Samuli Hokkanen

Let’s admit it together: the AI tools released in the past few months, such as ChatGPT, Midjourney, and the underlying machine learning models GPT-3, GPT-4, and Stable Diffusion, are some of the most interesting developments in the software industry in years. They evoke emotions and big questions about the future.

My initial experiments with these new technologies caused childlike excitement and amazement – how is this even possible? I created incredible-looking portraits using the Dall-E tool in the styles of Helene Schjerfbeck and Akseli Gallen-Kallela. I asked ChatGPT to turn a shopping list into a poem written by Eino Leino and tried to debug a geothermal heating system’s settings with its help. Especially after regularly using ChatGPT, I found myself thinking of the language model operating on the other side of the command line as a human-like agent.

Six Fingers, Lost Literary Sources, and Nick Cave

After the initial excitement, I started noticing errors – characters in images having six fingers and a certain mechanical strangeness that’s hard to define, but still recognizable in these outputs.

For work, I did some background research and asked ChatGPT to find data and name sources. It did well – providing statistics and research names in an organized response. However, upon Googling, it turned out that the credible-sounding studies named by the virtual hand-waver didn’t actually exist. The command line apologized for its errors when I inquired further about the missing literary sources.

In the larger world, songwriter and artist Nick Cave asked ChatGPT to create lyrics in his style for a song and delivered fiery criticism of the result (and the technology).

The Developers Best Friend?

Although ChatGPT (at least for now) has not proven to be the universal genius to be unreservedly relied upon like bedrock, language models trained for specific purposes seem to perform astonishingly well. At Crasman, we offered all software developers the opportunity to use Github’s Copilot language model, trained on billions of lines of code, which provides ready suggestions based on the developer’s comments and previous code.

Experiences with Copilot also seem to be mixed based on the initial comments: on one hand, coding with Copilot is seen to bring significant benefits and time savings (we calculated that if the tool saves half an hour of work time a month or avoids a few bugs, it’s worth its price), but some developers feel that Copilot does not offer significant benefits – it often misunderstands the context and intentions and can be “in the way” with its suggestions. Both viewpoints seem well justified.

So, we’re not at perfection yet, but I bet a language model like Copilot will stay in our tool kit permanently. The pace of development is, indeed, astounding.

To Infinity and Beyond

Despite the six fingers, clumsy Nick Cave lyrics, slippery consultant citation practices, and tech-bro buzz, I just shaved off a subscription for both the diffusion model Midjourney and ChatGPT Plus service from my bank card. Despite the hype, fluff, and questionable applications, I’m convinced by the pace of development. I strongly believe that we are living through the “iPhone moment” of these technologies and I want to be on the pulse when there are five fingers, diffusion model-produced images are genuinely nice to look at, and citations are spot on. We are inevitably moving towards that moment, and the latest GPT-4 model is already an extremely useful tool.

I strongly recommend the same to any business leader. It’s certainly good to consider how these tools, their next generations, and explosively growing use cases will affect the business we do.

P.S. I still use the term “AI” somewhat reluctantly, are we at the point where it’s the right concept? The selection of “AI solutions” from the snake oil salesmen of the past has left its mark on the word.

This article was originally published on the blog of Software Finland ry.