Old Fucking Punk @lwriemen@librem.one

In a recent earnings call Sundar Pichai claimed that at Google now 25% of Code is AI generated (“and then reviewed and accepted by engineers”). In the AI boosterism parts of the web (so basically X and LinkedIn) this number was celebrated: Even Google does AI code generation. So if your whole startup is just ChatGPT in a trenchcoat, you’re basically at the industry standard, right?

Let’s not be cynical here and point at Google’s not exactly stellar recent track record when it comes to great products and software, but let’s ask us where that number comes from and what it means.

Code completion

If you’ve ever written code in an advanced editor or even an IDE (Integrated Development Environment) you already know code completion or code snippets. Code completion is a function of your editor/IDE that has some semantic model of your already written and imported code and can sometimes save you a lot of typing by suggesting a function you want to call. This isn’t magic or super advanced, it just knows that the object named “A” has only one method that starts with “b” called “bwhyisthisfunctionnamesolong” so when you type “A.b” it will suggest to autocomplete the missing “whyisthisfunctionnamesolong” for you. Saves a lot of typing. It’s neat. Snippets are something similar, they are code templates for common structures that you can trigger and that you only fill in the parts that change for your specific case. Also quite a time saver.

So do we think that Google has – so far – tracked how often its users used those functionalities? How much of the code before the LLM revolution already was generated by other code? I’m willing to bet a thousand bucks that they did not. So we don’t exactly know if 25% is a lot more than what their editor/IDE did suggest before. Code completion/snippets don’t get you to 25%, that’s sure but it shaves some percentage off. We just don’t know how much.

What kind of code?

Not all code is created equal. A lot of programming is just doing what developers often call “boilerplate”: Code that sets up the environment, code that does some cleanup, code that translates from one data structure into another to use another external library, code that does data import and whatnot. That code is not irrelevant, quite the opposite (in machine learning the boring code that cleans and structures data is probably more important than the functional code of the neural network!), but often it’s not exactly complicated or requires a lot of thought: It’s just what you need to do to be able to do what you actually want to do. Kinda like putting on pants in the morning.

So when people hear “code” they mostly think of the core, functional code that makes up an application. What developers sometimes call “business logic”. This is the code that crystalizes the developers’ domain expertise and understanding of the problem space and a potential solution into an executable workflow. Did Google generate 25% of that code? The code that determines how the search engine calculates the relevance of a certain website given a certain query? Or was it boilerplate? We have no idea, given what we know about the abilities and qualities of “AI code assistants” my gut feeling is that it’s way more on the boilerplate side that not. Their spicy autocomplete is probably just an expensive way to not have to set up snippets for your code. Which again, will save some time. But it’s also not really as impressive as the sheer number makes it sound.

Side note for those who don’t program: While boilerplate is annoying to write you have no idea how much harder you can make your life by just making a few bad decisions with your boilerplate. A few bad calls can make programming a rather straight forward thing the equivalent of running a marathon in high heels: Doable but probably not exactly fun.

The bonus

When corporations give out those numbers it’s usually to show that their strategy works, that everything is working as intended and that they are growing/innovating/etc. The metrics a corporation tracks are (ideally) directly aligned with their business goals and strategy.

Goodhart’s Law states that a measure that turns into a target ceases to be a good measure. Because people. If you tell people to reach a certain measure they will try to find shortcuts to achieve that measure. If you for example demand your developers increase their code output measured in lines of code, they might start to express their code in very bloated and complex ways to game the metric. The measure is no longer useful.

So Google (like basically all of big tech) has bet the house on “AI”. Everything needs to have “AI”. “AI” is without alternatives. Resistance is futile. But those claims only get you so far, you need to show results. You need to show data. So you tell your engineers to use AI assistants for programming, you even tell them to track how much of code it generates. And to make sure that people actually do it, you tie their bonuses to this measure.

So if my fat bonus check this year depends on my team reaching 20% AI code generation by October, what do I do? I tell people to use the generator for fucking everything to ensure I hit the mark. Whether it makes sense or not. Even if it costs more in cleaning up the generated code in review than to just write it. Because of the goals and incentives that were set.

So of course everyone at Google tells Sundar that they use so much AI to generate code because Sundar told them that that’s what he wanted to hear and that that’s whet gets them a bonus check. He bought himself a useless piece on information. At least with regards to understanding how good AI code completion actually is. As a PR data point it’s great of course.

Conclusion

Google’s numbers don’t say much. 25% can mean that their AI can basically code better than any human on the planet or that it just replaced the use of IDE snippets. And the number is juiced because everyone inside of Google got told and got paid to use AI for anything.

So for us who actually want to understand what the abilities of AI code generators, their use cases and issues are, this number is just a useless distraction. Because it offers to little actual information and data. It’s not much better than just asking an “AI chatbot” how much code it could write. It will give you a number with great confidence. That number is just bullshit.

Liked it? Take a second to support tante on Patreon!

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

https://tante.cc/2024/11/01/what-do-25-mean/

#ai #code #google

Remember folks, we're the good guys (i.e. the US) and the good party is in power (i.e. the Dems) so obviously, this must be good.

"Once again, the United Nations General Assembly has overwhelmingly voted to condemn the US’ illegal economic blockade of Cuba, now in its 62nd year. In total, 187 countries voted in favour of lifting the blockage and just two against: the US and Israel."

https://www.nakedcapitalism.com/2024/11/javier-milei-fires-foreign-minister-for-opposing-us-66-year-embargo-of-cuban-economy.html

How do you organize mutual aid in a city without power, water, and internet?

Don't miss our latest episode of This Is America, featuring an interview with someone from Rural Organizing and Resiliance (ROAR) and volunteers @firestorm in #Asheville.

We discuss how hundreds of people across the city came together following #HurricaneHelene in mass meetings and helped to organize autonomous disaster relief and mutual aid. ROAR speaks about the challenges of mobilizing in rural areas.

We also speak with the two hosts of 'The Dugout' a Black anarchist podcast, roundup resistance news, and talk about how Trump is already moving to contest the next election.

Listen here on Mastodon or go here: https://itsgoingdown.org/this-is-america-202

At this point the US is not just an enabler of Israel but a full partner. We are committing genocide. We are war criminals. Biden and other top US officials should face ICC charges along with Netanyahu.

"Genocide as Colonial Erasure": U.N. Expert Francesca Albanese on Israel's "Intent to Destroy" Gaza

https://www.youtube.com/watch?v=gDeOUFPQf3o

#UN should consider suspending #Israel over ‘genocide’ against Palestinians, says special rapporteur - https://www.theguardian.com/world/2024/oct/31/un-should-consider-suspending-israel-over-genocide-against-palestinians-says-special-rapporteur "Francesca Albanese, who was stopped from appearing at US Congress this week, said Israel defied UN resolutions" #gaza #lebanon

Help, #rustaceans !

What does "unsafe" mean?

I want to pass around raw file descriptors (to image buffers), so that they can be consumed in many places. There are different kinds of descriptors on Linux, though.

Loading a regular file by accident instead of an image won't cause a segfault, but it may error out later when you try to use it.

fn load_image(fd: File) -> Result<ImageFd, LoadError>;

Should the above be an unsafe operation?

#rustlang #askfedi

If you followed the latest #FOSDEM news,
you may have already learned that the
FOSS on mobile devices devroom
is back for another round in 2025 🚀

You can view the Call for Proposals here:
https://gitlab.com/fosdem_mobile/devroom/-/blob/main/README.md

It should also show up soon on the mailing list archive:
https://lists.fosdem.org/pipermail/fosdem/2024q4/date.html

We are looking forward to all submissions 🎉
Submit today! 👌
or at least no later than the first of December ⛄

#MobileLinux #Mobile #FOSS #Mobian #postmarketOS #phosh #GNOME #KDE #sxmo #embedded #FOSDEM25

Software Freedom @conservancy: "Open Source AI Definition Erodes the Meaning of 'Open Source'"

"With this announcement, we have reached the moment that software freedom advocates have feared for decades: the definition of “open source” — with which OSI was entrusted — now differs in significant ways from the views of most software freedom advocates."

https://sfconservancy.org/blog/2024/oct/31/open-source-ai-definition-osaid-erodes-foss/

#OpenSource #FreeSoftware #OpenSourceAI #AI #OSI #OSAID

Every time random assholes call asking if I want to sell my house I tell them I"ll sell it for two million dollars. It makes me feel all warm inside when they hang up in disgust.

Elon #Musk skips court hearing in case challenging $1m swing state giveaways - https://www.theguardian.com/us-news/2024/oct/31/elon-musk-court-hearing-voter-lottery it's almost as if there is one law for him, and a rather different law for us

It's Halloween! Try a podcast about a spooky modern fantasy set in rural Virginia, with magic, creepy rag things, and people who may or may not be faeries...

Episode One: Arrival
https://eviscerati.libsyn.com/the-points-between-episode-one-arrival

Episode Two: Dust to Daylight
https://eviscerati.libsyn.com/the-points-between-episode-two-dust-to-daylight

(New today!) Episode Three: The Chapel of Light
https://eviscerati.libsyn.com/the-points-between-epiode-three-the-chapel-of-light

#mazon donates to group backing hardline anti-#abortion Republican - https://www.theguardian.com/us-news/2024/oct/31/amazon-doordash-cvs-fund-anti-abortion-republican wow, #bezos really wants to be loved, doesn't he?

EU emissions fall by 8% in steep reduction reminiscent of Covid shutdown - https://www.theguardian.com/environment/2024/oct/31/eu-emissions-fall-by-8-in-steep-reduction-reminiscent-of-covid-shutdown "Decline over 2023, helped by switch to renewable power, means greenhouse gas pollution is now 37% below 1990 levels" good to see, but much more work to do #ClimateCrisis

The #Forgejo monthly update was published ✨ It is a high level overview of the project activities.

Forgejo is two years old and has been a lively human adventure, a story worth telling. A hackathon organized by Codeberg generated thousands of new translations. Forgejo v9.0.0 was published, as well as a security patch release which was backported to Forgejo v7. A kubernetes cluster was created to replace the current infrastructure, running Forgejo from the Helm Chart.

https://forgejo.org/2024-10-monthly-update/

Dear #GitHub users, if you prefer the orange-themed heatmap for #halloween and would like to see it all year long: feel invited to move to #Forgejo or #Codeberg (and choosing a forgejo theme in the appearance settings). It's the default colour here.

I skated right up to the deadline, but I managed to submit my proposals to @socallinuxexpo on time! The call for presenters closes November 1 so you had better hurry if you want an opportunity to speak at one of the best Linux conferences around!

https://www.socallinuxexpo.org/scale/22x/cfp

#scale22x

"A societal transition from fossil fuel dependency to reliance on low-carbon energy sources will be impossible without substantially reducing overall energy usage and maintaining this lower rate of energy usage indefinitely. This transition isn’t just about building lots of solar panels, wind turbines, and batteries. It is about organizing society differently so that it uses much less energy and gets the energy it uses from sources that are sustainable over the long run."
https://www.nakedcapitalism.com/2024/10/what-would-a-real-renewable-energy-transition-look-like.html

For my newsletter I decided on Ghost as a platform. This blog post from its co-founder will help you understand why:

https://john.onolan.org/democratising-publishing/

Your daily reminder that if the Biden administration had done anything resembling this -- Trump regime blocked aid to Puerto Rico after its disastrous hurricane -- it would have dominated the news for days or weeks.

As it is, one outlet cared.

Remember as well: This is how a new Trump regime will handle disasters: States that voted for Harris won't get help, period.

https://www.nbcnews.com/news/latino/new-probe-confirms-trump-officials-blocked-puerto-rico-receiving-hurri-rcna749