Day one focused mainly around successfully implementing AI. Responsible AI (Trustworthy & Ethical) were topics mentioned during the presentations and discussions. Here are some highlights and learnings from three key sessions:
AI Strategy manual for CTOs - A corporate journey to production
Mistakes often made when going to production
Overestimating the applicability of off the shelf solutions
Underestimating computing cost
Rushing from prototyping to production -> build a full stack team and define complete software specifications upfront
Underestimating the importance of change management -> design and implement re-skilling strategies for your workforce
High level steps:
Opportunity mapping
Tech assessment
Prototyping: set KPIs
Go-to-production strategy: build P&L of the project
Product development
Adoption, maintenance and improvements
Panel Discussion: Getting to Production-Ready: Challenges and Best Practices for Deploying AI
Define carefully the use case
Understand the data
Understand how to move from development to production environment
Align expectation across stakeholders and ensure buy-in
Understand how the model becomes part of the larger ecosystem
Plan for data acquisition -> data curation -> data guardrails
Models fail if guardrails not right
Test and try to break the model
Avoiding bias
Responsibility starts at the collection of data and understanding the sources
Ensure data is up to date and keep it up to date
Data leaks and privacy are still an issue
Explainability
Data needs to be timely = only possible with automation and orchestration starting at pre-production
• Ensure that data is sovereign (research shows 65% of data is managed onprem but 90% of organisation don’t have the infrastructure to run it)
Putting Structure into Chaos - How to develop trustworthy AI-Powered workflows
The key to successfully deploying AI is evaluating the AI model = does it fit my purpose?
Classification is easy to assess vs GenAI is more difficult
Sampling is the most misunderstood method
Ask what is the confidence interval of the sample size?
How much is enough?
It is easier to evaluate the workflow
Back propagating from customer journey
Get advice from someone who has built the workflow for your industry
Comentários