Gareth James

CEO

[Gareth James @garethdjames]

@SamanthaDixonMP please could you help, thousands of people across Chester affected by BT internet down since Friday. No resolution promised

[Gareth James @garethdjames]

@Go_CheshireWest when will the roadworks in Handbridge be finished? Road dug up a week ago and left, is causing traffic chaos!

[Gareth James @garethdjames]

@SmythsToysUK if you see an influx of orders for Colin the Caterpillar, this is why. They are a hit with Fox Red Labradors.

[Gareth James @garethdjames]

@StoryhouseLive is your internet working? Most of Chester internet is down, will come to Storyhouse if your is ok

[Gareth James @garethdjames]

In my next post I will discuss how NNs are used in chatGPT in the same basic way, but on a much larger scale. I’ll also dip into NNs that produce images from a text prompt, which surprisingly uses the same basic principles.

[Gareth James @garethdjames]

How does chatGPT really work?
I thought I would try and write a post on how chatGPT works, but in the style of ELI5.
There will be no mathematical notation or computer science and only one diagram.
The aim is to reach as many people as possible.

[Gareth James @garethdjames]

Since we had lots of past data and the resulting decision for each lender, a NN could be built they would predict the result - Moneysupermarket Smart Search was born.

[Gareth James @garethdjames]

We used a NN by feeding in all the past answers to questions that previous customers had submitted, then used the acceptance decision and APR as the items were were looking to predict (like recognising digits).

[Gareth James @garethdjames]

In a different project, we wanted to predict the likelihood that an individual would be accepted for a loan and the rate of interest they would be offered, without the customer applying for the loan (a mark on the credit score is added if you apply).

[Gareth James @garethdjames]

but making it impossible to use a simple image marcher to determine the value the image represented, a NN was used to solve this problem.

[Gareth James @garethdjames]

making it difficult for automated “bots” to read the price, the images were ever so slightly different each time they were displayed, not enough to cause a human viewer to notice any difference

[Gareth James @garethdjames]

Early on my career I used NN to solve two business problems for Moneysupermarket. Firstly during the great captcha wars of insurance comparison, insurance sites switched to showing images of the price of a policy

[Gareth James @garethdjames]

For example, in the late 80s early 90s speech recognition did not use NN, along with handwriting recognition (apple Newton…), these solutions largely failed when applied to the general population of all kinds of speech/writing.

[Gareth James @garethdjames]

This tech is not new, in fact it has been around for 40+ years, the advancement of readily available training data along with computational power has made them a viable everyday solution.

[Gareth James @garethdjames]

And that is how I’m a very basic way a NN is created.

[Gareth James @garethdjames]

Going back to the NN we have made to recognise hand dawn digits that has been trained on a sample of known images. This NN should then be capable of recognising the number value of an image that it hasn’t seen before, ie it was not in the training set.

[Gareth James @garethdjames]

though there are known “types” of configurations of layers and formulas for the neurons that are best suited for what you want the NN to do.

[Gareth James @garethdjames]

In reality, choosing the right number of layers for the NN, the range of weights and the formula (rules) for each neuron (each layer can and should have different rules) is an art and a science, much work is being done to automate this

[Gareth James @garethdjames]

The weights of the connections (input multiplied by the weight) is what is doing the work, this is the “intelligence”.

[Gareth James @garethdjames]

As the weight adjustment for a different digit image could throw the NN off for previous images. The training images are tested multiple times until it produces the correct result for all images (or a threshold, eg 99.5%)