Insights from our Conversational-AI roundtable v.1
DataSeries |VRT300620
In June, DataSeries, an OpenOcean led initiative hosted a Virtual Roundtable about the future of “Conversational AI”. The session was moderated by Mike Reiner, Venture Partner OpenOcean.
INSIGHTS GATHERED:
FUTURE OUTLOOK:
4 notes/requirements on our journey to maximize total value out of virtual assistants/chatbots.
- We are on the cusp of creating a global Virtual Assistant paradigm that will be able to rule entire processes instead of having thousands of spread out assistants specialized in their fields. Having 10,000 or even 100,000 specialized domain assistants is not the most effective way to reach a truly groundbreaking level within conversational AI that is capable of navigating through complex ecosystems. We’re trying to create a conversational assistant that can delegate specific tasks for you — that allows you to compute in many more ways.
Build better voice apps. Get more articles & interviews from voice technology experts at voicetechpodcast.com
2. The ability to access the assistant from all devices (same cognitive load). In the current market, we still have limitations on what we can say to our assistants. The market seems to be getting crowded with conversational ai solutions, however, each solution has its own set of conversational rules. The moment the user is required to talk to more than one conversational ai solution confusions will appear.
3. The ability to access every service in a democratic way and being able to break down the barriers that are currently predefined by companies that are trying to integrate additional use cases in an effective manner. What is needed is that third-party services can dynamically adapt your de facto standard and therefore enrich interaction capabilities. Hardwiring all these various logics is very complex. It would be more effective to have a free-flowing and task orientated assistant that can interact and integrate services from the entire web. Currently, this is being solved by large established players such as Alexa, Google Assistant, Bixby, etc. to give third-parties exactly the same tools and platform access that they have used to internally build on their domain capabilities (usage of fully-fledged capabilities). There should be room to compete for the users’ affection and attention. Today this space is very fragmented.
4. Personalisation that allows to customise the experience of use cases. Samsung is giving “enrichment-tools” to their third-party developers to solve various speech problems in order to avoid the aforementioned user confusion. We already have it today with Siri, Alexa, etc. but they are not really good yet. Users are restricted in what they can say and don’t want to feel stupid — so they tend to be more conservative rather than truly having the ability to challenge their assistant with a decent comeback. Enabling user Discovery is a trend that is often not talked about.
Challenges that need solving are improvements in dialogue, having the ability to combine services from different data sources automatically, and by having a platform that every developer buys into. Trend: Siri, Alexa, Google Assistant, Bixby, etc. all live in the cloud and where the processes are being digested, but Adam from Samsung believes that there will be a trend towards moving some of the processing down to the device.
Example: If you are trying to have a hundred million users building a different language model of the services they want; e.g. Magic and Samsung services with your VA, then this is quite specific to the user. So, if every user will customize skills for their VA. Doing this in the cloud means that we need to do machine learning computation for a hundred million users and every time a new service updates its capabilities, then a large chunk of these users will need a new version of the service. However, if things are more from a client perspective, you can use their computer resources to solve particular requests. Now it makes sense that I can have a different language model constructed just for me running on my phone or home server (even distributed server).
ADDITIONAL INSIGHTS:
A truly horizontal assistant?
We can lock in an achievement the moment the VA will manage to help you with your 100 closest tasks and personalise your areas of expertise. In design, it will be similar to an App Store where you will be able to upgrade the VA with use-case specific knowledge. This requires a dynamic program generation where the AI is able to write the logic for every single use case.
How far are we away? (Estimates vary from very close to decades)
We are probably 5–10 years away in order to create an assistant that has human-like conversational skills. One vision to perhaps achieve this is to create a virtual version of yourself that will truly understand you and learn from your decision-based patterns (similar to the “home- egg” that we have seen at the Black Mirror series).
Dialog-Management is a classic challenge that we are currently seeing. Not just the individual exchanges but also the whole conversational aspects dealing with a single intent. We often look at the intent from a bot's point of view and not the human perspective (e.g. how does the bot marshal its decision to achieve this requested intent).
Natural Language Expectation — being able to create sentences from scratch moving to a more grammatical based natural language understanding. There seems to be not too much progress in this space as Machine Learning as a whole is sucking up all the PHD talent, not allowing the talent to look at more niche segments.
AI being able to write specific use cases for you and that integrates the relevant data. Once the intent has been set, all the logic requires hardwiring. Creating a platform that does this for you — will emerge as a definite winner in the space.