Three years ago, Cal, Shahrad and I were sitting at a little off-centre cafe in Parramatta when we decided to go for it.
Looking back, we were naive to the extreme in knowing exactly what it was, but we knew that Artificial Intelligence stood at the intersection of neuroscience, mathematics, computer science, philosophy and psychology and that it would be one hell of a ride.
Having discussed the idea over many coffees and other more alcoholic drinks, we knew that something like Remi had to exist. Sure, there were pure A.I Research companies like Deepmind, but there wasn’t a single A.I. company out pushing the boundaries of A.I. Research, with an absolutely crazy long-term mission, and then turning their cutting edge breakthroughs into kick-arse, beautiful solutions and products.
The crazy long-term mission? General Artificial Intelligence.
So in the Spring of 2014, we threw caution aside and went for it.
The reason we went with Remi.ai was because we sincerely desired to not be clumped in with the unfriendly companies that litter the tech startup space. Our long-term ambitions were human-level Artificial Intelligence, therefore we wanted a friendly human name. Another contributing factor was that Remi.com was quoted at $130,000.00 whereas Remi.ai was available at $19.90 at onlydomains.com.
In 2014 there wasn’t a wealth of companies in the A.I. space and there wasn’t the huge volume of papers or training material on deep neural networks out there. Nonetheless we threw ourselves into to the task of learning and building everything available. We started with simple neural networks and RNNs, tested on a wide variety of tasks like predicting sales revenue and inventory movement. We explored image recognition for a few different projects, and even won a kaggle competition along the way.
But after a few months it became very clear to us that Reinforcement Learning was the key to truly advancing A.I. and developing Agents that learned like Humans.
For any readers who are inexperienced in the field of Artificial Intelligence, there are currently three different sub domains. Supervised Learning, Unsupervised Learning and Reinforcement Learning. I won’t give much detail on the first two as that can wait for another post. But Reinforcement Learning can be defined as building A.I. to learn as mammals do. Mammals have a simple, yet extremely effective learning mechanism in their ability to try something, learn from the consequences and most interesting, pass their learning on to their peers and offspring. This learning mechanism is heavily informed by internal rewards that have evolved over millennia and Reinforcement Learning gave us an opportunity to develop artificial intelligence that could mimic this, obviously in a rudimentary manner to begin with. Life isn’t entirely about reward and punishment, there are higher cognitive functions that arguably sit about the reward mechanisms in the mammalian brain. But it sits as the bedrock of our intelligence and is an important step toward human-level A.I.
The second fundamental point is that many of the reinforcement learning algorithms are general purpose algorithms, in the sense that the same algorithm can be applied to many different tasks with oft incredible results. If you think about it, humans are the finest example of this ability to learn any task through trial, error and practise, as long as there is a sense of reward and progress in the training period. You can do at least some, if not all, of the following: read, write, tie your shoelaces, do long division, ride a bike, play chess, snowboard and countless other skills. This all comes from one or perhaps more than one learning algorithm in your brain.
And just as you possess the ability to learn a wide variety of tasks, the algorithms we work on can be applied to numerous tasks: from driving cars, to playing computer games, to learning how to understand language and many other applications.
This became the entire focus of our Group’s research. We started testing reinforcement learning agents in both artificial environments (computer games) but also real world problems. We built Agents to play the Space Invaders, Pong, and other computer games. We built an Agent that could rearrange code as it saw fit to meet a Users request, streamlining any chat interface by removing the need to make requests of individual API’s, one at a time. We built a group of Agents that were in an environment that developed their own language to communicate about their environment, learning how to warn of the arrival of a Threat Agent, in much the same way meerkats do. We built an Agent that could kick arse at Google Adwords bidding campaigns and massively increase ROI for companies and we also built a really cool language agent that learnt by itself how to understand requests from users to complete tasks.
And throughout all this, so much has changed. We’ve learnt more than any current university degree could provide. In this time, A.I. has gone from a small, offhand subject of interest amongst a few to being one of the most talked about technologies at present, heralded as the way of the future, not for industry but society at large. At Remi, we have gone from testing small algorithms on small games such as tic tac toe, to building agents for Enterprise Companies that control the spend of millions of dollars each month.
Yet so much remains the same. We’re still reading a lot of books on childhood neurodevelopment, we’re still drinking too much coffee, we’re still 100% bootstrapped, and we still have the same crazy long-term mission that really excites us.
These next couple of years are going to be really damn interesting for Remi. We’ll be launching two products in the coming year, we’ll be continuing to work on a Semantic Network Reinforcement Learner, which is in its early days and is showing extremely strong promise in advancing A.I. Semantic Understanding. We are in talks with a VR Dev Company to build an A.I. inside a VR game. There are also some more hush hush projects which will hopefully see the light of day later this year too.
There really is no experience like starting a company. Finding a group of people who are crazy smart and motivated and throwing all your collective intelligence and willpower into trying to build something that no one has done. It has so many highs and lows. Hours upon hours of staring at equations and papers trying to get your head around a new idea. So many nights where you’re trying to find the energy to keep working. But then there’s the moment when an Agent starts to work, when you see it surpass a human at something, that cautious excitement that it might actually be working, and lastly those beers with the team to celebrate the fact that it’s working.
There really aren’t many experiences out there that compare.
Stay tuned on the Remi AI blog as we build out the complete supply chain offering!
Or, if you're ready to start seeing the benefits of A.I-powered inventory management, start the journey here.
Who are we?
Remi AI is an Artificial Intelligence Research Firm with offices in Sydney and San Francisco. We have delivered inventory and supply chain projects across FMCG, automotive, industrial and corporate supply and more.
Want to know more? Sign up to our newsletter for the latest information from us and other knowledgeable folk in the market.