Ian Mulvany

June 30, 2021

Are we seeing the promise of AI in augmented productivity systems?

We have known for a while that machine learning is good at very targeted specific tasks. The following is one of the best presentations on the topic: https://www.cs.princeton.edu/~arvindn/talks/MIT-STS-AI-snakeoil.pdf

The key conclusions in that presentation are:

- AI excels at some tasks, but can’t predict social outcomes.
- We must resist the enormous commercial interests that aim to obfuscate this fact.
- In most cases, manual scoring rules are just as accurate, far more transparent, and worth considering.

That latter is interesting. For many situations around decision trees, a simple scoring rule can be just as useful. 

We have seen real progress in things like perception, image detection, voice to text. 

But what of more cognitively complex tasks? 

A very cognitively demanding task is around software engineering. At the same time it is a task that is done with a lot of supporting formalism (the default patterns of use of a specific programming language, for example).

Microsoft, the owners of Github and Visual Studio Code, have been using the latest large scale language models to create tools to help programmers. By training these models on a huge corpus of programming languages they now have an assistant that can go much further than just recommending best practices for a coding language. 

The is now available for early sign up - https://copilot.github.com and some of our developers at BMJ have applied for access. I will be looking on with interest to see how this tools helps them. 

One of the side advantages of training on a corpus of computer code is that you can avoid baking in much of the sexist and racist bias that our language and literature is permeated with. 

We can imagine other cognitively complex tasks that could be helped by training on a dedicated corpus, e.g. writing abstracts, writing referee reports. But would we be confident that we could avoid biases in the training set? Another question to think about would be who could introduce a tool like that into the market that could get reasonable uptake? 

In any case, this new tool from GitHub looks like it might be the first broadly mass market AI assistant for reasonably cognitively complex tasks, that goes beyond just object detection.

Over the next few years applications like this are going to be more common across all of the domains of work that we do:

- spreadsheet creation 
- financial analysis 
- hiring 
- report writing 

I'll leave you with a final question - what skills might we want to focus on, so that we can most take advantage of when there is an opportunity to offload some of the schleppy mechanics of work that we are embedded in?