In the past month, you have almost certainly read many headlines describing the latest artificial intelligence, ChatGPT – OpenAI’s chatbot assistant based on their large language model GPT3 – or perhaps Google’s large language model BERT, which has helped to power Google Search since 2020.
[https://searchengineland.com/google-bert-used-on-almost-every-english-query-342193].
The latest news suggests that to remain competitive in the search engine market, Microsoft is now in talks to acquire a $10 billion stake in ChatGPT, in order to augment the capabilities of Bing search with advanced AI. On top of that, our favourite Office apps like Word, Powerpoint, and Outlook are also potentially getting the GPT treatment. Soon, we’ll be able to create all type of content: documents, slides and emails with a ChatGPT-like experience. How amazing is that?
https://www.theverge.com/2023/1/9/23546144/microsoft-openai-word-powerpoint-outlook-gpt-integration-rumor [https://www.cnbc.com/2023/01/10/microsoft-to-invest-10-billion-in-chatgpt-
Even with all the buzz around ChatGPT, we are likely to be blown away by the upcoming release of GPT4 – OpenAI’s successor model rumoured to be coming out early this year. It’s expected to be an order of magnitude bigger than its predecessor GPT3 and possess capabilities far beyond ChatGPT
You may rightfully be wondering, just what are large language models and what explains the current hype surrounding them and how can my company start benefitting from such technologies? Well, you have come to the right place!
What are Large Language Models?
Put simply, large language models are beefed-up text predictors – think a much more powerful (maybe even useful!) version of autocorrect on your smartphone. Today’s large language models are trained on massive quantities of text data collected from the Internet. The models are trained on a variety of text prediction tasks (e.g. fill in the missing word, predict the next sentence, etc) using the latest machine learning techniques. (For the machine learning enthusiasts among you, this is an example of self-supervised learning, which allows the training to be carried out on an enormous scale using completely unstructured data!).
Training large language models from scratch requires vast data storage and computational capabilities, which are only accessible to large tech companies – think Google, Meta, and Microsoft. Fortunately for those of us outside Silicon Valley, using and fine-tuning such models for other tasks is orders of magnitude faster, and it can be done by companies like us, Effixis! As Sam Altman (CEO of OpenAI) said during an interview with Greylock “There will be a whole new set of startups that take an existing very large model … and tune it … to create the model for [insert your sector here] . … Those companies will create a lot of enduring value.” Our goal at Effixis is to adapt these state-of-the-art models for your specific use case, whatever it may be! If you have not yet witnessed it, the capabilities of today’s large language models keep growing by the day and show no signs of slowing down.
What are their capabilities?
Though trained on text prediction tasks, the capabilities of large language models go well beyond simple fill in the blank exercises. When exposed to enough data during training, these models learn a rich representation of the underlying languages – which includes grammar rules, syntax, semantics, writing styles, and more. This can lead to many surprising capabilities in the finished model. (For instance, OpenAI’s GPT model, when trained over web data, ended up learning how to write code in several programming languages thanks to the abundance of Stackoverflow posts in its training set.)
These capabilities include, but are certainly not limited to:
- Text summarization, extracting only the important information of conversation threads, articles, or dense scientific papers
- Question-answering and generation of human-like responses in chatbots for customer service
- Content generation (ad copy, essays, templates, poems, jokes, etc…)
- Rephrasing written text in a different tone or writing style
- Programming purely with natural language
- Debugging code
How can your business investigate the ROI of such technologies with Effixis?
People have already started building apps on top of technologies like ChatGPT. At Effixis, we believe that every company should be trained on the topic and should have a go.
Founded in 2017 at EPFL, Effixis has a team of 15 highly skilled data scientists specializing in data analytics & reporting, advanced machine learning, and natural language processing (NLP – the technology behind ChatGPT). We specialize in the application of NLP (extraction, generation, and understanding of documents) to text management problems such as:
- News and emerging tech monitoring in the manufacturing industry
- Feature extraction from electronic health records (adverse drug reactions, allergies, vaccines, etc)
- Research paper summarization in the healthcare industry
- And so much more!
Luckily, we have been working mainly with the Azure Suite from Microsoft on our AI projects, acquiring extensive expertise in this environment. Given the latest news, we can expect a streamlined integration of OpenAI features in the Azure suite in the coming months, with the Azure Open AI Services.
https://azure.microsoft.com/en-us/products/cognitive-services/openai-service
Let’s get practical …
To help you get over the hype, understand what is at stake, and leverage the language models features to get business results, our AI team proposes a demystification workshops of ChatGPT / DALL-E / GPT3, during which we address the following among other:
· Understanding ChatGPT and GPT3:
- Defining and explaining the models and their underlying technology and architecture
- Gaining insight into the training methodology and training sets
- Reviewing ChatGPT and GPT3 capabilities
· The Power of Language Models
- Differentiating between Generative AI and Analytical AI
- Understanding how these models can be used for tasks such as Natural Language Processing
- Highlighting direct and potential benefits
· Practical Implications & Considerations:
- Identifying what tasks ChatGPT excels at
- Addressing potential concerns, limitations, biases, and data privacy
- Outlining best practices for implementation with successful use cases
It can be followed by an AI-strategy workshop to explore, identify, and prioritize together the various opportunities for implementing a solution with ChatGPT-like features in your company. Finally, PoC’s can be developed in order to showcase how your teams can be empowered with these technologies.
If you’d like, let’s hop on a quick phone call together to see how we can help you uncover the secrets of large language models for you and your team. You can book some time to chat with the team using the “Contact Us” button.
Looking forward to speaking with you soon!
Authors:
Dr Zachary Schillaci & Guillaume Vanhalst
