Open Source AI Models for Regulatory Reporting, Compliance and Data-Driven Decisions

Using free AI software from Big Tech suppliers is chasing your tail.

Microsoft released Phi-3, an AI model that runs on personal devices. This means data remains private rather than being sent to the cloud.

Apple released eight open-source models for developer use. Four are pre-trained and four are instruction trained.

That sounds good, what does it mean?

Why are Big Tech AI Models free?

Open source software has no licensing fees. There are lots of options depending on whether you work with text, image or speech. RemoraTech uses the best-in-class to deliver low cost AI-enhanced workflows to financial services clients.

Microsoft and Apple released these models to enhance the value of their existing products. They broaden reach to those sceptical of the cloud who want to try AI. They also head off competition from lower cost models.

AI models are ranked by parameter count. A parameter is a variable with an arbitrary value which is then trained for a purpose. The number of parameters is a proxy for model complexity.

OpenAI’s GPT-4 is believed to have over a trillion parameters across eight linked models. Google’s PaLM 2 has hundreds of billions. By contrast, Phi-3 has 3.8 billion and is tailored for smartphones and laptops. It is an alternative to using Apple’s models, which range from 270 million to 3 billion parameters.

Apple’s four instruction-trained Efficient Language Models (ELM) are for specific tasks. They are suited to AI assistants and chatbots. If you provide customer service via Apple devices it is now easier to add AI-powered question and answering.

Which AI Model is right for Financial Services?

The financial services industry requires data security. Financial and reputational damage from data leaks are a constant concern. Cloud-based models that share data, such as ChatGPT, are non-starters.

Thereafter models must be fine-tuned for specific tasks. This provides higher capacity than prompting, token-savings (a measure of running costs) and faster responses. Custom workflows include answering client questions about contracts, document processing and report writing.

AI is excellent for pattern recognition and spots correlations and overlapping risks in funds and securities that human’s won’t see. McKinsey reports that only 2% of AI use at work is for data science. This indicates an opportunity for smaller firms to level the playing field by using open source AI models trained to compare and contrast investments.

Fine-tuning for Specific Workflows

It’s important to have a tried and tested workflow to automate. Examples we’ve delivered include Health and Safety advice for remote workers, chatbots that enable managers to interrogate investment data, and forecasting with proprietary methods and metrics.

The more specific the workflow the less complex the model required to run it. Hence why Microsoft and Apple release smaller models for particular tasks. But the wider world of open source software is better suited to financial services workflows.

Big Tech companies’ free AI models are designed to keep you within their environment. Pre-trained models lack the power of open source alternatives, while instruction-trained models define workflows for users.

Big Tech will always want to sell subscriptions to the latest models and their free software will lag. Running workflows on these free models means playing catch up with competitors who harness more power.

To get ahead of the game, choose an important workflow to automate. Then fine tune the most appropriate open source model with company-specific knowledge.

If you’d like to discuss the use of AI in financial services please contact Harsh Patel or Tanisha Lakhani.