Few pieces of technology – ever – have enjoyed a deeper love-fest from the media and consumer public than ChatGPT and other endeavors built on top of vanilla GPT-3. And enterprise IT managers have been busy developing homegrown apps based on GPT-3.
So far, so good.
But as we’ve seen before – think of the mid-1990s internet boom or blockchain more recently – companies can easily get ahead of the game by making big investments in things other than strategic goals.
I remember talking to an executive about trying to launch a website in the very early days of the World Wide Web and asking, “Why? What for? What are you trying to achieve?” Instead of getting a litany of concrete goals and objectives, I heard variations of “One of our board members read about it and insists we get one,” or “Our CEO’s son can’t stop talking about it.” talk” or the dumbest: “Everyone seems to be doing it.”
Those are exactly the comments I’m hearing about GPT-3 today. To be clear, many of the industry’s most hyped technologies ended up being strategically important. Not all, but many.
ChatGPT has an impressive set of capabilities, but it’s really just a huge database with an interface that effectively mimics human communication. Think of it as a hyper-powered intranet.
The information presented in most ChatGPT exchanges is nothing that couldn’t be found with a decent Google search. “Found” is the key point. A user may have to look through dozens of Google search results to find the one answer that ChatGPT finds.
Another advantage – and this is perhaps where the most IT value can be found – is the human interaction. In theory, this could eventually lead many coding projects to skip the more basic programming endeavors. Most programming projects start with a business leader or manager saying, “We need the system to do XYZ. Go make that happen” to technical talent.
What if ChatGPT could bypass some of that coding talent and create code directly from industry instructions? Some coding is very creative and resourceful and will continue to need a human touch. But frankly, a lot of programming is painstaking and repetitive. Can GPT take over that part?
On the other hand, we’ve all seen the ridiculous errors and outright fabrications that GPT-3 systems have produced. Until that is resolved, use of GPT-3 will be limited. As enticing as the natural language interface is, letting a GPT-3 chat program speak to customers on your behalf is asking for disaster.
So how can is it used? There are two ways to explore that critical question: prescriptively and openly. Depending on your business and objectives – not to mention budget – either approach can be very attractive.
The prescriptive approach is simpler and more likely to yield more short-term results. What are you trying to achieve? What can GPT-3 do today to help your business and perhaps enable some product/service implementations that you’ve wanted for a while but couldn’t achieve.
The open approach is more interesting. That’s where you give your team a lot of leeway to play with GPT-3 and get creative and see what it can do. But that approach must have limits.
CIOs need to figure out what they want to do with this, or developers will endlessly spin their wheels on crazy ideas, said Scott Castle, the chief strategy officer at analytics firm Sisense. “CIOs have to filter strategically, otherwise you’re just Willy Wonka in the chocolate factory,” he said.
One of the top analytics experts in the industry is Roy Ben-Alta. Ben-Alta left Meta/Facebook as director of AI last month to start his own company. Before Meta, he served with Amazon for 11 years, ending with the title of director, analytics/machine learning, data streaming, and NoSQL databases.
CIOs “have to ask themselves, ‘What’s this going to do to my business?'” said Ben-Alta. “The best way to approach that is to work backwards from the customer. What problem are we trying to solve? Here’s the catch: to jump in, you have to spend a lot of money. Training requires a lot of CPUs. Any use case requires specific data sources and if they don’t have the data available, they need to determine how much it will cost them to acquire that data.”
The most powerful element of GPT-3 is the encryption, the interface. But for companies trying to build on top of all that, the problem isn’t coding. It will definitely be data.
“The Achilles heel of any analytics system is data quality,” says Ben-Alta. “Most of the work consists of the data. Data integration is always the problem and the most challenging element. The format of the data and the type of data to be used evolves. The analytics model only becomes good when the data becomes good.”
The data issue is important, but much of the analytical complexity materializes because of data interactions. Waqaas Al-Siddiq, the CEO of medical analytics company Biotricity, offers a powerful example of how interactions can undermine the best of large language models.
“Anything that’s a peak or an anomaly — like in three or four standard deviations from the mean — is going to be a lot of trouble. The more variables, the more challenging it gets, because you need more data,” Al-Siddiq said. they’re anomalies, you can’t provide enough data.”
Al-Siddiq gave an example of inventory logistics: “Let’s say there is a construction project that causes people to divert and during those same two weeks there is a heat wave. That made people stop and grab a drink. Now there are multiple variables. You will never have enough data to handle that anomaly in an autonomous or predictable way unless you make sure you keep track of those variables. The more variables you track, the more complex your AI model.”
There is enormous potential value in exploiting these large language models, but it is clearly a good idea not to let emotions get the better of you.
“This whole buzz is because of one product from one company: OpenAI. This society runs a lot on the bandwagon effect, the fear of missing out,” said Jay Chakraborty, partner at PwC (formerly Price-Waterhouse-Coopers). “This is another version of the California gold rush, the dotcom euphoria, that whole Y2K ‘the whole world is going to crash’ thing.”
Chakraborty encourages CIOs to just do some sandbox experimentation and “prompt the company to come up with ideas and use cases. If I’m a hedge fund, why should I? not thinking about what I can automate? It could easily knock out investment letters much more efficiently. It generates the analysis automatically and that is another crucial step towards the finish line. It is writing the final piece.”
Forrester analyst Rowan Curran, who specializes in data science, machine learning, artificial intelligence and computer vision, agrees that GPT-3 has great potential, but said executives should see it as just another strategic effort.
“The the first thing to do is step back from the public eye and ask, ‘Where can we actually apply these where we can take advantage of their strengths and play down their weaknesses? How can we use it?’” Curran said.
While GPT-3 “is potentially a great way to innovate, it’s also really important to focus on what’s practical in the short term. There absolutely is a need to educate yourself about what is even possible. This is a new and dynamic space,” Curran said, adding that he sees serious boundaries. For example, he thinks the idea of using it for chatting in a customer-facing application is “extremely irresponsible”.
Large language models are nothing new, but the human-mimicking frontend that GPT-3 created has made the IT world aware of the possibilities and made many dream. That’s great, as long as they wake up just before investment decisions are made and project directions set.
Copyright © 2023 IDG Communications, Inc.
