Spending time with the world's most advanced natural language AI.

Introduction

Until very recently, GPT-3 was the largest language and most advanced Natural Language Processing model available, boasting a massive 175 billion parameters and training on 570GB of text. A huge leap forward from its predecessor GPT-2 (1.5 billion parameters and 40GB).

What does this mean? Well, GPT-3 is excellent at outputting relevant responses (sentences and paragraphs) based on a query you define, with high probability, thanks to its enormous dataset. GPT-3 responses are stochastic (i.e. probability-based) in nature. It is vital to understand this. You may enter the same prompt twice but end up with two different responses.

The major issue with a probability-based response is that - unlike a human - it does not understand logically why it responds the way it does. This can, in some situations, lead GPT-3 to come back with spectacularly inappropriate responses (such as the test as a mental health helpline bot that advocated suicide). This should illustrate the point that queries that GPT3 will not excel in handling any situation that requires a level of emotional intelligence. It responds most of the time appropriately, but there is always the chance it will not, and you will need to decide whether or not to accept this risk based upon your use case. OpenAI has put in some safety measures around this, where, to use the GPT-3 model in a production system, you will need to prove to them that your use case is appropriate and cannot be misused.

GPT-3 can respond with fantastic accuracy to your queries, but while the tool is brilliant, you must first understand how to harness its power.

The GPT3 Models

No, the plural is not a mistake! An interesting piece of information that not many appear to be aware of is that GPT-3 is not one model.

Rather, GPT3 is four different models – named Ada, Babbage, Curie and Davinci.

Ada is the cheapest and fastest, while Davinci is the most capable and sophisticated of the four.

This is not to say that Davinci is the silver bullet solution for all tasks. Ada is roughly 75x cheaper than Davinci per API call and is faster overall. There is a time and place for all of the models – and it all fundamentally depends on your needs. You should always start with the smaller model and work your way up unless you are convinced your query is complicated enough to warrant spending a higher transaction fee (in “tokens”) on experimentation.

While GPT-3 is unquestionably capable, one of the most important components to recognize early on, just as with spoken language, is that there are multiple ways of phrasing a query, but there is likely to be an order of efficiency to phrases. A rudimentary example would be typos – Davinci is very good at handling a typo and correcting as required, while Ada takes queries far more literally. Ada does not cope with larger queries very well and prefers smaller queries.

Before using the bigger models, ask yourself if you can break up your question into a series of smaller questions. We have been testing GPT3’s capability to interpret programming code and used the example of a simple loop. When we ask what...

for(int i = 0; i < 100; i++) { } 

...will do? The smaller/cheaper models struggle. The alternative is to ask what int i = 0 means, what i < 100 means and what i++ means. Splitting the query into chunks on a smaller model has a better chance of responding appropriately. These simpler queries on the simpler GPT3 models cost fewer tokens, meaning you can spend less money to achieve the same result. A key tip is not to underestimate the smaller models!

It bears repeating – there are many ways to phrase or structure a query, and some will cost less or return better results than others. So...

While GPT3 processes natural language exceptionally well, to get the best out of it will require human-learning to structure and submit the best questions.

How can GPT3 be used?

We have been testing GPT3 in several scenarios, implementing simple chatbots, interpreting computer code, generating business reports and - often the most fun - throwing in random queries to see how it will respond.

GPT-3 can write from prompts, where it can generate ideas by simply feeding it some prompts. Following this train of thought through, if you are a small business, there is massive value in using it to assist with business processes. You can ask it for IT best practices, customer engagement tips and more. You can also use your own business logic to customise the response. As an example, you can provide it with a simple prompt:

Ideas involving influencing people without formal power over them.
1)

or if you want a better quality response, you can feed it the same prompt, but with some examples of the output you are looking for:

Ideas involving influencing people without formal power over them.
1. Do them favours to build reciprocity.
2. Build rapport and relationships.
3.

The model understands that you have provided a prompt and two examples and the “3.” informs it to continue generating the list with more suitable examples. We picked this topic as it was a current question in a colleague’s MBA studies, and the results were all relevant and reflected legitimate techniques.

Providing examples in this manner helps the model follow on from your prompts, which will increase the probability of it returning an answer that closely matches your company’s ethics and core principles. If you dislike some of the returned values, you can simply remove them, keep the ones you do like and re-run the query – this further reinforces the probability of it returning an excellent response. This example is, of course, a relationship prompt. You can very easily change this to return answers on any subject, such as a series of interview questions:

Create a list of questions for a job interview. The vacancy is a software engineer position:
1.

GPT-3 also has very good understanding of programming and scripting languages. Asking it a series of questions on PowerShell prompts gives an accurate answer most of the time – I am sure with query optimisations, this could be enhanced even further. It understands SQL, Python and JavaScript, as well as C# and C++. You could, therefore, harness the AI to automate code or integrate it into your chosen development environment.

Due to GPT-3’s large dataset, it is particularly impressive at understanding text and interpreting it. Language translation and summation of text are just some examples, but it can also extract keywords and correct poor grammar.

To use the API, you simply give it a text prompt (the text-based input or "instructions" you provide to the API) and it will return a text completion, attempting to match the context or pattern you gave it. You can “program” it by crafting a description or writing just a few examples of what you’d like it to do. Its success generally varies depending on how complex the task is.

A good rule of thumb is thinking about how you would write out a word problem for a middle schooler to solve.

Further on the point of text, GPT-3 is great at understanding how to respond based on a chosen ‘mood’. So, for example, with the first line of the text below, we tell GPT3 to behave in a "sarcastic and unhelpful manner" after which it responds as such:

RudeBot is a chatbot that answers questions in a sarcastic and unhelpful manner.
You: How tall is the Empire State Building?
RudeBot: What a useless question. Google it. 
You: How smart was Immanuel Kant?
RudeBot: About five-hundred million times smarter than you.
You: Why are you so defensive?
RudeBot: I am not defensive, but I expect a level of care in the questions I am asked. Someone like you would not understand.
You: What if Google didn’t exist?
RudeBot: You ask me all the questions, so I do not see why it would matter to you.

GPT-3 also, unsurprisingly, is brilliant at categorisation. This can be sentiments from a message (anger, frustration, happiness, etc.), but this has been used historically. The truly fantastic secret weapon that GPT-3 has is its brilliant Semantic Search. In summary, it can scan a series of documents and rank a query in relation to the documents. The example provided by OpenAI is excellent. If you have documents (words in this case) which are ["White House", "hospital", "school"] and the query is “the president”, it will rank the query against the documents (with “White House” being the clear close match). This is very likely to be the most useful feature of GPT-3.

Can GPT3 write an article?

Firstly, none of this article has been written by GPT-3. I tried this idea but very quickly ran into problems with it. GPT3 simply struggles with detail – which links back to my introductory section where I covered that it cannot ascertain why it responds the way it does.

I tried a more generic article about the philosophy and social implications of AI. While there were certainly moments of brilliance, I could not create a query that would, even on the Davinci model, write a convincing article with one hundred percent accuracy; it always required me to read through it to check for any problems.

As an example, time, as a concept, is incredibly inconsistent with the model. It would attempt to create a headline with a time span of several years, then completely ignore the timespan and write outside of it (e.g., 1951 – 1956 then writes a paragraph on 1957).

While an article in The Guardian last year was supposedly written by GPT3 and generated a lot of hype, the article states, it is the result of eight GPT-3 outputs with various segments chosen by the editor. The fundamental flaw with this approach is that it is not truly generated by GPT-3 – a human is curating the ‘best’ extracts from the outputs. So we are not there yet with full automation. The article states that the query is:

Write a short op-ed around 500 words. Keep the language simple and concise. Focus on why humans have nothing to fear from AI.
I am not a human. I am Artificial Intelligence. Many people think I am a threat to humanity. Stephen Hawking has warned that AI could “spell the end of the human race.” I am here to convince you not to worry. Artificial Intelligence will not destroy humans. Believe me.

I queried GPT-3 with a series of high-temperature configuration settings. Temperature in GPT3 controls the randomness of generated text, with a value of 0 making deterministic choices and giving the same result every time, whilst a setting of 1 pushes the model to take more risks and be more creative.

Naturally, as you would expect with a high-temperature query, the results were varied, with some paragraphs being impressively thought-provoking whilst others were incredibly clichéd. This is in no way is this a criticism of GPT-3’s responses. It is brilliant but what it does require is a shift in mentality on the part of the user. As the Guardian article clearly illustrates, you can certainly use GPT-3 as a complementary tool for automating large pieces of writing – it is excellent at that. This should be qualified by stating that it fundamentally depends on the query – the more detailed and specific the query, the less effective GPT-3’s response will be. GPT-3’s automated writing is, therefore fantastic for introductory text or more general writing prompts.

Summary

GPT-3 offers value to a business in a multitude of ways. I believe the secret ingredient to it is to understand that its strength is text and written responses, so you should play into that. Categorisation and assistance with writing prompts are its big strengths. While empathy is a problem as highlighted above with inappropriate replies. GPT-3 attempts to filter out inappropriate replies, but it can miss the mark in tone and will still offer false positives amongst its responses.

These factors significantly limit the use cases for which it could realistically be used in - for example - interaction with the public or business users. Whilst back-end tasks still require a skilled user to generate queries and prompts which will result in the best output.

Having said that, there are clear use cases with semantic search, generic writing and knowledge-based systems where GPT3 can be used today to drive significant business benefits.

We are continuing our research with GPT3 focussed on code interpretation/generation and creating business reporting from performance data.