Get the latest on how the workplace is changing, delivered to your inbox daily. Sign up for the WorkLife Daily Newsletter here.
Artificial intelligence developments are moving so fast that sometimes it feels difficult to keep up. Last week, the latest version of GPT, the language model behind ChatGPT, was released, with significant advancements compared to GPT-3 and 3.5.
GPT-4 is the company’s newest language model that can receive both text and image inputs, compared to GPT-3 and 3.5 which were just text-based. OpenAI spent the last six months developing GPT-4 using lessons from its adversarial testing program, a technique used in machine learning to fool or misguide a model with malicious input, as well as ChatGPT — the tool which catapulted this rise of AI into the mainstream. The company responded to two areas of ongoing concern and said it’s 40% more likely than its predecessor to produce factual responses and 82% less likely to respond with “disallowed content,” like providing answers for how to create a bomb or how to steal from someone.
“The best way to describe it is the difference between a college student and a high school student,” said Mike Murchison, co-founder and CEO of AI-powered customer service automation platform Ada.
Companies are quickly building new tools with GPT-4, like accounting giant PwC, which just signed a 12-month deal with startup Harvey that will give its lawyers access to an AI tool built on the newest technology. The multimodal model passed a simulated bar exam with a score around the top 10% of human test takers. In contrast, GPT-3.5’s score was around the bottom 10%. Other early adopters of GPT-4 include Duolingo, Microsoft, Stripe, Khan Academy, and Snapchat.
“For the individual user, it’s more likely that they will be accessing it [GPT-4] through technology they were already using,” said Cliff Jurkiewicz, vp, strategy at global HR tech company Phenom. For instance, instead of going to ChatGPT Plus, the GPT-4 language model will come to more and more products we already use in our everyday lives and will better assist us directly, removing the need to open up a new tab.
There is currently a GPT-4 API waitlist to help build more applications and services, where those interested can integrate it into an existing product or build from scratch. “ChatGPT was like the iPhone moment that the mobile world had,” said Murchison. “It took a while for everyone to build a great iPhone app. That’s sort of the moment we’re in right now with large language models. It will take a little while for us to get the killer apps, but they’re coming and they’ll be everywhere.”
Early versions of GPT, and especially the introduction of ChatGPT, helped show workers just how much more they can get done if they have their own AI assistant at hand to ask questions. While everything that workers were doing in the former versions still applies, things are taken up a notch with GPT-4.
ChatGPT – the tool which helped make AI so easily digestible and accessible to the general public – runs on GPT-3.5. To be able to access the functionality of what GPT-4 can offer, people and businesses need to pay for the upgraded version ChatGPT Plus, which costs $20 a month.
“I think that any worker at any job that doesn’t have a ChatGPT window open is doing it wrong,” said Tal Lev-Ami, co-founder and CTO at media experience cloud company Cloudinary, which is on the API waitlist. “If you haven’t found a use case yet, you need to think a bit more.”
We spoke to a number of experts, who have been playing with GPT-4 since its release on Mar. 14, to learn more about how workers will begin to use the new model for their everyday work, aside from how it has been implemented into other apps.
“It’s another new kid on the block,” said Nigel Cannings, CTO at Intelligent Voice, a company that provides speech to text services.
1. Ask GPT-4 to read a 50-page contract and tell me what the top areas of concern are.
OK, maybe we aren’t all dealing with large contracts, but the point is GPT-4 can consume up to 50 pages of text in ChatGPT Plus. That’s double what the prior models could do. Instead of spending your time reading a large document, whether it’s 20 or 45 pages-long, a contract or a white paper, you can have GPT-4 do the work for you.
“It’s going to take a while for users to catch onto this, but soon people will start to realize that actually, they are the limiter to the model’s performance,” said Murchison. “The model isn’t the limiter. That’s where things start to get really exciting.”
It shows how important it is for workers to understand not only the full capacity of GPT-4, but also how to best prompt it to get the right outcomes for the work. For example, after it gives you the output of what the top areas of concern are in the contract, you can very easily ask it to then write an email to the person who shared it, expressing the concerns.
Overall though, the larger context window is extremely useful and creates endless opportunities for inputs. GPT-4 can remember and act on more than 20,000 words at once. Other ways in which workers might take advantage of it is by asking for summaries of long transcriptions, turning those into a 10-slide presentation, and so on.
2. Upload images for social posts and auto-generate captions.
One of the best parts of GPT-4 is that it can take in both text and image outputs. However, it is only available in the API. That means that you won’t be able to access this through ChatGPT Plus. It will only be an option in other products. During OpenAI’s demonstration, the founder took a picture of a napkin sketch of an app and had GPT-4 turn it into code for a website. It’s pretty mind-blowing.
However, in our day-to-day jobs, we’re not always going to be building websites. That’s why Rachel Woods, founder of media and education company AI Exchange, suggests uploading images. Instead of describing what the image is and then asking for a caption, it knows exactly what it’s looking at with the quick upload. Taking a photo cuts out even more time than it would take to describe an image using text and helps provide a first draft for some of those social captions.
3. Consider treating GPT-4 as another employee, who is more reliable and “hallucinates” less than GPT-3.
There is so much more nuance possible with GPT-4. The prior models would provide outputs that seemed like they could be a real answer, but in fact, were not. For example, there was a trend where people would ask ChatGPT “Who is [insert your name]?” and it would often make up different jobs that person held, say they had accolades that they didn’t, and were wrong about where they studied. It’s a small example, but ChatGPT would lie consistently in a way that appeared authentic.
With GPT-4 being 40% more likely than its predecessor to produce factual responses, it’s easier for us to trust it. “It’s lying to me a lot less, which is comforting,” said Lev-Ami. “Its ability to structure information and process it has been improved.”
One of the problems with generative AI in the past is that it likes to fill in the gaps even if it doesn’t know the answer. “It would literally hallucinate,” said Cannings. “It comes up with answers which are just factually wrong. GPT-4 definitely is hallucinating a lot less. It seems to be trained on a wider variety of data and a lot more reinforcement learning has been used. It also seems to be a bit less toxic as well.”
Murchison says that’s why he can see folks using GPT-4 as if it were “another employee” contributing to the conversation.
“This morning, we were asking what we should name one of our new products,” said Murchison. “We’re trying to figure out how to best articulate it to make sure it lands. There was a whole discussion, and one of the colleagues copied and pasted the conversation into ChatGPT. It said ‘well, look, here’s actually a better way to think about it.’”
Benjamin Netter, founder of Riot, a cybersecurity training platform for employers and employees, says that GPT-4 is extremely accurate at calculating impact. “It doesn’t sound as crazy as before that you would use it as real advice,” said Netter.
However, Jurkiewicz says we shouldn’t let our guards down entirely. It is necessary for users to have a sense of skepticism around the output. “It’s still not 100% reliable,” said Jurkiewicz. “People should know that its responses in certain areas should be tested with other sources. There is still a liability there. It hasn’t learned everything yet, and it will be a long time before it does. You don’t want the liability of the technology to become the liability of the enterprise or the individual’s work output.”
4. Train GPT-4 to learn your tone of voice.
Nearly every expert raised the importance of training GPT-4 to learn your tone of voice and even suggested journalists consider doing so, to ensure articles are written in their own individual style. That may be a bit of a leap for some (or many) journalists, but regardless, training GPT-4 to know your tone of writing could be helpful in other instances.
It goes back to being able to input more information than its predecessor. If you provide several emails or social media posts that you’ve written before, it can better understand how you talk and provide a better output when you ask for copy later. Before, it was necessary to edit answers from ChatGPT pretty heavily for it to look like it’s coming from yourself or to ensure it’s the same tone of your company. Now, it can be trained easier.
“It’s much better at dealing with nuanced style,” said Cannings. “You can say ‘generate a 500-word article based on the transcript above in the style of the 10 articles I uploaded.’ It learns from you and how you sound and how you write and tries to use your voice in generating the article.”
However, it’s important to note that GPT-4 being able to learn someone’s voice can heighten the risk for phishing attacks. Someone can learn how individuals from your HR department speak, for instance, and then reach out looking for personal and financial information.
5. Ask GPT-4 for pointed, personalized interview questions.
It’s true that AI has changed the game for recruiters. They need to be more adept at knowing when a candidate embellished their resume or cover letter with AI. However, GPT can also help recruiters in that they can create automated messages quicker. With GPT-3 and 3.5, that might have meant asking for a short message to send to potential candidates on LinkedIn in say, for example, a data analyst job in New York City. However, usually it would lack any real personalization.
Now, you could input information from a potential candidate’s LinkedIn profile and website and other social media, to create an extremely personalized message asking for an interview. Now, GPT-4 can output something with the understanding that one candidate has worked in the industry for 10 years, held senior positions, but has also expressed interest in diving into a more consultative role. That personalization can help the recruiter secure more roles than they previously might have been able to and in a shorter amount of time.
However, like all things mentioned on this list, reviewing it is key.
“It’s only good if you can review what the output is,” said Netter. “It’s not perfect, but as long as you review it, it will save a lot of time.”