Tag Archive for: GPT-3

A recent trip to CCW 2023, Berlin

By Björn Gülsdorff

CCW is the “international conference and trade show for innovative customer dialogue”, which is a bit bulky a title, but much better than the original ”CallCenterWorld” which is no longer a good fit. Unfortunately, this is now easily confused with ContactCenterWorld, also called CCW. But I managed to get to the right event to support our German partner SOGEDES in matters related to Conversational AI.

One very non-technical but nonetheless important takeaway from CCW is that Real life ain’t dead. CCW sported a hybrid concept, with all talks and product presentations available through live streams (and recordings after the event).  And still, many people came to Berlin, and I think that nearly everyone who had come was ready for business. The quality of the conversations at the SOGEDES booth was outstanding.

Unsurprisingly, ChatGPT was a big topic. However, although it was hot in talks and booth side chat, it was evident that few real applications had been built yet, except for the obvious use cases of summarising texts. This was not only due to the short timeline between the launch of CHATGPT and CCW, but what I heard many people saying was that ChatGPT is not ready for customer dialogue. This is correct of course (as GPT4 was not out then) but I was still very surprised how educated and relaxed people were about it.

I had expected a buzz similar to the machine learning hype some years back. At the same time, ‘AI’ has become a commodity. Everyone has it – because everything is now called AI. Similarly, but to a lesser extent, Conversational AI has lost leading edge appeal and simply become something that’s used for automated conversations i.e., chat bots – but those, as a word, have fallen from grace.

The top two top – Agent Assist and Voice Bots

After years (decades?) of promoting Digital Self Service, voice is still a strong channel and companies are now looking to automatise calls and are no longer avoiding this. The handover/routing of calls from bots to agents is an important ingredient in the requirements.

In a way, organisations are now looking for something that Creative Virtual has been suggesting and recommending since its founding. That is, virtual agents are a part of an overall conversational strategy and there should be collaboration between human and silicon agents.

We have been delivering solutions that enable human/digital agent collaboration for decades.  In fact, “agent assist” is the name of some of the internal projects we deliver to our customers that, well, assist agents, so they can provide a better service. Our recent work with smart and audio-codes is also testament to our credentials as leaders in conversational AI, and the best choice for customers.

It is worth saying, however, that this new focus on voice brings with it very simplified conversations, or at least very rigidly structured conversations, something that is no longer prevalent in text.  You can’t get very complex in a phone conversation, you can’t show images to let users choose the right product, and you can’t play a video to explain something.  To deliver the experiences customers expect, organisations do need to ensure there is integration, with seamless interfaces for these simple flows.

The Overlooked Feature of GPT : Vectorisation

By Olaf Voß, Lead Application Designer

These days everyone is stunned by the generative power of the GPT models, including myself. However, today I want to discuss a GPT feature that is largely overlooked in media coverage: the embeddings endpoint in the OpenAI API. This feature ‘translates’ any text into a 1536-dimensional numerical vector. Personally, I prefer to use the term ‘vectorisation’ instead of ’embedding’.

The idea to turn individual words into vectors is about 10 years old now. These word vectorisations were trained on large corpora – or what we thought were large corpora 10 years ago. They are useful because the way words are distributed in the vector space represents relationships between them. The most famous expression of this concept is probably the equation king – man = queen – woman, which holds approximately true when creating word vectors with tools like GloVe or Word2Vec. These word vectorisations have been in use ever since, forming the basis for much of the progress in machine learning for language-related tasks.

Now, GPT offers the same level of analysis for entire texts. It’s not the first model that can do this beyond single words, but its high quality and affordability make it highly attractive. If you experiment with it, you can quickly see how useful it may be. For example, a question and its answer are matched to fairly similar vectors. Additionally, a text in one language and its translation into another will be mapped to close-by vectors.

This vectorisation can be used for anything that can be done with vectors. Text comparison and thereby text search is one obvious use case. We at Creative Virtual will be using it this way in our upcoming Gluon release for intent matching – still keeping rules as the fallback option if and when needed. Another way we are already using it is for text clustering. Finally, you could use the text vectors as the input layer for a neural network and train it for whatever task you want, thereby ‘inheriting’ many of GPT’s text understanding capabilities.

So, if you have access to the OpenAI API and if you are running out of ideas for what to do with the chat endpoint, give the embeddings endpoint a chance. Vectorise away!

​Generation AI: Growing up side-by-side with our silicon-based contemporaries

By Olaf Voß, Lead Application Designer

I was born in 1966. That means I’m usually sorted into Generation X. But these days, looking back at the past 57 years, I think we should really rename it to Generation AI. It has been my generation having witnessed AI from its infancy to the breakthroughs we’ve seen in the past few years. And with a bit of luck most of us will witness how AI will be reshaping our societies – for good or bad – in the next 20 years.

So let me give a recount of my encounters with AI throughout the decades.

There’s no way around it: I have to start with ELIZA, which Joseph Weizenbaum developed in the year I was born. I was too young to be aware of this when it was new of course, and even less aware that chatbots one day would play a major part in my professional life, but merely 16 years later I had access to Commodore computers at our computer club in secondary school. The ‘large’ one had 16 kb RAM AND a floppy drive! And we had an ELIZA clone running on them. But I admit I didn’t spend much time with her, I was far too busy with freeing princess Leia in an early text adventure or writing my own very simple games.

​I had my first chess computer at some time around 1980. It could analyze to a depth ​of 5-6 moves. I was an OK, but not great, player and I won against it when giving it up to 30 seconds thinking time and lost at above 2 minutes. In the early 90s I became a club player and soon after I didn’t have the slightest chance against the chess programs of that era even with 10 seconds thinking time. No need to be embarrassed about that I guess, since Gary Kasparov lost against Deep Blue in 1997.

Around 2005 I started playing Go. I was a convert from chess, and by that time I was used to having great chess programs available for training and analysis. Go has a much greater branching factor than chess and is much less suitable for static board state evaluation. With the available technologies of those days, programs could play Go at a mediocre level at best. Well, I still lost to them, but they were not strong enough to rely on their judgement. At that time most Go players, including myself, (apart from thinking that their game is much better than chess) thought it would take at least 50 years until computers could crack Go. It took about 10.

​How did they succeed​ so quickly? Deep neural nets. I read about those first as a university student in the 80s and was thrilled. I played with them a bit on the first computer I owned, an Atari 520st. I quickly thought about applying them to chess. My ideas were not very far from what is done in that field today, but of course I hadn’t heard about reinforcement learning at that time. I very much like to believe my ideas were extremely clever and would have worked. Alas, we’ll never find out, because it was clear very quickly that the hardware (especially mine!)  at that time was totally inadequate for tackling this problem.

​With what I’ve given away about myself so far nobody will be surprised to hear that I ended up becoming a software developer. Around the turn of the millennium I started to work on chatbots. We must have been one of the first few companies worldwide to tackle this commercially. ​At that time I was reluctant to say that I was working in AI. We were using pattern matching, and even a pretty simple form of that. Of course pattern matching IS an AI technique, but I was aware that with our focus on building chatbots that were useful for our customers in their restricted scope and on doing that efficiently, what we did would have bored any AI researcher. I wasn’t ashamed of what we were doing – quite the contrary: I was proud about what we could achieve with our pragmatism. I just wanted to avoid the pointless discussion if what we did was ‘real’ or ‘interesting’ AI.

​Fast forward another 15 years or so.​ Word embeddings came up and made it possible to tackle natural language problems with artificial neural nets. So I welcomed those back into my life. Only now the hardware was somewhat better plus I got paid for playing with them. Heavens!

Then comes 2020 and GPT-3. That was mind-blowing. I’ve heard people characterising deep neural nets as ‘glorified parameter fitting’. And sure, parameter fitting is all there is to it. But these parameters, each by itself just a dumb number, let something pretty astonishing emerge. I am not an expert on these topics and I am not even sure how much sense it makes to compare human and artificial intelligence. But sometimes I feel provocative and want to ask how we can be sure that our own intelligence is more than just ‘glorified synaptic strength fitting’. Once again I think the discussion about ‘real’ intelligence is far less important than considering what can be done. And, since there’s so much more that can be done today, how that will change our world.

Apart from being fascinating and super-relevant to the field I’m working in, GPT-3 is also a lot of fun of course. I remember a conversation I had with it in the early days, before OpenAI put the brakes on it to avoid harmful responses. (Which I applaud, in spite of it spoiling some fun.) I asked it  – actually before the start of the war in Ukraine – for a couple of suggestions about ‘how to achieve world peace’. One of the suggestions was: ‘Kill all humans.’ Well yes, job done … I’m still glad you are not yet in charge, you know!

I want to mention two more recent developments, even though they relate to me personally in a tangential way at best. Being a physicist by education I follow scientific developments closely, which brings me to AlphaFold 2. When I was 4 years old, the first successful DNA sequencing attempts were made. In 2020 AlphaFold 2 predicted and published 3D structures of thousands and thousands of proteins based on DNA sequences alone. Another loop closed during my lifetime. I make the prediction that in 2035 more than 50% of all newly approved drugs will have been developed using the results of AlphaFold or its successors at some stage in the process.

The second one is CICERO. As an avid board game player I reluctantly admit that I have never played Diplomacy, though I did play similar games like Risk or Civilisation. Diplomacy is a conflict simulation game in a WW1 scenario. It involves tactical moves on the board and lots of diplomacy around it – pacts – betrayals – revenge. CICERO can play this game on par with human experts. Apart from making clever moves on the board – easy-peasy these days for AI – it has also to negotiate with other players in natural language. So it needs to bring together strategic, natural language and social skills. Even though this is a model with a  niche application scope, I think it is at least as impressive as GPT-3, if not more so.

​​We are living in exciting times. And I think it’s important to understand that we are seeing the beginning of something, not the end. What will be possible in 20 years? Many things will happen, not all of them good. I’m not much of an expert in AI risks and besides, discussing them here in detail would go far beyond the scope of this blog post. Still I’m asking myself, how we as a society will cope with these – at the moment still largely unknown – changes.

My role at Creative Virtual involves looking at all the new technologies that pop up and evaluate if and how we can use them. So I have a bit of a front row seat in watching this unfold. I encourage you to check out our ChatGPT, GPT-3, and Your Conversational AI Solution blog post for a closer look at how we see these recent developments fitting with the work Creative Virtual does in the customer service and employee support space.

I think it is of the utmost importance that as many people as possible have a basic understanding about what’s going on. An ignorant society will not be able to react. I am trying to play my small part by sharing my knowledge with as many people as possible. Just recently, when an old friend of mine called, my wife burst out: ‘Great, now you can talk his ears off!’

ChatGPT, GPT-3, and Your Conversational AI Solution

By Chris Ezekiel, Founder & CEO

Since the official announcement in November 2022, there has been an enormous amount of buzz and excitement about OpenAI’s ChatGPT. Industry experts are publishing articles about it, social networks are filled with comments about it, and local, national, and global news organisations are reporting stories about it. From students using ChatGPT to complete assignments for class to me getting a little help from ChatGPT to write my latest ‘Virtual Viewpoint’ column, it certainly seems like everyone is testing it out.

As a specialist within the conversational AI space, Creative Virtual is excited about what ChatGPT and the technology behind it bring to our industry. We’ve been having lots of discussions with our customers and partners, as well as internally, about how this can deliver value to businesses using our V-Person™ solutions.

ChatGPT is an extremely powerful language model that is changing quickly and will continue to get more sophisticated. However, like any deep neural network, it is a black box which is hard – if not impossible – to control. Using it as a generative tool means you can’t steer in detail what it’s going to say.  You can’t deliver reliable, accurate self-service tools if you can never be certain what response might be given.

These limitations don’t mean you should write off ChatGPT or GPT-3 (and future versions) as completely ineffective in the realm of customer service and employee support. In some cases, one might be willing to accept a certain risk in exchange for very efficiently making large chunks of information available to a chatbot. Also there are ways to use the language power of GPT in a non-generative way, as we’ll explore in this post.

In any case, ChatGPT can only ever be used as just one piece of the puzzle, like content management, integration, user interface, and quality assurance. ChatGPT alone cannot replace all of that.

One of the design features of Creative Virtual’s conversational AI platform is the flexibility to integrate with other systems and technologies, including multiple AI engines such as transformer models like GPT-3. We are currently exploring the best way to interface with this model and use it to deliver value to our customers and partners.

Let’s take a closer look at ChatGPT, how it works, and the ways it can be used to deliver customer service and employee support.

 

What kind of AI is ChatGPT and how is that different from how V-Person works?

ChatGPT is a transformer model, a neural network, and is trained to predict text continuation. It uses a variation of GPT-3 which is OpenAI’s large language model (LLM) trained on a wide range of selected texts and codes. It is extremely powerful with respect to language understanding and common world knowledge. However its knowledge is not limitless and so on its own it will not have large parts of the information needed for specific chatbot use cases. Also its world knowledge is frozen at the time it was trained – currently it doesn’t know anything about events after 2021.

V-Person uses a hybrid approach to AI using machine learning, deep neural networks, and a rule-based approach to natural language processing (NLP). The machine learning component is integrated with workflow functionality within our V-Portal™ platform so enterprises can decide the best configuration for their conversational AI tool to improve in a controlled and reliable way. At the same time, natural language rules can be used as an ‘override’ to the machine learning part to ensure accuracy, resolve content clashes, and deliver very precise responses when needed.

We developed this approach to give our customers control over the AI to create accurate, reliable chatbot and virtual agent deployments. The use of natural language rules as a fallback option to fix occasional issues and finetune responses is much more efficient than trying to tweak training data.

 

Can businesses use ChatGPT to directly answer questions from customers and employees?

At the time of writing, ChatGPT is still in a research preview stage and highly unstable with no clean API available, so it’s not possible yet for businesses to use it in this way. However with its predecessor, InstructGPT, it is. It’s also worth noting that GPT-3 is high quality only in English and a few other languages which is another potential limitation for global use.

The biggest issue with using ChatGPT to directly answer questions from customers and employees is that it does not give you control over how it will respond. It could give factually incorrect answers, give answers that don’t align with your business, or respond to topics you’d prefer to avoid within your chatbot. This could easily create legal, ethical, or branding problems for your company.

 

What about simply using ChatGPT for intent matching?

There are two ways in which GPT-3 could be used for intent matching.

The first way just uses GPT-3 embeddings and trains a fairly simple neural network for the classification task on top of that. The second option also uses GPT-3 embeddings and a simple nearest neighbour search on top of that. We are currently exploring this last option and expect to get some quality gains from that approach.

 

Can I just provide a few documents and let ChatGPT answer questions by ‘looking’ at those?

Yes, this is absolutely possible. In fact, we have offered this functionality with V-Person for several years without needing GPT but none of our clients have been interested. GPT-3 improves the quality of this in most cases, but also comes with a higher risk of being very wrong. If an organisation is interested in using GPT-3 in this way, we can support it within our platform but what we currently offer already enables us to deliver document-based question answering.

It’s important to keep in mind that using ChatGPT to answer questions from documents is only addressing one aspect of the support expected from a virtual agent. For example, no transaction triggering API will ever be called by GPT looking at a document.

 

Is it possible to give GPT-3 a few chat transcripts as examples and let it work from them?

You can provide GPT-3 with sample transcripts and tell it to mimic that chat behaviour. But unless you want a chatbot with a very narrow scope, a few transcripts won’t be enough. If there are complex dialogue flows that need to be followed, you’ll need to provide at the very least one example of each possible path – most likely you’ll need more.

This raises some difficult questions. How do you maintain those if something changes? If you try to use only real agent transcripts, how do you ensure that you have complete coverage? How do you deal with personalised conversations and performing transactions that require backend integration? It may not be too difficult to train the model to say ‘I have cancelled that order for you’ at the right time, but that doesn’t mean GPT will have actually triggered the necessary action to cancel the order.

When you really examine this approach it becomes clear that this is not an efficient way to build and maintain an enterprise-level chatbot or virtual agent. It also doesn’t address the need to have integration with backend systems to perform specific tasks. Today our customers achieve the best ROI through these integrations and personalisation.

 

What other key limitations exist with using ChatGPT to deliver customer service or employee support?

Using a generative ChatGPT-only approach to your chatbot does not give you the opportunity to create a seamless, omnichannel experience. To do that, you need to be able to integrate with other systems and technologies, such as knowledge management platforms, ticketing systems, live chat solutions, contact centre platforms, voice systems, real-time information feeds, multiple intent engines, CRMS, and messaging platforms. These integrations are what enable a connected and personalised conversational AI implementation.

With ChatGPT there is no good way to create reliable and customised conversation flows. These flows are regularly used within sophisticated conversational AI tools to guide users step-by-step through very specific processes, such as setting up a bank account. This goes a step further than just creating a conversational engagement to employing slot-filling functionality, entity extraction, and secure integrations.

You also won’t have the ability to optimise the chatbot for the channels and devices on which it will be used. This includes using rich media – such as diagrams, images, videos, hyperlinks – within answers. For example, you can’t include an image carousel to display within a messenger platform. You won’t be able to show photos or drawings to help with a new product set-up. You don’t have the ability to display clickable buttons with options for the user.

 

As ChatGPT continues to change and moves out of the research preview stage, our expert team at Creative Virtual will stay on top of new developments and opportunities this technology offers. Our mission is always to innovate in a way that will help companies tackle their real challenges and deliver real business results – and our approach to this language model is no different.

If you’re interested in discussing more about how ChatGPT and V-Person might fit with your conversational AI strategy, get in touch with our expert team here.