Everyone’s talking about AI. What do we here at Visual Assist expect? It’s clear it’s incredibly relevant – debate on that ship has sailed – but it’s hyped. We take a brief non-hyped look at where we think things will go.
An AI-generated image built at simplified.com with the prompt ‘A laptop with an AI onscreen’. Impressive it’s possible, but equally it has a long way to go. Images and movies are harder than text. Are they harder than code?
Today’s AI
- Runs on remote servers / the cloud
- Shares knowledge (unless you pay 10x as much) which makes it inappropriate for confidential data like source code
- Is computationally very expensive: the hardware required is expensive and you need a lot of it
- May not be relevant: it’s trained on wide data, but you may have something more specific, such as your own libraries and source
However:
- There’s a lot of work being done reducing the computation cost: for example, Facebook’s Llama and related models have a large number of tweaks that bring it down to single-computer levels at reasonable speeds
- Once it can run locally, accidentally shared knowledge is not an issue (at the same order of magnitude of cost, you could also have an affordable in-house server, shared only with employees)
- Once it can run locally, it can be used any time, not just with a net connection
- It may be much more relevant: if it’s trained on your data, it will be able to suggest domain-specific solutions
What’s key?
For commercial software development, three things are key:
- Keeping your source code confidential when getting AI input
- Low cost
- Giving domain-specific, your-app-specific useful results
We see local and private AI trained on your own codebase becoming a relevant reality quite soon. Compared to a remote cloud-based ChatGPT (for example) instance, that seems far, far more useful, practical, low cost, and safe.
Visual Assist
Tools like Copilot or ChatGPT, which are cool but legally risky for the owners and potentially dangerous for you if your source is added to their knowledgebase, are not the way forward. Plus, an AI that knows your tech is far more useful than an AI that knows generic programming. Local, private, trained-on-your-code AI is where we see industry relevance.
The above is our view on where AI will go for development. As for what this means for Visual Assist, none of this can be taken as a statement of product direction. It’s best to say that we are interested in the topic. VA already provides industry leading refactorings and other tooling powered by a unique code understanding engine, a tool developed to be non-compiler-like and more programmer-like. AI’s potential for features based on code understanding syncs very well with what we provide. Without hype, and moving carefully, we may see movement in this direction. If we do, as always it will be with the Visual Assist ethos: an eye towards true usefulness, not headlines; great performance; and the features we choose will be designed by devs, for devs.
