The "limited ROI on AI investment" articles will continue to percolate slowly into the brains of the LinkedIn hive-mind until we hit a tipping point, and then we'll finally shut up about how a handy dev tool with some decent use-cases is the dawning of the singularity that will replace all white collar labor and get back to actually building business value.
Linux's LLM policy allows AI-generated code because the project was never vulnerable to this DoS attack in the first place. They kept the social defense they've had for decades.
The killzone here is Github culture, where people decided it was normal to accept code from anonymous randoms with anime avatars. They're doomed.
I’m not saying it’s good for everything, but we’ve gone very far in terms of capabilities in the last 3 years. Thinking otherwise will make me question others’ experience on how much they’ve used it so far.
it would be more convincing if you had concrete examples of technical merit and quality/speed improvements that worked for you or your team that justify going all-in
Priorities.
I manage several teams of developers who use them every day professionally and use them for personal projects privately. Just last week, at the prompting of said devs, we had a working agreement conversation about curtailing the use of AI in our codebases because of rapid erosion of our teams' ability to operate, update and maintain codebases that had started to spill over with slop.
We have had multiple incidents of credential leakage, integration tests wiping live databases, comically broken code that passed vibe-written tests, documentation and code comments that were hallucinated and/or fake, and most importantly developers saying "we no longer know how this code works but it's massively bloated and unreadable and we can't tell you with a straight face that we can maintain it or fix it if it breaks." We have seen a flood of vibe-PRs from engineering adjacent teams that suddenly think they can code shipping prototypes into production that do not work do not scale and cannot be maintained. I am personally writing the tickets to decom one of those today. Which is great, I love telling business "the progress that was reported to you was a lie, this shit never worked, don't shoot the messenger but also don't let this happen again."
I embrace GenAI as a productivity tool for people who know what they are doing. It's a +10-15% velocity boost. That's great! That's a big deal, devs are expensive, and that might push some kinds of business model over the threshold into viability. That's great!
It is not, however, transforming the profession as I know it, it is rather making my job harder and less pleasant and it is making my leadership dumber by the second.
Absolutely no comment on vibe business decisions / vibe OKRs / slop reports or the host of other garbage that has started to creep into professional life. I have had to have some very uncomfortably direct conversations with peers in leadership about using complete bullshit to make decisions, and it is very, very frustrating. Do you know how hard it is to convince someone that metrics their bot hallucinated don't exist and would be meaningless if they did? You can't convince someone of something they are incentivized to not understand. It's been very eye-opening in terms of who I can trust to actually make sense when it matters. I'm grateful for the clarity.
Meanwhile we are rapidly losing brainshare from the top because our principal/staff engineers are pissed off and have the bankroll to just leave. We aren't hiring and training younger engineers to keep the talent pipeline moving. Which boy howdy is THAT going to cost us unbelievable sums of dollars to fix in the medium-term future.
I am in the uncomfortable role of trying to make the best of this but it would be a metric ton easier if the narrative from the c-suite aligned with reality in any meaningful way.
</rant>
when i brought this up with my manager they just said "you need to engineer a better harness" and "aren't you cultivating your agents.md file? thats probably your problem"
which is to say, apparently we're holding it wrong...
y'know, I am sincerely trying, in good faith, to let my teams guide me in how best to intentionally leverage GenAI. I'm trying to use it in good faith myself. I really do reserve the right to be wrong about it.
It's just... the problem was never the ability of GenAI to produce code that works on a superficial level. The problem was always that GenAI would push engineering teams at companies yet further away from what they were building. Generations of compilers, interpreters, frameworks and ORMs have kept pushing engineering teams out there. Cloud infrastructure has pushed engineering teams out there. Giant data analytics platforms have pushed engineering teams out there.
Engineering teams need people who love learning about how things work, but those people often have a rough time in the industry. People who are passionate about software and systems engineering, who love learning how systems work, who love diving deep into niches of reality... those people have been getting hit in the head since they entered the work force.
They're essential, and they've had decades of this crap to learn from. History has taught them (us) that businesses are really, really good at opting in to bloated, obtuse, unmaintainable solutions if some sales monkey can convince leadership that they no longer need to hire smart people who speak a language they don't understand. And they (we) are used to dealing with the consequences. Every single experiences engineer has their war stories about it.
Speaking personally:
I had a partner who worked at a startup in Seattle. They were a novice programmer. I saw what happened when their employer tried to build a Ruby on Rails stack with only junior engineers out of college. It ended horribly.
I worked as a first hire on a clean-sweep purge of a .NET team that had forgotten how to program. The company was bloated and nobody was accountable for anything for years. Then one day they get a mandate to actually build a thing for a big new online initiative. Nobody knew what they're doing. Wrappers around wrappers around wrappers of copy-pasted stackoverflow code. The company hired a hard-ass director to reboot things because nothing was getting done. He mandated a total shift to Golang/Kubernetes over .Net/some managed Microsoft stack, and said "you'll prove you can do it or you're out". I was hired because I could. I was substantially younger than everybody, but asked to write architecture docs and establish design patterns and sit on hiring panels for all the new devs. Because I don't know, I was a nerd who learned how things work and gave a shit and had somehow picked this knowledge up. They fired 8/11 of the devs on that team, as well as the project managers and the dev manager. We grew a new team behind it that was 1/3 the size and moved immeasurably faster. I hate everything about how that whole story played out. I'm not proud of that chapter of my life. I resent in particular the leadership who allowed engineering fundamentals to erode over time because it wasn't their problem.
I've worked at companies that fled to the cloud from local infrastructure. They did it with this almost insane desire to stop paying people to solve complex infrastructure problems. Balls-out, fuck these nerds who actually know how hardware works, they're weird and we're sick of them, we're going to pay Amazon a billion dollars to never have to hire them again. And then they hired some (often quite competent!) devops folks to migrate everytrhing... those people did... things got wildly expensive... meanwhile businesses locked in to stupid technologies like DynamoDB and Lambda because some insufferable sales idiot connected with a CTO in a sweat-yurt... and et voila, here we are writing the proposals to unfuck our cloud migration and return to local infrastructure. Yay CapEx. Boo OpEx. Venture Capitol batteries possibly included. Just gross incompetence by a generation of LinkedIn leadership buffoons who don't know the first fucking thing about how computers work.
I've been there. You've been there. LLMs are one more way of being there. We feel it in our bones. We have noses and we can smell. Trust was lost long ago. This shit is old. And we're really tired of it.
And this time it's especially annoying because it's big and tied in with an economic bubble and companies laying people off and general major disgruntlement. It's a really bad vibe right now for a technology leader, and if you're running a company I hope you get it, at least a little bit.
I really give a shit about being a good leader in technology and I'm really, really frustrated right now because there an almost un-navigable divide between what I believe to be true (from my experience) and the narrative I'm expected to sell. That's a hard position to be in. But I (and my peers) try our best to occupy it with grace.
</rant>
To anyone reading this: If you couldn't see this coming three years ago, you don't deserve your job.
> We have seen a flood of vibe-PRs from engineering adjacent teams that suddenly think they can code shipping prototypes into production that do not work do not scale and cannot be maintained.
What a nightmare. Never forget that AI is for idiots.
I haven't written a line of code myself in 6 months, but the point is that _you shouldn't be able to tell_
What I tell people at work: Using AI is good. AI can help you do things that you otherwise wouldn't have. It shouldn't be a crutch for thinking.
In that sense, people shouldn't be able to tell that you're using AI unless the tell is that it's _higher quality_ than if you had implemented it by hand.
A good example here is a well-written document that has a _ton_ of deep research behind it. Or code that has extensive tests that would have taken too long to write by hand for the task at hand.
It's bad when it's just "do this thing and post a PR" but I write my code with AI, look at it as a reviewer, offer suggestions, put those into my steering if needed, and I refine it. The way I see it, I'm the first reviewer on everything now before passing it off to another human for review.
There have been points where having AI rewrite something was slower than me doing it, but I'm already in the harness. The cost of waiting is almost zero. What stops me from contributing to the JVM using AI written code that I understand and spent time manicuring? Can a human really tell that it's AI written?
It would be interesting to see how this effects the copyleft Licences with contributors are using AI for the PRs
Which recent verdict?
> It would be interesting to see how this effects the copyleft Licences with contributors are using AI for the PRs
I know Linux and GCC have been diligent about tagging and tracking LLM based contributions. In the worst case they can chuck it all out and handwrite it back.
There is this old one around contents like images and all, the ripple affect is pretty much everywhere.
Its sort of coupled with the recent penalty of $1.5B on Anthropic. The catch 22 is that some portion of the LLM training data can be classified as IP theft. Even though Anthropic has been fined the data still exists and can be used by LLM to generate code for you. So if you are claiming something as an IP and it has stolen part in it then it leaves you in a hard place.
Language translations may save you in some cases, though the whole definition of cleanroom has been in debate recently too where people are trying to rewrite opensource/famous libraries in different language and claiming IP rights over them.
Settlement not penalty.
Anthropic will have judged the benefits of the settlement, not just the headline cost. It could have been a strategic move by Anthropic: we can't know without information we don't have. https://news.ycombinator.com/item?id=49014389 1.5B looks like ~2% of funding/income.
Secondly the future is still not here yet, the people who came forward to sue are mostly in the category of book publishers/authors. The tech companies are are not actively searching for copyright thefts as of yet, however I am sure its just a matter of time when the big blobs of codes get rediscovered specially in case of any publicly visible code
Other copyright lawsuits against other major companies are happening.
If a similar lawsuit costs OpenAI more than 1.5B then perhaps Anthropic is better off in their competitive race?
> the people who came forward to sue are mostly in the category of book publishers/authors
Irrelevant. The financially savvy parties are the lawfirm L.L.P.s trying to win money for themselves.
As for AI, certain core teams, notably the Database team, are forbidden from using LLM coding tools.
Refusing to use AI honestly seems more like vanity to me. Like "no machine could ever do what I do".
That said I would not necessarily accuse people who don't want to use AI of vanity - there are good reasons not to want to use it - but if someone else did it I'd find it easier to understand.
Imagine if I exposed my API keys and just said “sorry boss, I took a pill before the job and I hallucinated”