Business of AI is hard

The business of AI is hard

Ref Article

Key Challenges

• Heavy Cloud Infra Training needs
• Ongoing human support
• Problem of edge cases
• Commoditization of AI challenges (You build someone gives for free API)

Observation - AI companies resemble traditional IT services

Hidden debts

• Retraining is increasingly recognized as an ongoing cost
• Operate on rich media like images, audio, or video
• Human-in-the-loop systems take two forms, contributing to lower gross margins for many AI startups.
• Social media companies, for example, employ thousands of human reviewers to augment AI-based moderation systems. 

Accuracy vs Cost

• Financial resources are tied up until model accuracy reaches an acceptable level.
• Visibility into the distribution of the customer’s data and eliminates some edge cases prior to deployment

While there is no silver bullet to reaching this ideal state, one key is to understand as much as possible about your customers – and their data – before agreeing to a deal

Keep Exploring!!!

Computer vision challenges - Data Challenges / Updates

• Dataset challenge - Easier to build age/gender with kaggle. If you need to build age gender for one country you can't find the dataset easily
• Implementation challenges - Segmentation models in the real world do not work 100% perfectly. Github code to expected accuracy takes time
• Post-processing - Color correction, noise detection, depth estimation multiple techniques we need to have to have a final smoothness
• Maintenance / Update issues - mediapipe was released in 2020, The architecture/datasets are not shared in the general forum. There is no updated model at this point

The learning list consists of

• Vision models awareness
• Vision image processing techniques
• Domain related aspects
• MVP product knowledge
• Making it work at scale 
• Research Landscape
• Data collection / techniques / models

Keep Exploring!!!