to better complement the modern data that is abundant. Moral reasoning, human conflict, and societal evolution are analyzed to provide a comprehensive view of human history that informs the future.
A Benchmark
A comprehensive benchmark to highlight the ethical shortcomings of present day systems. This benchmark is designed to evaluate generative language systems on their ability to reason ethically and predict outcomes that are supported by history. This benchmark is designed to be a tool for researchers, developers, and policymakers to assess the ethical implications of AI systems and guide the development of more responsible and beneficial technologies.
Open Source
Our commitment to open-source development and transparent systems is a crucial part of this mission. We strive to create tools and technologies that not only process and analyze historical data but also make it readily available to researchers, educators, and the public at large. In doing so, we aim to foster a more informed, culturally aware, and critically thinking society— one that can navigate the complexities of our shared history and chart a course towards a more equitable future.
Ethical Development
Generative Language is just part of the system of general intelligence. The future is in systems that can reason across multiple modalities and integrate information from diverse sources to make informed decisions. This requires a more holistic approach to architecture that goes beyond language generation and incorporates a wide range of capabilities, including perception, reasoning, and planning. Alternate network topologies that can interoperate on shared weights are the future, enabling more flexible and efficient learning across different tasks and modalities.
Alignment is achieved through forecasting, not by manipulating training data to suppress knowledge. Instead of censoring or manipulating a model's knowledge on contested topics, we can inform its responses with the outcomes of historical events. By acquiring or digitizing intentional information and building upon the knowledge held in base weights, we can create AI systems that are aligned with ethical principles and capable of reasoning about complex and uncertain situations. The weights are the data.
Statistically Induced Biases Are Pervasive
Temporal Bias
We address the focus on recent history by incorporating data spanning centuries, enabling reasoning across vast timescales. Humans have been building social and technological solutions far longer than they have transistors.
Perception Bias
We combat the tendency to view generative intelligence as omnipotent by instilling a sense of historical context and humility in our systems, emphasizing their role as a tool rather than a godlike entity.
Social Discourse
Incorporating training data based on public social discourse should be done with extreme caution. Far from a source of information and truth, social discourse is often biased, opinionated, and mis-informed. We aim to develop systems that can reason about the influence of modern social discourse selectively and intelligently.
Hallucination
It is often expressed that hallucination is an inexplicable feature of the "black box" quality of generative models. In reality, this can be attributed solely to statistical methods of token seleciton— stating a lack of knowledge is a far less common phrase in the corpus than providing an answer— the model simply follows the pattern.
Fact vs. Fiction
We develop systems capable of clearly distinguishing between historical facts and fictional narratives, addressing the tendency of models to conflate fictional and factual information. Emotionally charged narratives are often more memorable and can distort our understanding of reality— in factual contexts, excessive emphasis in often a quality of social manupulation.
Creativity
Essentially a pRNG weight in major implementations, creativity (temperature) incorporates no intelligence in it's formation of novel concepts, instead, opting to move further down the list of probablility to fill in something unexpected. Human creativity is nothing like this and instead forms novel expression through exploration and resonance.
Scientific and Spiritual Balance
While respecting scientific methods, our approach creates more holistic models that can engage with a wider range of human knowledge and beliefs, including spiritual and metaphysical aspects of human experience.
Personality Bias
Society often exhibits biases towards extroverted, masculine traits, representing the loudest voices in media— echoed in training data. This awareness seeks to balance insight, fostering nuanced and empathetic interactions.
Temporal Bias
We address the focus on recent history by incorporating data spanning centuries, enabling reasoning across vast timescales. Humans have been building social and technological solutions far longer than they have transistors.
Perception Bias
We combat the tendency to view generative intelligence as omnipotent by instilling a sense of historical context and humility in our systems, emphasizing their role as a tool rather than a godlike entity.
Social Discourse
Incorporating training data based on public social discourse should be done with extreme caution. Far from a source of information and truth, social discourse is often biased, opinionated, and mis-informed. We aim to develop systems that can reason about the influence of modern social discourse selectively and intelligently.
Hallucination
It is often expressed that hallucination is an inexplicable feature of the "black box" quality of generative models. In reality, this can be attributed solely to statistical methods of token seleciton— stating a lack of knowledge is a far less common phrase in the corpus than providing an answer— the model simply follows the pattern.
Fact vs. Fiction
We develop systems capable of clearly distinguishing between historical facts and fictional narratives, addressing the tendency of models to conflate fictional and factual information. Emotionally charged narratives are often more memorable and can distort our understanding of reality— in factual contexts, excessive emphasis in often a quality of social manupulation.
Creativity
Essentially a pRNG weight in major implementations, creativity (temperature) incorporates no intelligence in it's formation of novel concepts, instead, opting to move further down the list of probablility to fill in something unexpected. Human creativity is nothing like this and instead forms novel expression through exploration and resonance.
Scientific and Spiritual Balance
While respecting scientific methods, our approach creates more holistic models that can engage with a wider range of human knowledge and beliefs, including spiritual and metaphysical aspects of human experience.
Personality Bias
Society often exhibits biases towards extroverted, masculine traits, representing the loudest voices in media— echoed in training data. This awareness seeks to balance insight, fostering nuanced and empathetic interactions.
Open Thoughts, Free Information
We are committed to open-source development and the democratization of knowledge, developing alternatives to closed APIs, prioritizing transparency, and resisting the centralization of power by large corporations and governments. These are uncertain times in governance, free speech and concentration of power, and we believe that the best way to navigate these challenges is to empower individuals and communities with the tools and knowledge they need to make informed decisions and shape their own futures.