Cogniti is one of a new generation of software that uses generative AI to augment human intelligence and work.
Steering a large language model #
Driving Cogniti’s abilities are powerful large language models (LLMs) developed and hosted by leading AI organisations such as OpenAI, Microsoft, Anthropic, and Google. There are many LLMs out there, including GPT-4, Claude 3, Gemini, and others. These LLMs drive many of the generative AI tools used by the general public, including ChatGPT (powered by GPT-3.5 and GPT-4), Microsoft Copilot (powered by GPT-4), Claude AI (powered by Claude 3), and Gemini (powered by Gemini).
Those general purpose generative AI tools are powerful, but there are also some issues around using them in an educational context.
One of the issues that Cogniti overcomes is by allowing educators to ‘steer’ the LLM. This is done by providing a special message directly to the LLM which it respects each time it receives a user’s prompt and generates a completion. This ‘system message’ is crafted by educators, and not seen by student users, meaning that Cogniti ‘agents’ (customised chatbots) can behave in controlled ways that are more desirable for learning.
Giving the AI resources to draw upon #
Another important aspect of Cogniti is its ability to allow educators to provide ‘resources’ to the generative AI. These resources could be lecture notes, an open educational resource, marking rubrics, assignment guidelines, and more.
These resources can be considered by Cogniti when it generates completions in response to a user’s prompts – find out more about how this works.
The way this works technically is called ‘retrieval augmented generation‘. This means that when a user types in a prompt, a Cogniti agent will search through the resources available to it, and find snippets from those resources – ‘retrieval’. After the snippets are retrieved, it will add the snippets to the user’s prompts in the background, so that the LLM can consider the information from the snippets when it generates its completion – ‘augmented generation’.
Is this ‘training’? #
Not quite. ‘Training’ of the AI models is what OpenAI, Google, Anthropic, and other AI companies do to prepare the underlying AI models like GPT-4, Gemini, Claude, etc. This involves using a huge set of data to establish the strengths of the connections between the artificial neurons in the artificial neural networks that make up each AI model.
Data you enter into Cogniti is not used for training like this. Cogniti uses AI models that are already trained by these AI companies.
Instead, what Cogniti allows educators to do is set instructions and resources that are provided to these AI models (that are already trained) to influence their behaviour and their grounding knowledge.
Secure and private #
Cogniti can be set to connect to a range of different generative AI models, depending on the needs and contexts of your organisation. These include the GPT family (created by OpenAI, and hosted by OpenAI or on the trusted Microsoft Azure platform) and the Claude 3 family (created by Anthropic, and hosted on the trusted AWS Bedrock platform). Gemini support is under development.
You can experience Cogniti for yourself from a student perspective:
If you’d like to create your own agents, get in touch and we can discuss collaborations.
