In this episode, I review last weeks conversation with ChatGPT, an artificial intelligence language model developed by OpenAI. I discuss the technology behind ChatGPT, its potential risks and benefits, and the ethical and societal questions it poses. I also talks about my personal experience with ChatGPT and my intention to use it for podcast scripting.
Or listen at:
Published: Wed, 21 Jun 2023 15:00:04 GMT
Links
- OpenAI - About
- OpenAI - Model Version
- Mashable: Whoops, Samsung workers accidentally leaked trade secrets via ChatGPT
- BBC: Sam Altman: CEO of OpenAI calls for US to regulate artificial intelligence
Transcript
Hello and welcome back to the Better AI from Software Development Podcast.
In the last episode I started Look at ChatGPT. It will have been difficult to escape the news of ChatGPT over the last few months, so I wanted to take my first look at it. In the episode I asked ChatGPT a number of questions directly and used the transcript for the majority of the episode.
In this episode, I want to review how I felt the interaction went, why I asked some of the questions, explore some of the answers received, and look at the future of ChatGPT and related services.
But let's start by unpacking a little what ChatGPT is. ChatGPT is a service that uses an artificial intelligence language model to interact using natural language. Natural language in this example was asking simple questions expressed in everyday language. ChatGPT then use its language model to understand the question and to formulate a meaningful response again in natural language. Underpinning all of this is the language model.
As the ChatGPT said, the language model is developed using machine learning and a considerable amount of training data such as news articles, books and online content. The machine learning is similar to how you will have learnt as a child. It is exposed to language and it learns in a similar manner, building up an understanding of relationships and concepts within the language. The output of the machine learning is a model, in this case, one that is trained against language text.
That model is then used to provide an API, an application programming interface, a service which accepts text requests, processes against the model, and then provides the response. The language model and the API have been developed by OpenAI.
OpenAI describes themselves as being:
"OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity"
The API that they provide can then be used by software developers to add into their own products. And in the case of last week's conversation, I used a website called https://chatgpt.org, which provides a nice interface over the API to make it easier for me to have that conversation and then to export into a transcript which I included in last week's show notes.
The https://chatgpt.org website is independent from OpenAI - they have just built their website to take advantage of the API that OpenAI provide. It is an excellent example of showing how the technology can be used.
Okay, let's move on to the conversation itself.
I have to admit to being pleasantly surprised with how well ChatGPT was able to understand my request and provide a meaningful response. It felt very natural and arguably a better experience than I've had with many human powered online webchat services.
During that conversation, I asked a couple of specific questions that I want to talk about here.
Firstly, I wanted to talk about why I asked ChatGPT which version it was. I asked that question because OpenAI are continually developing new models and evolving them. Based on the documentation, I believe last week's chat via the https://chatgpt.org website was using version 3.5, whereas OpenAI are now in beta test with version 4, which is described as being "A set of models that improve on GPT-3.5 and can understand as well as generate natural language or code".
So much of the conversation and news that you will have heard about ChatGPT will have been based on version 3.5. It is interesting to see that the next generation is already being previewed. This should be seen as an indication that the technology is moving quickly.
The second question I wanted to go over was why I asked it if it was provided proprietary information, was there a danger it would be disclosed to others. And I asked this question for a reason. Interestingly, the ChatGPT response indicated that it could not disclose proprietary information. However, the Korean version of The Economist reported three separate instances of Samsung employees unintentionally leaking sensitive information to ChatGPT.
It was reported that employees use the service to help them with their tasks, one of which was preparing notes for a presentation only later to find that the proprietary information was then exposed by ChatGPT.
One article covering the subject commented that this was to be expected. From ChatGPT its own data policy, it says that it uses data to train its models. Unless you request to opt out in the ChatGPT usage guide, it explicitly warns users not to share sensitive information in conversations.. I'll include a link to the article in the show notes.
For me, this accidental exposure of proprietary information is an example of operation controls not keeping up with technology. Technology will continue to innovate and provide helpful and beneficial services. Busy employees will use those services if it helps them to get their job done better. But often this is done before organizations have taken the time to assess the risks of those services.
The same thing happened with the explosion of the use of cloud. Anyone with a credit card could spin up a service, put their valuable customer database in some random database, often referred to a shadow it, and run the risk of exposing that valuable data because the due diligence hadn't been done.
So while the risk of accidental exposure isn't new, I'd suggest it's possibly no less dangerous than using online services like Grammarly, Gmail, or Dropbox. If used correctly, it can be amazing things for an organization, but it should be something that organization leaders are aware of and making sure they understand the risks.
You want to get the benefit in a risk controlled manner.
So let's talk about where we see the future of these services.
In short, these services are here to stay and will continue to advance over time, but they do prompt real questions.
I'm recording this episode about a week after Sam Altman, the CEO of OpenAI, testified before a US State Senate committee to talk about the possibilities and pitfalls of the technology. Surprisingly, during this, Mr. Altman called for a regulator to be formed to license AI companies. Mr. Altman said:
"I think if this technology goes wrong, it can go quite wrong...we want to be vocal about that," "We want to work with the government to prevent that from happening."
So will there come a time when we need to bow to our AI overlords as predicted by countless Sci-Fi books and movies for the past 70 years?
Personally, I don't believe we need to worry about a Terminator style future where a Skynet deems all human life unnecessary. Artificial intelligence is not sentient. While it can appear to breed at level sentience, it is purely extrapolating from data that is built upon.
There are, however, ethical and societal questions that the use of AI poses. Is it ethical to use AI to make certain decisions such as those around life and death decisions? And as a society, we need to adapt to the changes in job roles and our ability to earn a wage.
Now, arguably, this is nothing new. Like all technological revolutions that came before it, society will need to adapt to the different ways of working and what that means for our populations.
If we do reach a future where AI and robotic automation drastically reduces the need for a human workforce, then we may need to adjust our reliance on a capitalist society which is ultimately based on a demand for labor - we work for money - we use that money to achieve a standard of living. Without that demand for labor, we have an unutilized, and possibly starving, workforce.
We have seen this as a problem in isolated industry sectors as automation and technology takes hold. The question here becomes, is this AI change enough to impact a large enough part of industry that capitalism is no longer a viable model?
Going back to the US State committee hearing, Mr. Altman said:
"There will be an impact on jobs. We try to be very clear about that," he said, adding that the government will "need to figure out how we want to mitigate that".
Democrat Senator Richard Blumenthal commented that the AI dominated future "is not necessarily the future that we want".
"We need to maximize the good over the bad. Congress has a choice now. We had the same choice when we faced social media. We failed to seize that moment"
I'll include a link in the show notes to a BBC article covering the hearing.
I do, however, think that Congress and Mr. Altman may already be too late and possibly naive to think this is just a US only concern. The technology exists and it exists outside of US borders and controls. Simply regulating it within US borders feels a little like shutting the door after the horse has bolted. It's only a few weeks ago that the AI godfather, Geoffrey Hinton, quit Google over the dangers of AI. But again, maybe a warning message too late.
The ethical and societal impact are likely to be something of much debate over the coming decade.
More immediately, though, I do have a concern of a diminishing of original content and thought. Our natural tendency is to use technology to help us with our everyday lives - the Samsung employees are a great example of just trying to get their work done - why wouldn't you use a tool to help you get started?
I can well understand the temptation of students to use ChatGPT to provide them help in assignments, even if only to help them get past that initial blank page.
But by using a large language model like ChatGPT, does that effectively limit free thought?
AI models are built on the same content that they could potentially eradicate. Are we in a danger of a future AI model being built on the content of prior AI models? Are we recycling the same stale source material over and over without any advancement? We maybe get stuck with the AI approved way rather than having disparity of thought.
And given that the AI models are unlikely to provide attribution to the original source, what benefit is there for someone to take the time to even create that original thought and then to write it down?
This may seem like an academic concern, but with learning communities such as software development, I can see this being a real and impending risk. While I'll talk about it more in future episodes, advancements in software development relies heavily on disparity of thought. New ideas to address old problems and then being able to receive kudos for that contribution to the overall practice.
So given my initial experience and the possible futures of it, will I be using it?
Well, yes.
I can definitely see myself using ChatGPT or similar tool for helping me with podcast scripting, for example. I think it would be difficult for an audience to tell the difference between the ChatGPT script rather than my own - if anything, the ChatGPT is likely to be better written. As such when starting with a blank page, using a tool like ChatGPT can really help to get basic content in place, and at this stage I'd probably only intend to use it for basic content - with that being reworked into final episode script by hand. I suspect factual content being provided via ChatGPT and opinion being provided by myself, but time will tell.
To keep myself honest, I will include any ChatGPT transcripts used to produce an episode in the show notes.
And obviously in some future episode I'll revisit my experience with ChatGPT and related services and give you some sort of feedback on how I feel it works for me.
In this episode, I took a look back at that initial conversation I had with ChatGPT and my initial impressions. I've talked about how organizations are struggling to keep up with having the correct governance around these services as the disclosure of sanctions proprietary information demonstrates. I've talked about my own personal concerns over diminishing of original content and thought if everything is delivered by AI powered models. And I've asked the question what will the long term impacts be on our society? And finally, I've given my own commitment to transparency and the use of ChatGPT or similar in the production of this podcast.
ChatGPT and artificial intelligence is here to stay. Any technology once invented cannot be invented. It has the potential for great good and great harm. And it's something that I'll undoubtedly touch on regularly as we map its effects on software development, the organizations in which we work and the societies in which we live In.
Next week's episode, I want to take a look at a related subject - the growing use of large language models, such as ChatGPT, in helping software development - something like an AI coding assistant.
Thank you for taking the time to listen to this episode and look forward to speaking to you again next week.