Aira Technologies – AI in the real world.

A good use case for a limited technology

• I have long held the view that under certain circumstances, today’s AI can improve performance to a superhuman level, and I think that the proposition of Aira Technologies, (see here) a small start-up, is one of them.
• Artificial Intelligence is often heralded as the greatest thing since sliced bread and in some instances, it is, but a lot of the time it massively overpromises and underdelivers.
• Nowhere is this planer to see than in autonomous driving where we were promised self-driving vehicles powered by AI in 2019 but still, they are nowhere to be seen.
• I do not consider vehicles with remote safety drivers to be self-driving as there is still a human element, but it does incrementally move the proposition closer towards true autonomy.
• I have long defined true autonomy where the passenger can be asleep in the back seat of a car with no human monitoring and have no fear for his life.
• The problem can be boiled down to one fundamental flaw which is that AI is not an artificial brain but a sophisticated statistical pattern matching system that is able to find correlations in data patterns.
• In practice this means that an AI system has no causal understanding of what it is doing meaning that it has no ability to adapt should things change.
• Put simply, this means that AI will do a great job in any task where the dataset that defines it is both finite and stable.
• This means that it is feasible to show the AI almost all permutations of possible events and that the dataset will never change.
• This is where using AI for autonomous driving hits a brick wall because the dataset of the road is both almost infinite as well as constantly changing.
• However, this also means that tasks where the dataset is both finite and stable, then a deep learning AI system can perform better than humans.
• Aira Technologies is a good example of this because it uses deep learning to optimise radio packet transmission between a wireless terminal and the base station.
• While the radio environment constantly changes as the terminal moves, the laws of physics for radio transmission remain constant making this an ideal use case for AI.
• Typically, when a packet fails to arrive at the terminal, the base station retransmits the whole packet and will keep doing so until it arrives or the software in the base station tells it to give up.
• Aria Technologies uses deep learning in the digital part of the radio to optimise the retransmission of packets and subpackets such that a substantial improvement in signal range and throughput is achieved.
• Aria has started with Bluetooth as it is open and easy to work with and it has demonstrated a 6db improvement with a bookended solution.
• Bookended means that its technology is in both the terminal and the base station and 6db represents roughly a 4x improvement in the signal.
• This is demonstrated in Bluetooth by achieving a much greater range before the sound quality degrades or disappears.
• This improvement in signal to noise can also be used to provide better battery life at the same range or greater throughput for the same energy budget depending on what the device maker desires.
• Starting with Bluetooth should enable the company to win some customers and generate some cash before considering the bigger opportunities of WiFi or even cellular.
• I have long believed that there are many such examples of where the advantages of AI can be used to their fullest extent without the very limiting drawbacks that deep learning presents.
• This is where I think that the real money in AI is going to be made over the next few years because other areas such as autonomous driving or digital personal assistants are badly suited given the nature of their data sets and hence are likely to continue to struggle.
• This is why outside of these specific areas, I think that AI will fail to deliver on its promises leading to lower investment and falling valuations.
• The hype and chatter on AI have gone pretty quiet since the pandemic and outside of some specific use cases such as this, I expect it to remain that way.