Listen to this story
GitHub Copilot was the first product to show that LLMs could be applied in real world use-cases, prompting a gold rush of sorts to add AI to developer tools. While many incumbents powered by traditional technology still exist in the market, few challengers have emerged with AI-powered features.
The future, however, is ripe for disruption, with the burgeoning creation of an AI-fuelled Dev Tools 2.0 ecosystem. From setup, to implementation, to deployment, AI tools are making human work more efficient while eroding some of programming’s biggest inefficiencies.
Disrupting programming with AI
GitHub Copilot was a ‘eureka’ moment for developers, as they realised that AI had matured to the point where it can be integrated in developer workloads without too many issues. This phenomenon grew to the point that around 46% of code today is written by the AI pair programmer, helping developers code up to 55% faster.
Similarly, OpenAI also wished to gauge the impact of LLMs on labour markets. Recently, it released a research paper determining the effects of GPT-like technologies on labour markets. Through their research, they found that 80% of the US labour force could see 10% of their work impacted. When adding LLM-powered tools and software to the mix, researchers found that up to 56% of all tasks could be completed “significantly faster at the same level of quality”.
As any programmer will know, writing the code is just the first step, with the process only becoming more complex as the life cycle continues. To aid this process, companies have created various offerings to make debugging, committing and reviewing code, hosting, and observability a lot easier. While few offerings have set the gold standard, such as AWS for cloud hosting, GitHub for asynchronous programming, Stack Overflow for debugging, and Asana for issue tracking, these so-called standards are falling behind the times.
For instance, Stack Overflow, which has built up a database of known errors through human interaction, along with a community of veteran developers, is ready to be replaced by Adrenaline. Instead of going to Stack Overflow for bug fixes, Adrenaline allows devs to chat with an AI assistant to find and fix bugs in their code.
These advancements suggest that software might be the first field to be disrupted by AI. In the words of Paul Kedrosky, an investor at SK Ventures, “The current generation of AI models are a missile aimed, however unintentionally, directly at software production itself. Such technologies are terrific to the point of dark magic at producing, debugging, and accelerating software production quickly and almost costlessly”.
AI outprices human programmers
The reason for software being ever present in the current day and age is a result of reducing costs for running them. From a room-sized computer in the 1960s to a pocket-sized computer today, technological advancement and the economies of scale have done wonders for infrastructure. The graph below shows the plummeting cost of resources such as storage, bandwidth, and CPU compute.
A significant datapoint in this graph is the relatively high value that is ascribed to programmers when compared to other infrastructure needs. While developers have generally been the most expensive part of the value chain in software development, AI holds the potential to change that.
Shashank Kare, a software engineer with over 25 years of experience in the field, had this to say about the rise of AI in software,
“Software will be developed by small teams going forward and AI will handle all day to day work. I think it is the end of the line of frameworks, OSes, languages, platforms etc. They were being designed for humans to work smarter and faster. With machines there is no such challenge.”
Indeed, this is an important point to note. AI pair programmers don’t need the things that make life easy for human developers, and have shown themselves to be capable of interfacing with code in a more ‘organic’ manner. In a research paper titled, ‘Sparks of Artificial General Intelligence’, researchers from Microsoft found that GPT-4 was able to execute some pseudo-code directly after inferring its purpose. Moreover, it was able to do so without translating it into other programming languages.
In the future, an advanced enough AI pair programmer could function completely off natural language prompts. This is just the impact felt from LLMs being deployed in an extremely narrow use-case. A world where the entire software development life cycle was carried out along with AI assistants is what dev tools 2.0 aims to be.
Moreover, with innovations like ChatGPT plugins, devices can simply plug and play the components they need for their architecture, allowing various AI models to interface with each other under the supervision of a small team of software developers.