More than a quarter of all new code at Google is generated by AI
twitter.comOne thing that worries me about AI-generated code is that if there's an obscure bug that pops up later, there's no engineer to think "ok hang on, I remember something strange happening when I first wrote that code... let me have a look." Instead, there are only engineers who reviewed the code, which is of course a lot different from writing it.
> One thing that worries me about AI-generated code is that if there's an obscure bug that pops up later, there's no engineer to think "ok hang on, I remember something strange happening when I first wrote that code... let me have a look." Instead, there are only engineers who reviewed the code, which is of course a lot different from writing it.
I feel this is a very weak argument against AI. Professional software development rarely values crafting good code. You get it in to meet a deadline to make management that is technologically clueless happy. Even orgs that value good code have people leave because of the take a new job merry-go-round to get a good pay raise of years past. One of the reasons open source surpasses most closed source software despite a lack of funding is you a variety of individuals with different goals that are focused on making a maintainable and usable solution.
> Professional software development rarely values crafting good code. You get it in to meet a deadline to make management that is technologically clueless happy.
While this is very common, you also have professional software developers with a deep sense of ownership about a system: they animated it, so when it’s being quirky in particular ways, you have an almost supernatural sense of what branches it’s following. You don’t really get to internalise the logic of a program by reading it. It’s a byproduct of having to come up with it. When a part of that thinking is outsourced, some logic internalisation is lost.
Engineers in tech move between jobs/teams often enough that this benefit doesn’t exist anyway
How common is this situation? While it's nice to look at the history of some line of code and contact the person who wrote it, in a company with a lot of turnover or promotion, that person isn't going to be available or want to help you.
It's pretty handy if they _are_ there, and/or if they've trained others on the codebase.
CEO of company with AI product says, in earnings call, that nontrivial amount of code is generated by AI..
Do people think an earnings call is some kind of secret meeting where the marketing stops?
Nearly a quarter of HN submissions lately are just spammy ads for the AI sector. I seldom open them anymore, but I’ve noticed most of them get hundreds of comments. I wish people spent their time on something useful.
At least people are not openly hostile to the hype train like with blockchain.
PDF of Earnings (31 points) https://news.ycombinator.com/item?id=41988811
Coverage from NYT (11 points) https://news.ycombinator.com/item?id=41989256
Coverage from Verge (7 points) https://news.ycombinator.com/item?id=41989674
Coverage from CNBC (2 points) https://news.ycombinator.com/item?id=41989727
Any good examples of such generated code? Anything to look at on Github, Gitlab, etc?
80 lines of licence comment
30 lines of class boilerplate
One explanatory comment for every statement
The revolution is here!
AI code is the new offshore code. Have fun maintaining it.
Better source article: https://blog.google/inside-google/message-ceo/alphabet-earni...
When function names autocomplete in my IDE, is AI writing my code?
When vim shows me where a matching paren/bracket goes, is that AI that put them there?
When we use code validators like lint or Purify, is that AI code review from the 1980s?
It sounds like a boastful lie for stock reasons, but if not, the AI overlords may decide to write algorithms creating effectively a social credit service and not tell any humans.
Would CEOs of companies selling AI lie about how useful AI is?
In unrelated news, Amazon's CEO boasted that AI saved the company 4,500 years of engineering effort during a JDK upgrade.
Because saying "more than a quarter of all new code at Google is generated by free crowdsourcing from internet scraping" doesn't roll off the tongue as easily ;)
Google has two billion lines of proprietary code, conformant to their style guides and proprietary requirements. I can't imagine they'd poison their model with non-conformant third party source.
More than a quarter of the new code is generated by AI while I need to clarify each time with my PM what does the ticket they opened mean
With so many projects abandoned, is code speed really an issue for google?
Has there been any novel or interesting code written or is this code block completion that saved people keystrokes?
Like most inherently greedy corporations these days, they're figure out yet another way to demoralize, fire, and drive away talented engineers.
It could also be they're using AI to propose and do large-scale refactor code mods, but that would be a smart use-case for it.
Google is a big company which buys other companies to grow.
Does this AI actually improve their M&A somehow? Does it allow them to AI-write new software instead of buying it?
I don't get it.
Your entire premise is pretty obviously wrong. Acquisitions are not the main driver for any kind of growth, whether revenue or the code base
What kind of code? That’s pretty important to know right?
This would be more interesting if it were supported by representative examples from Google open source projects.
Are they using models that are public, or are these customized private models.
Mission critical code? Unit tests? Context matters.
RIP Google
Eh, I would imagine that more than a quarter of new code at any FAANG is boilerplate which is ideal for current AI systems. I'm pretty anti-current-AI, but I can't say that I'm not impressed at how well it handles boilerplate code.
and it shows