Are we forgetting the Why in AI?

Reading the news recently I saw that Amazon have launched Q, the latest in the line of great chatbots that will change our world. Since ChatGPT we have had Anthropic, Bard and Meta AI to name but a few. To keep up with these Tech titans it made me think whether Elemental Concept should launch “BimBot”. This would be a great idea and really put us on the map, however the only problem would be is it wouldn’t do much. It might be useful to spit out the odd banal and slightly eccentric blog – but even that is questionable.

So you might ask why wouldn’t ‘BimBot’ be special (It would be because it’s a great name). The answer to this is pretty complex but in my simple view we can’t replicate the tech titans because of a combination of the following;

• The models used are Large models (the number of parameters of the models is way larger than what we have seen before),
• They have had years of reinforced (human feedback) and supervised learning, and
• They need immense amounts of computing power
• They have been trained on very large amounts of data.

They all have points of IP differentiation – e.g. are they using a Generative Pretrained Transformer (“GPT’), and they are trained on differing data.

It’s the bit about the data that they are trained on that is the actual reason for this blog .

At EC we are often engaged to perform technical due diligences (see https://www.elementalconcept.com/tech-due-diligence/ ) These may be requested by the companies themselves, potential investors or acquirers.

In the DDs we are essentially having a thorough look at the company’s existing technology and the plans to improve the tech to meet a businesses’ aspirations.

For us this is a privileged position as we get to see some amazing technology and understand how it has been put together. The outcome of the review is a report which details our thoughts on short, medium and long term technology improvements. It was the result of reading one such report that helped this blog.

As you would guess a number of the companies we look at have implemented some form of AI. We will of course look at how its implemented and the integrity of the data being used, but the thing that surprised me in the report was how incredibly perceptive our team was.

The question that was being raised in our report was whether the output of the models reflected the company’s (that we were reviewing) values.

Company Values and whether they exist and people know what they are is of course another debate, but for the purpose of this blog lets assume they are how a company presents itself to the world. They are the ethical or moral code that their customers and stakeholders understand they adhere to and influences their decisions.

In this case it wasn’t clear to us whether the models truly reflected said company’s values and purpose.

The reason we were asking this question was because our review looked in depth at the data that was being used to train the company’s AI. What we found was that the historic data that had been used for training did not reflect where the company wants to be. There is no doubt that historic data is biased, it’s what the outcome was historically. It isn’t a reflection of where things are now or likely to be. There have been a catalogue of failed AI models in Justice, Recruitment, Social Media and Finance. It isn’t quite garbage in = garbage out – its more simply perpetuating a bias.

What we did well in this report was to suggest other sets of data that could be incorporated. We also recommended ways that the AI Team could become more educated about the ethical AI dilemma and be more in tune with the Company’s Values.

The data being used is key and our understanding of the circumstances that produced that data has to be reflected in what we expect the AI to do. This means understand your business historically and make sure your team understand how the business and the environment in which it is operating is changing. Without this you won’t get any real benefit from using AI.

So bringing this back to the never to be launched ‘BimBot’. This can’t happen as we don’t have the IP and infinite computing power. Furthermore, there isn’t enough data (even if you include my English GCSE coursework) that will reflect the values I believed in then and how with experience my views will change.

What I can say though is at Elemental Concept we have a team that can help you review and improve the implementation of your AI and ascertain whether the outcome will be in line with your Company Values.