For various reasons, it's been a while since I've been able to present at conferences or workshops. However, this year I've had the pleasure of giving the keynote at 3 events: the Systems Research Challenges Workshop, MakeIT & JCON and the Cloud Control Workshop. They've all been to different audiences (researchers, open source developers, academic and industry collaborators) but on a common theme: has AI made us/them irrelevant? I figured I'd try to write up some of what I was trying to convey.
Everyone is asking the same thing lately: are we finally redundant? The hype is everywhere. You see these demos of AI spitting out code at a speed no human can touch, and it feels like the game is over. However, after looking at the reality of it, I’m not losing sleep over it. Not yet, anyway. Here’s the thing: there is a massive divide between syntax and architecture. AI is brilliant at syntax. It can churn out boilerplate and individual functions by pattern-matching training data faster than you can grab a coffee. But it doesn't understand the "why." It doesn't understand the architectural vision. For example, it’s like having a mason who can lay bricks perfectly but has no clue what a cathedral is supposed to look like. AI struggles with the big picture, complex system design, avoiding conflicts with existing patterns and keeping codebases from becoming bloated, unmaintainable messes.
Given my background, in the limited time my current job allows, I still spend a lot of time thinking about fault tolerant distributed systems. If you’re trying to build a high-performance database with ACID properties and "five-nines" uptime, you’re in one of the hardest domains in software. This is where humans still win: we understand trade-offs between performance, consistent and availability. What about AI? It’s estimated it would need 500,000 extremely high-quality, expert-level training examples to even begin to be trustworthy in that space and that kind of data isn't available in the public repositories which have been used to train the frontier models. Much of that code is closed source and was written years or even decades ago.
Furthermore, the "other side" of the AI story is pretty messy. We don't talk enough about the accountability gap. Have you actually read the disclaimers for these tools? For example, in one version of the Copilot licence Microsoft literally says it is for "entertainment purposes only." (OK, they later walked that one back!) However, typically all of the AI vendors make zero warranties. If the AI hallucinates a non-existent library or leaves a massive security hole, which happens in about 45% of generated snippets, it’s your head on the block, not theirs. Then there's what I like to call the "Dark Star Effect": the total lack of a moral or ethical framework. It can reproduce copyrighted code or build-in biased patterns without a second thought.
Fortunately it’s not all doom and gloom. AI is a powerful tool: it can save 50-70% of your time on the boring stuff, such as automated testing, code reviews and finding bottlenecks. But it’s a co-pilot, not the captain.
Our jobs are changing. We’re becoming translators of ambiguity. We take messy human problems and turn them into precise technical requirements. AI can help us write the grammar, but we’re the ones who have to tell the story. Don't let the hype convince you that you’re irrelevant; we’re the ones dealing with the legacy systems and integration debt underneath. Stay focused on the architecture. Let the machine handle the syntax.
No comments:
Post a Comment