We recently hosted our latest roundtable to hear from our local CTO and VPE peers on how the future of generative AI is going to impact the developer landscape.
In the next few years, will the impact of generative AI on development represent productivity gains or full replacement? Where is it offering the most value to developers today? What are the subsections of generative AI that still have significant gaps? And further: how are engineering leaders currently driving education and experimentation inside of their teams? How are employees being provided with the guidance required to accelerate the desired experimentation while avoiding new risks? What are good leadership practices around this today?
The experts chimed in with some great feedback, and early use cases around what they’re seeing in the market.
For Now, It’s Definitely Only Productivity Gains, Developer Jobs are Safe – Github CoPilot and other similar solutions still generate a lot of errors and bugs – they also need more analytics than they have today. Still, productivity gains are already coming down the pipeline. These tools are starting to impact:
- Dev Iteration
- Test Automation
- Code Quality
- Code Security
What Kind of Code You’re Writing Matters – GenAI logic is very general. Business logic tends to be different. Segregating code base architecture and separating those two logic structures could make AI solutions much more impactful
Everyone Becomes a Manager – It won’t replace people, it will augment them. Even junior people have to “manage” the AI. This may eventually lead to a restructuring of teams and empowerment of junior employees
This Will Help Ops and Business Teams Leap Into Technology – E.g. product will start contributing to engineering, other teams will lean into these advances. Documentation and text-based work is the most real right now – marketing is a bigger fan than the devs
Expectations from CEOs are High – The reality is a lot more mixed, but we all have to get on the train
The Focus Needs to be on Inexact Use Cases – Great for avoiding the blank canvas problem – gets you 30% of the way there, etc. Things that need binary answers like aspects of dev and security are not good use cases. The tech is still too fuzzy and optimization is tricky
Source Code is Not Your Secret Sauce – It’s not necessarily a big deal to share this with genAI tooling – proprietary or customer data is another matter. Much of this is already in the cloud – is this tooling going to be any different?
On-Prem is a Niche That Will Need to Get Filled – Major competitors like OpenAI and Anthropic are avoiding it
If Models are Commoditized Then Data Becomes the Differentiator – Protect your data
Today’s Guardrails Are Really Soft – Most companies do not centralize access to software through IT – ultimately, developers are going to be downloading and accessing what they want. It’s a speed issue. Larger companies are going to have larger guardrails. LangChain may become the defacto standard to prevent data from leaving the building
Engineering Productivity Will Be Rewritten – DORA 2.0 is going to emerge over the next few years – the companies at the vanguard of generative AI will be reshaping these metrics (e.g. Accuracy, Freshness, etc.)
Code Generation May Not Be the Killer App – Business applications today look a lot more promising. There will also be better communication between business and dev – translation of requirements from what a business person is looking to produce in terms of code, and vice-versa
Early Days Use Cases
Dev and Math: Copilot, Code Interpretation, Source Code Documentation
Generative Art: Midjourney, Adobe Firefly, DAHL-E
Operations: Predictive Maintenance, Access to Ops Data Sooner, Test Case Generation, Logs, Tech Documentation/Internal Knowledge Bases, Document Summarization
Security: Security Operations, Interpretation of Video Footage, Security Design Review, Threat Modeling
Human Resources: Employee/Engineer Onboarding, Resume Upscaling and Generation
Data: Analysis on Disparate Data, Reduction of Manual Work for Data Stewards, Stringing Data Sets Together