2 A ITM 
Researching the mind, what it does, and where it's headed.

Georgia O'Keeffe | Ladder to the Moon | 1958
The VictorTM Project - AI vs 2AI

AI is a very hard problem. The best analogy for human progress in AI thus far is a ladder to the moon. Adding rungs feels like progress, but these may never suffice to reach the actual goal.

At issue is that all traditional AI methods model historical data to create an expert system. The goal is to be fast, reliable, and accurate. Tactical, precise, and consistent. Predictable. For passive classification tasks these encyclopedic solutions are ideal, and very useful. Successes abound.

However, when humans see these AI solutions in action, no matter how impressed we are with the results, we hesitate to call them AI. They seem engineered, like a magic trick. Seeing behind the curtain dispels the illusion, and the goal is forever out of range.

At 2AI Labs we believe the future of AI will ultimately hinge upon its capacity for competitive interaction. That ability is the primary mechanism by which humans learn from each other, improve over time, and innovate. We believe only this approach brings the moon within reach.

Jules Verne | De la terre à la lune | 1865

To see the gap, consider what happens when passive AI is given a competitive task. It does not suffice to model more historical training data. If you try that against a real opponent, your responses will eventually be anticipated, bluffed, gamed, and defeated. The mind of the opponent is part of the exercise. To win, you must take into account not only your estimate of their skill level, but also their estimate of your skill level.

Our research suggests that competitive interaction involves selecting something other than just "good moves". Instead one must select a sequence of solid tactics that maintain an inscrutable underlying strategy. Accordingly, we have developed an entirely new theory of algorithms, from the ground up, specifically for interactions with potentially sentient adversaries. The name of the theory, and the AI itself, is VictorTM.


Examples of passive vs competitive tasks:

  • Solving captchas. [Google]
  • Mastering trivia games (Jeopardy) [IBM]
  • Mastering scroller video games. [Atari] [Mario]
  • Playing board games with maximum odds of not losing. [IBM] [Google]
  • Playing board games with maximum odds of winning quickly.
  • Playing combat games with maximum odds of not losing. [OpenAI]
  • Playing combat games with maximum odds of winning quickly.
  • Winning One-Card poker. play
  • Winning Rock-Paper-Scissors. play.
  • Playing optimal low-limit poker. (fixed raise amounts) [Cepheus]
  • Playing pot-limit or no-limit poker.
  • Making small wagers on sporting events.
  • Setting or moving a wagering line.
  • Daytrading, or managing a small investment portfolio.
  • Managing a large investment portfolio. (Large enough to move the market.)
  • Estimating the market price of an asset. [Zillow, Airbnb, Invinio]
  • Negotiating the price of an asset (car, house, work of art, business, etc)
  • Estimating the fair market value of goods and services.
  • Negotiating the price of a contract (salary, compensation, service agreements).
  • Preparing citations for a lawsuit.
  • Negotiating a legal settlement.
  • Split testing optimal logos and sales letters.
  • Creating a marketing plan.
  • Auto-aiming weapons.
  • Planning a military incursion.
  • Indexing a fixed archive of documents. [Baidu, Bing, Google]
  • Language translation of documents. [Google, Microsoft]
  • Interpreting voice commands. [Apple, Google, Microsoft, etc.]
  • Predicting customer preferences. [Amazon, Netflix]
  • Selecting relevant click-through advertisements.
  • Image matching. [Wolfram]
  • Age guessing. [Microsoft]
  • Face recognition. [Facebook]
  • Emotion recognition.
  • Screening humans for drunkenness, illness, or bad demeanor.
  • Screening humans for bad intent. (security guards, bouncers)
  • Diagnosing illness from medical images, charts, and tests. [Enlitic, IBM]
  • Diagnosing illness from a patient interview.
  • Prescribing antibiotics.
  • Prescribing pain medicine.
  • Interpreting polygraph exam data.
  • Conducting a polygraph exam.
  • Interpreting meteorological data. (predicting weather)
  • Interpreting geological data. (predicting earthquakes, finding oil)
  • Interpreting polling data.
  • Reacting to polling data.
  • Fending off a zombie (DDOS) attack. [Cisco]
  • Fending off a live hacker attack.
  • Estimating your credit score. [Experian]
  • Deflecting credit card fraud. [Kount]
  • Warehouse automation. Automated parcel delivery. [Amazon]
  • Predicting traffic jams. [IBM]
  • Driverless aircraft/drones.
  • Driverless ships. Docking. Ocean crossing. [Rolls Royce]
  • Driverless trucks. Parking. Highway driving. [Mercedes]
  • Driverless cars. Parking. Highway driving. [Google]
  • Driving in traffic. Knowing when to speed up, slow down, merge, or yield. [Google]
  • Delivering a speech, or auditorium lecture.
  • Taking a standardized test. [Allen Institute]
  • Preparing an exam.
  • Teaching a class. Knowing when to speed up, slow down, ask, or allow questions.
  • Diplomacy. Knowing what to say, or not say, and when.
  • Conversation. Knowing when change topics, pause, listen, or interject.
  • Retelling a joke.
  • Being comedic. Timing. Knowing your audience.


2AI © 2018