How Can a Company Remain Principled?

Talk is cheap. It is easy for any young company to say they are going to do things differently. How do we actually make sure our mission statement is lived out each and every day?

Align what keeps the lights on with our values

When push comes to shove, every company is beholden to what keeps the lights on. This is what makes a company possible, but it is also what can corrupt your mission.

For most companies this includes both investors and customers. By choosing to not take outside investments, we have protected ourselves from the former. But that is not enough.

We have aligned the products that we sell to match up closely with our artificial intelligence goals. Our products and offerings must always be things that are primarily differentiated through incremental steps towards artificial general intelligence that benefit humanity.

What are our values?

Our values are represented by how we should ACT and where we should AIM.

• All-Star team members only - Everyone must be inspiring to others
• Communicate frequently, honestly, and constructively
• Treat everyone with friendliness, empathy, and extreme respect

• Attempt the impossible
• Innovate by continuously challenging conventional wisdom
• Make the future an artificial intelligence powered utopia

Unlike most other companies, these values not only represent how an individual should act, but also the heuristics people should be considering when making more strategic decisions. Every decision must be weighed on how we "AIM" to make sure we are staying true to our mission.

How do we keep the lights on?

We commercialize our incremental success through software as a service (SaaS) offerings. We find areas of intersection where there is a large demand and a product that can only be solved by "real" artificial intelligence.

Specifically, that area is natural language generation. We view text as the best representation of knowledge, so creating services that write high quality, useful, or informative text necessitates ambitious and general artificial intelligence research.

Currently, we have two products, Article Forge and WordAi. Article Forge takes as input a keyword and returns an automatically written, high quality informational article about that keyword. WordAi takes as input an entire piece of content and turns it into completely unique rewrites that maintain the same overall value and meaning as the original.

We differentiate our products by focusing on generating text that is factually accurate and insightful. This is in heavy demand and our work in this field incrementally moves us towards our long term research goals.

We are succeeding in this effort - we have thousands of paying customers and are growing 10% month over month.

Our Research

As a startup, we must make tough decisions for how to pursue our research roadmap. Our starting goal is to be the research leader in a handful of areas, and increase our scope as we grow.

Factually accurate natural language generation

Our primary area of research is writing factually accurate long form text given very short inputs. For instance if asked to write about the "history of deep learning", we should write content that is coherent, accurate, and useful.

Our research is focused primarily on creating generative models that are able to write in a factually accurate way. There is also heavy work in how to find, categorize, organize, store, and retrieve all different types of knowledge.

We view factually accurate natural language generation as a way to incrementally work towards conveying intelligence. Writing accurately about any topic is a pre-requisite to writing insightfully about any topic. And we view being able to insightfully write about anything as a key feature in a generally intelligent system.

Information retrieval

A core part of creating factually accurate and useful content is to know what the facts are. We maintain our own index of the internet and focus on the best ways to represent, store, and retrieve this data. This includes research in areas such as information extraction, knowledge graphs, and efficient retrieval.

A key part of our research is making sure that our data is accurate, unbiased, relevant, and up to date.

Scalable machine learning

Machine learning models like transformers are getting larger and larger (tens of billions of parameters, hundreds of billions of parameters, and beyond!) and text lengths are getting longer and longer. With that there are a host of research and engineering challenges present to make sure we can continue to scale.

Our research covers topics like efficient attention and approximate attention, model distillation, and all other forms of improving memory consumption and performance for very large models.