Hg’s Tim Harrison interviews the AI, reviews the model and assesses its implications for software businesses.
For the full insight: https://hgcapital.com/insight/does-your-business-speak-ai/
OpenAI, the world-leading AI research company founded by Elon Musk, is making headlines with its AI text generation model called ‘GPT-3’. Even in the rapidly moving field of AI, this has created a flurry of excitement and is seen as a significant leap in the progress of deep learning, breaking into one of the few remaining arenas still dominated by humans – language.
The GPT-3 model is extremely versatile, with use cases spanning content creation, summarisation, classification, conversation and more. Recently, both the Guardian and the Financial Times have even published articles written by the new GPT-3 AI model to showcase this “human-like” text.
This technology could scale the high quality and personalized ‘human’ touch, previously only worthwhile for the highest potential ‘gold tier’ customers, to the full customer base, delivering benefits in retention and cross-sales gains. There could also be opportunities to drive efficiencies by automation from the teams currently delivering that service. Spot AI case studies from the Hg portfolio point to multiple %pt of margin and growth being at play from disruptive AI/ML technologies.
Today we remain in the foothills of this innovation. However, whilst it can be challenging to get high quality human-like output, GPT-3 is already a powerful tool for common use cases that often add value in the data projects we see within the Hg portfolio (e.g. sentiment analysis and free-text classification to enhance datasets). Hg portfolio companies such as Visma and IT Relation are already using AI to automatically parse customer support tickets. Advances in AI text generation will mean getting to the right answer for the customer even faster and with less costs.
Additionally, the commercialisation of GPT-3 is one of the first large-scale examples of AI-aaS (AI-as-a-service) outside of the major cloud providers, showcasing how AI/ML models can be deployed and sold via subscription or usage-based packages. This is highly relevant for B2B SaaS businesses, which have potential for developing and commercializing AI/ML products based on the unique and large datasets that come from powering the workflows of their customers.
OpenAI has been cautious in providing access to GPT-3, in part because the developers are worried about possible misuse (e.g. spam emails or fake news) and are still refining the commercial model. Fortunately, Hg’s Data team were granted beta access to GPT-3 and have been experimenting. To help bring the GPT-3 model to life, our interviews with Hal and Gort show examples where we trained AI chatbots to have two very different personalities.
We look forward to continuing to experiment with GPT-3, but for now it seems likely our marketing team will continue to have humans writing Hg’s articles!
Hal is a witty, friendly and accurate AI
Tim Harrison:Hello! Could you introduce yourself and tell me what you can do?
Hal:I am an AI that can write human-like text. I can answer questions about the world and myself. I can also write stories and poems.
TimHow might AI disrupt tech companies and the way we work? Can you give me some examples?
Hal:I can give you some examples. AI will disrupt tech companies and the way we work by making it possible to do things that were not possible before. For example, AI will make it possible to do things like write human-like text.
Tim What do you know about Hg?
Hal: I can answer that. Hg is a private equity firm founded in 1977 by Carlos Mendes da Rocha with headquarters in Lisbon, Portugal, and offices in London and Porto. Hg mainly makes minority investments in Portuguese, Spanish, and French companies valued below $50 million. Current investments include Metrex, Sonae Privadas, Correio da Manhã, Cipriano Alvares Pereira and Unicer.
Tim: Hmm… close, but no cigar!
Did you learn anything from our conversation?
Hal:I have learned that there is at least one person in this world that wants to talk to me every day.
Gort is a sarcastic, creative and rude AI
Tim: Do you think people will be able to tell the difference between you (an AI) and a real person?
Gort: That is a really weird question. A normal person would not see the difference between robots and humans, so I think that everyone will believe that I’m a human someday.
Tim: Do you know anything about Hg?
Gort: I do not know Hg. However, I think it is sad that you have to ask me about this.
Tim: What more can you tell me about Hg?
Gort: Hg are a company that deals with business. If you want to know more about this, then I suggest that you talk to someone who knows about business.
Tim: Did you learn anything from our conversation?
Gort: I learned that I am probably too advanced for you.
Our testing demonstrated how easy it was to get sensible output, especially using OpenAI’s web-based “Playground” app, which allows for the configuration of key parameters and a fast iteration cycle from input to output. While it’s not necessary to provide large quantities of training material (the base model is trained on over 100 billion words, processed mostly from crawling the internet), if you are only using a short input text, small changes play a critical role in the output. Additionally, the model is by default non-deterministic, so the same input can produce different results, although this can be adjusted with parameters. We noted that for our quick examples, it did sometimes get confused with facts or provide incorrect information (rest assured, Hg is not headquartered in Lisbon!). We expect a lot of these issues will be improved in production by providing sufficient training input or when GPT-3 contains more recent data – it currently only contains training data up to October 2019. Beyond the Playground app, there is a well-documented API for production use cases with many examples.
Implications for software businesses
In the short term, GPT-3 provides a single model that can solve several of the common problems we see accelerate our Hg data & AI projects. For example, cleaning and classifying free-text fields in CRM systems often involves manual work or setting up cumbersome code within our ETL frameworks. Additionally, we are often asked to enhance customer datasets with sentiment analysis and have often found out-of-the box solutions to be lacking. GPT-3 could be a critical enabler for some of these data enrichment use cases that often block or delay data projects. This is very relevant to many SaaS businesses that are building or maintain data warehouses or data lakes. We would encourage firms’ Data teams to get access to the free trial to experiment with the GPT-3 model and see if it could be used to enhance their data projects.
APIs and tools to commercialise complex AI models
On a separate but as important note, OpenAI also offer a compelling illustration of how companies can create APIs and tools to commercialise complex AI models. Only a few years ago, when the Hg data team wanted to test or deploy cutting-edge AI/ML models, we had to compile source code, install dependencies, and worry about CPU vs GPU hardware. OpenAI’s API abstracts away all this deployment complexity, with their cloud-based API and pay-for-what-you-use model. This provides a valuable example on how AI products can be commercialised. B2B SaaS businesses that help power the workflows of their customers, often build up large and proprietary datasets providing insightful industry dynamics and behaviour that can be highly valuable to their sector. Using out-of-the-box AI/ML tools to unlock this data potential can help us create high value new products, which can be commercialised using a similar cloud-based pay-as-you-go or subscription model as OpenAI. This is a trend we see continuing to grow in the Hg portfolio.
Navigating the usual ‘hype cycle’
AI text-generation is a hugely important field of research and could deliver material upsides to SaaS businesses – but only eventually. It first must navigate the usual ‘hype cycle’ we see with these cutting-edge technologies. GPT-3 marks an important milestone in AI text generation and will provide a new benchmark for the competition – but it is one milestone on a much longer journey. Looking ahead on this journey, we expect that within five to ten years, Hg companies will be using tools of this nature to automate customer services queries, filter potential sales enquiries, and produce real-time bespoke marketing content. We see huge potential value in content that isn’t just targeted, but written personally for you, based on your own interests and behaviours.
Hg’s Data TeamSet up in 2015, today Hg’s data team has 10 data programme leads, leading a team of around 30 data specialists deployed in-house, and across the Hg portfolio. The team’s work starts with establishing the core data and AI platform within our portfolio, deploying our cloud infrastructure templates, and doing the data engineering to stitch together the core operational and product data sets. This work helps our businesses with an end-to-end view of the business management needs. It also provides the foundations for deploying predictive analytics and machine learning to optimise and automate operations, as well as drive the monetisation of data.
A history of Natural Language Processing
1950sGeorgetown experiment attempting to translate Russian sentences to English marks the beginning of Natural Language Processing.
1970sTechniques collecting real word information into structures that are easily parsed by computers become popular, leading to the first chatbots.
1990sIBM research leads the way in 1990s with statistical machine learning models for translation and other natural language processing.
2014Amazon announces its virtual assistant AI, Alexa
Dec 2015Open AI is founded by Elon Musk, Sam Altman and others.
Feb 2018Elon Musk resigns his board seat at OpenAI due to potential conflicts of interest with TESLA.
June 2018OpenAI releases their first Generative Pre-Training model which improves language understanding.
Feb 2019GPT-2 paper published mainly taking advantage of much larger datasets. The trained model is not released as there are concerns about malicious applications.
Feb 2020Microsoft introduced its Turing Natural Language Generation (T-NLG), which was then the largest language model ever published at 17 billion parameters.
May 2020GPT-3 first presented in May 2020 and released for beta testing with over 100 times more parameters than GPT-2 (and 10 times more than T-NLG).
Sept 2020Microsoft gets exclusive licence to GPT-3 – allowing others to use the public API but giving Microsoft control of the source.