Earlier this year, LinkedIn co-founder and venture capitalist Reid Hoffman issued an astonished warning about AI. “It’s literally magic happening,” Hoffman said, speaking to technology leaders from every sector of the economy.
Some of that magic is becoming more apparent in creative fields like the visual arts, and the idea of ”generative technology” has caught Silicon Valley’s attention. AI has even recently won awards at art exhibitions.
But Hoffman’s message was aimed squarely at executives.
“AI will transform all industries,” Hoffman told members of the CNBC Technology Executive Council. “So everyone needs to think about it, not just in data science.”
The rapid progress of MicrosoftCopilot AI, the automated code-writing tool from its open-source subsidiary GitHub, was an example that Hoffman, who sits on the Microsoft board of directors, directly cited as a signal that all companies should be better prepared for AI in their world. Even if they don’t make big investments in AI today, business leaders need to understand the pace of artificial intelligence improvement and upcoming applications or they will “sacrifice the future,” he said.
“100,000 developers accepted 35% of Copilot’s programming suggestions,” said Hoffman. “That’s a 35% increase in productivity over last year’s model. … Anything we do we’re going to have reinforcing tools, it’s going to get there over the next three to 10 years, a base for everything we do,” he added.
Copilot has already added another 5% to the 35% quoted by Hoffman. GitHub CEO Thomas Dohmke recently told us that over the past year in beta testing, Copilot has handled up to 40% of the coding among programmers using the AI. In other words, for every 100 lines of code, 40 are written by the AI, reducing overall project time by up to 55%.
Copilot, trained on vast amounts of open-source code, monitors the code being written by a developer and works as an assistant, taking the developer’s input and making suggestions on the next line of code, often multi-line coding suggestions, often ” Boilerplate” code which is needed but a waste of time for a human to recreate. We’ve all had some experience with this form of AI now, in places like our email program, with both Microsoft and Gmail proposing the next few words we might want to type.
AI can logically recognize what might come next in a text string. But Dohmke said: “It can’t do any more, it can’t grasp the meaning of what you want to say.”
Whether a company is a supermarket working on point-of-sale technology or a banking company working on the customer experience in an app, they are all effectively becoming software companies, all developing software, and once a C-suite has developers, they have to deal with the Developer productivity and how to continuously improve it.
This is where the 40 lines of code come into play. “After a year of Copilot, about 40% of the code was written by the AI where Copilot was enabled,” Dohmke said. “And when you show executives that number, it’s overwhelming for them. … calculate how much they spend on developers.”
With projects being completed in less than half the time, a logical conclusion is that there will be less work for people. But Dohmke says another way of looking at software developers’ jobs is that they do a lot more high-value work than just rewriting code that already exists in the world. “The definition of ‘higher value’ work is to eliminate the standard work that rewrites things already done over and over again,” he said.
The goal of Copilot is to help developers “stay with the flow” as they begin the coding task. That’s because some of the time spent writing code is really spent looking for existing code for browsers to inject, “snippets from someone else,” Dohmke said. And that can lead to programmers getting distracted. “Eventually they’re back in editor mode and they’re copying and pasting a solution, but they have to remember what they were working on,” he said. “It’s like a surfer on a wave in the water and he has to find the next wave. Copilot keeps him in the editing environment, in the creative environment, and suggesting ideas,” Dohmke said. “And if the idea doesn’t work, you can reject it or find the next one and edit it at any time,” he added.
The GitHub CEO expects more of these Copilot code proposals to be accepted – up to 80% in the next five years. Contrary to many other things happening in the computing field, Dohmke said of this forecast: “It is not an exact science… but we believe that it will grow enormously.”
After a year on the market, he said new models get better quickly. When developers reject some code suggestions from Copilot, the AI learns. And as more developers adopt Copilot, it becomes smarter, interacting with developers much like a new colleague and learning from what’s accepted or rejected. New models of AI don’t come out every day, but every time a new model becomes available, “we can make a leap,” he said.
But AI is still a long way from replacing humans. “Copilot can’t do the job 100 percent today,” said Dohmke. “It is not sentient. It cannot create itself without user input.”
With Copilot still in private beta testing among individual developers — hundreds of thousands of developers — GitHub hasn’t announced any enterprise customers, but expects to start naming enterprise customers before the end of the year. There’s no pricing information released for companies just yet, but in the beta test, Copilot’s pricing was set at a flat rate per developer — $10 per person per month, which is often spent by developers on corporate cards. “And you can imagine what they’re making per month, so it’s marginal cost,” Dohmke said. “If you look at the 40% and you think about the productivity increase and you take 40% of the Opex spend for developers, the $10 isn’t a relevant cost. … I have 1,000 developers and it’s a lot more money than 1000 x 10,” he said.
The GitHub CEO sees what’s happening now with AI as the next logical phase in productivity advances in a programming world he’s been a part of since the late 1980s. This was a time when programming was coming out of the punch card phase and there was no internet and programmers like Dohmke had to buy books and magazines and join computer clubs to get information. “I had to wait until I met someone to ask questions,” he recalled.
That was the first phase of developer productivity, and then came the internet and now open source, which allows developers to find other developers on the internet who have already “built the wheel,” he said.
Regardless of whether the coding task relates to payment processing or a social media login, most companies – whether start-ups or established companies – use open source code. “There’s already a huge open source dependency tree,” Dohmke said.
It’s not uncommon for up to 90% of mobile app code to come from the web and open-source platforms like GitHub. In a coding era where “everything else is already available,” that’s not what sets a developer or an app apart.
“AI is just the third wave of that,” said Dohmke. “From punch cards to DIY to open source to lots of code where AI is now writing more,” he said. “At 40% soon, as AI spreads across industries, the innovation on the phone will come with the help of AI and the developer.”
Today and for the foreseeable future, Copilot remains a technology that is trained on code and makes suggestions based on looking it up in a code library. It doesn’t invent any new algorithms, but at the current rate of progress it is “quite possible that it will create new ideas for source code with the help of a developer,” said Dohmke.
But even that requires a human touch. “Copilot is getting closer, but it will always take developers to innovate,” he said.