Bottlenecks for AI

AI applications: automation everywhere

All activities are susceptible to be automated, if there is enough data to train models.

The automation equation: 🦾 Automation probability = 🎼 Coherence * 📊 Data

Coherence: similarity between cases, all configurations are known. We can improve coherence by controlling the environment (ex: battery hens, factory, autonomous driving on private property, prefab in construction)
Data: the less coherent the environment, the more data we need to describe it (e.g.: driving in Florida on a private property, driving in US, driving in Toulouse).

Source

WORKING PAPER The Future of Employment & Q.Chenevier's own opinions

Perception & manipulation

2 unsolved challenges:

Perception in unstructured environments: houses, fields
New or irregular objects handling, with a soft grasp and through learning of mistakes

The human hand is a wonderful tool, full of sensors:

Position
Toughness / Roughness
Heat
Humidity

Manipulation learning experiment (from 19:48 to 21:11)

Social intelligence tasks require several hard-to-automate skills:

Recognize emotions
Being empathetic with our interlocutor ("compute its state of mind")
Exploit this context and the past interactions

"Recognizing emotions is a challenge, but the ability to answer in a smart way to this information is even harder."

Nevertheless, Woebot Health: Relational Agent for Mental Health, a CBT (Cognitive Behavioral Therapies) chatbot, demonstrates shows that many CBT therapists' sessions could be automated.

It doesn't replace the therapist, but can reduce significantly the needed amount of sessions to do a therapy.