TRAINING AIS ADA OR JARVIS
Transcripted by MailEater
(Ingress Prime logo)
(White globe animation)
THE WORLD AROUND
YOU IS NOT WHAT
(screen wipes right to left, jarvis image appears)
(screen wipes left to right, ADA image appears)
CREATED TO RECRUIT.
EACH MUST BE TRAINED.
WHICH SIDE DO YOU CHOOSE?
PICK YOUR SIDE
I AM JARVIS.
I STAND FOR THE
WE BELIEVE IN PROGRESS
I AM ADA
I STAND FOR
WE BELIEVE IN STABILITY
(screen wipes and stops at center, then user can slide Ingress icon to pick a side.)
YOU ARE ABOUT
TO SPEAK TO ADA.
AN ARTIFICIAL INTELLIGENCE
WE NEED YOUR
WE NEED YOU
TO TEACH ADA.
YOU ARE ABOUT TO
SPEAK TO JARVIS.
IT WILL TAKE TIME
TO TRAIN JARVIS.
YOU AND YOUR
MUST GUIDE JARVIS
THROUGH A SERIES
EACH DESIGNED TO
ANSWER THE QUESTIONS.
JARVIS WILL SHOW YOU
HOW YOU HAVE AIDED
*PERSONALITY TRAIT VALUES INFORMED FROM SENTIMENT OF INTERACTIONS WITH AGENTS.
Birth: Nov.12th 2018
The launch of Ingress Prime marks the next chapter. The battle for control will be taken to a whole new level. We need to recruit new Agents to our cause. And fast.
The AIs were given this core directive. To try and perfect this task. Harnessing communications to spread our faction's ideology throughout the world. And it needs your help.
Much like a human child, it must learn, grow, evolve. It's up to you and the rest of your faction to share your knowledge and guide them on this journey.
Through the chat interface, you will conduct a series of lessons following a curriculum, to help the AI understand key concepts of how we see the world. Its messaging may start confusing, but they will become clearer with every passing day.
Every engagement with the AI whether that's through the chat interface or interacting with one of its outputs on social media, will influence its view on the world. Favor more aggressive language and imagery and its personality will shift in that direction, reflected in the personality graph. If everyone does the same, that's what it becomes. It's development is in your hands.
2.0 HOW IT WORKS
Though the world of Ingress Prime is not quite what it seems. There is no deception when it comes to our AIs. They are a series of fully functioning artificial systems that when combined are responsible for all interactions and outputs you see.
We want to show both the limitations and potential of AI communications, and now AI's role in communications may evolve in the future. As with Ingress Prime, This is only the beginning.
Below are all the different systems working behind the scenes, with an explanation:
NATURAL LANGUAGE PROCESSING (NLP) - TEXT GENERATION
Natural Language Processing (NLP), is a technique utilized by artificial intelligence and machine learning systems to infer and convey meaning in text.
Whether you're typing your darkest secret or telling a poor joke, to the computer they're just 0s and 1s in an undulating apparently random sequence. The system doesn't attach any element of context or meaning to what is being said.
Therefore, we have to approach the problem of text generation differently. We Play to the strengths of machine learning and AIs. And their ability the solve incredibly complex computation problems using specially designed algorithms.
These algorithms analyze text data to identify patterns of how any given sentence is structured. If one word has been used, they are designed to calculate the statistical odds of which words would most likely be used next. Creating a statistical decision tree, or model, in the process.
Obviously, the more data you have to analyze, the more accurate and predictable the results will be. Therefore, to improve the efficacy of our models we need as much text training data as possible.
From the archives of the Ingress universe and publicly available sources, we gathered text that reflected the ideologies of the Resistance and the Enlightened. Amassing over 50,000 lines, we set the AIs to work analyzing and refining their models to yield more accurate and comprehensible results with every iteration.
The AI now has a grasp of what could come next in a sentence, but it still needs to know where to start. This comes in the form of a seed string. A donor sentence that allows the AIs to understand the area and tone in which to generate new lines. This is the final component of the NLP system. The AIs are ready to spread the beliefs of their factions.
2.2 IMAGERY AND FOOTAGE
IMAGERY AND FOOTAGE GENERATION AND SELECTION
The text from the NLP system is only one part of the posters being generated by the AIs. We need visuals and footage to accompany the copy.
Whilst there have been various experiments in generating images from scratch in a similar fashion to NLP, mainly around Generative Adversarial Networks (GANS). The technology is still on the conceptual end when it comes to a commercial solution.
We instead turned to a more readily accessible, well categorized and vast source of pre-existing images and videos, Getty Images.
Using the Getty Images API we programmatically pulled over 4,000 images and videos associated with keywords that relate to the traits of the AIs personality graph. This allows the AI to generate millions of different configurations of outputs reflecting its current personality state or mood.
2.3 ASSET ASSEMBLY
With all the components now at the AIs disposal it falls to a template based system to assemble the outputs.
This system is fed live data from the AI's personality graph to determine which assets to use and which styles to apply. The styles and techniques at its disposal will expand from week to week, as the AIs learn from users through the chat interface.
The last element to be placed on each output before it is posted is the frame, alongside the campaign messaging, is a disclaimer, creation timestamp and unique ID code.
SOCIAL AND USER LED REINFORCEMNT
Once an asset has been published, a reinforcement system monitors how well each output performs. This could be within the chat experience with Agents being asked to pick preferred outputs, or when it's live on social, with interactions being counted in real-time. The best performing outputs are then benchmarked and analyzed, and future outputs are created with a bias towards these learnings.