Am I suffering from a serious case of copium or is tech journalism seriously out of touch with reality when it comes to AI?

[-]

Raunhofer@reddit

It doesn't help that the people who get paid by the success of AI are literally lying about the capabilities of it. Thus I wouldn't blame it all on journalists.

[-]

Meta_Machine_00@reddit

What percent of people from that poor eastern European country did not get to do what you did? You sound incredibly lucky to have had that opportunity.

[-]

dats_cool@reddit

Sure bro. I'm just lucky. Luck plays a factor in all things in life but my hard work and determination is what got me here.

[-]

Meta_Machine_00@reddit

Plenty of people put in more hard work and determination than you and don't have any of the opportunities that you have access to. It is all luck.

[-]

dats_cool@reddit

LOL. I mean whatever helps you sleep at night. Nothing was luck, there was a lot of deliberate studying, all nighters, moving across the country for a low paying first gig.

That's sad that you have that perspective. So do you just not try for anything because you think luck is everything?

[-]

Meta_Machine_00@reddit

No. Free will is not real. We only do what the physical laws force us to do. Some people are lucky and are forced to do perceivably better things while others are forced to suffer.

[-]

dats_cool@reddit

Okay, I'm over this conversation.

[-]

Raunhofer@reddit

Achtually due to quantum events that occur in our neurons, free thought seems to be a real thing.

The FitzHugh-Nagumo equations and quantum noise - ScienceDirect

Our brains sense the future through a new quantum-like implicit learning mechanism - ScienceDirect

[-]

Meta_Machine_00@reddit

Randomness/non-determinism are not freedom. That just means you are random, not free.

[-]

Raunhofer@reddit

It means LLMs and our brains do not function the same way. Our thoughts are obviously not random gibberish, it's just that the process leading to them has a kind of "salt" that cannot be predicted, making it the closest thing to free will we can scientifically have.

Free thought is real, you are not a predictable machine. No-one and nothing knows how you'll respond, if you will at all.

[-]

Meta_Machine_00@reddit

Non-predictability is not freedom. Freedom is being able to choose independently from more than one option in a single instant. Occam's razor asserts non-freedom because any choice will be algorithmically assigned and there is no chance for the other options to be chosen in the instant. Your bias just makes it hard for you to abandon the illusion of "freedom".

[-]

Raunhofer@reddit

Freedom is being able to choose independently from more than one option in a single instant.

Non-predictability is the basis for free will. With our desires and intentions, this extremely complex, non-deterministic, and not fully understood mechanism is what drives us.

It's the near-magical nature of LLMs that is the illusion. You absolutely can reverse-engineer every character it types out. ML is said to be a black box because it's not practical to reverse-engineer, not because it's impossible.

You can now answer or not answer, use your free will in its full glory. ChatGPT can't, it will answer.

[-]

Meta_Machine_00@reddit

No. If you answer then you are forced to answer. Just because it is asynchronous and there are pauses in the middle does not mean it is not automated.

And you also explain that things are unknown, but then you leap to non-predictability = free will. You literally say that you know that the unknown is what drives us. That makes zero sense.

[-]

Raunhofer@reddit

You are cherry picking my words, non-predictability alone is not free will, just something that needs to occur, please read the full sentence.

You are not forced to answer even if you answer. You strip all meaning of the word forced if absolutely everything is forced, which is not the case due to pre-mentioned non-deterministic nature of neurons.

I understand that it is easier to grasp the concept of free thought by oversimplifying it into something totally deterministic and simple, but that's not how our brains work and that's why LLMs are not the way to how our brains work.

[-]

Meta_Machine_00@reddit

If you don't understand it as you've already admitted, how do you know that something is being oversimplified? You are justifying your knowledge with ignorance.

Nonetheless, randomness can be forced. Non-determinism is simply the appearance of phenomena with zero preceding information. Just because something is random does not mean that appearance of the phenomena at a given time was not a mandatory random generation.

[-]

General_Platform_265@reddit

take your meds

[-]

Meta_Machine_00@reddit

So you think I need to use medication to alter the operation of the brain to produce a different result? You are just validating what I am talking about.

poolpog@reddit

As an infrastructure folk, I too, say don't trust it. LLMs aren't even quite at the trust but verify stage yet for terraform or IaC

[-]

DeterminedQuokka@reddit

Honestly, I was surprised she likes it. But if it’s working for her, good for her.

Ugh. I’ve had no success with dashboards. It’s constantly making up metrics that don’t exist in datadog. It has never given me a single real metric that aws sends. It’s so annoying.

[-]

poolpog@reddit

LLMs are not just not good at terraform they are legit bad at it.

[-]

OskarSarkon@reddit

Imo it's a significant problem for stuff like Terraform that's more niche than your average JS frontend or Python API call script, as a related example I've been on (multiple!) incident calls that were derailed by people repeating made up but correct-sounding nonsense from ChatGPT about Elasticsearch

[-]

a_reply_to_a_post@reddit

I couldn't really leave the house last night to go to the dispensary for a 50% off sale they were having because my wife was at the darkroom printing photographs, and i was kinda annoyed about it and ended up downloading cursor and messing around with it to see what the hype was about

i've been paying for old ass servers to host a few old websites, my wife's portfolio site that I built over 10 years ago when she was my girlfriend, and actually did build her a new version that's like modern typescript/react but i just have some json files

i've had an idea in my head for years about how i can stop paying for these old servers and just store content on S3, and in about 2 hours with cursor I had a working Electron app that publishes to an s3 bucket...couple more hours and it was kinda robust with features like drag and drop reordering and static API generation

for quick prototypes, this shit is kinda fun if you just chat with it and be ok / guide it through improvements

i think for greenfield projects / one offs, shit is kinda sick, but in my day to day work i don't really want to lean too much into using AI to write my code because otherwise it makes the job basically reviewing junior code all day and asking them to make a bunch of changes

[-]

coworker@reddit

What you did is called the agentic AI workflow and like you said it works great for prototypes or smaller things. This is conceptually like being a manager delegating whole tickets to junior engineers.

Cursor also supports the manual AI flow where you are actively coding like normal but using AI in a very targeted manner for small, specific things. This is like working on a problem normally but having a principal engineer available to consult on a whim. This style works extremely well when for large, established code bases because you are in full control.

So which AI slop are you selling?

[-]

m0j0m0j@reddit

I’m selling a service where it’s not apparent at all there’s any slop in it, as it all is under the hood

[-]

AntDracula@reddit

Lmao of course

[-]

m0j0m0j@reddit

What “of course”? And Ronacher is building some email thing, where AI slop is only in development, not even in the product.

I'm using Cline (VS Code extension). You can bring your own LLM. I use Claude Sonnet 4.5.

I'm doing a lot of AWS infrastructure as code in recent but have also done a lot of Python with it. When it's wrong it's wrong but it's pretty rare. The irony is that I'm building internal AI tools so.. using AI to make AI.

As far as cleanup? Hardly ever.

[-]

SignoreBanana@reddit

When execs can tie AI to increased revenue, then I'll buy the hype. But so far, not a SINGLE company has.

[-]

RedWinger7@reddit

100%. This is the reality of the situation:

https://pracap.com/global-crossing-reborn/

Hue hue hue

Our main guidance at work “don’t trust AI, it will break all the rules you set”. It is great that it can write code. But if I am spending more time to describe and then clean up, than if I wrote the code myself, then AI is worthless.

[-]

Strict-Molasses4816@reddit

AI is a bloated, expensive, and unsustainable screen-scraping "macroservice" that, when all is said and done, summarizes and copy-pastes from stackoverflow and similar websites very, very quickly. Putting AI summaries at the top of web-searches robs these websites of traffic, even though the AI payload was distilled from their content.

I'm not worried about AI taking away any serious developer jobs, and it will never actually do DevOps or Systems Administration autonomously by any stretch of the imagination.

Maximum-Objective-39@reddit

I originally heard it called the 'Next Page Effect'

Article about something you know about - "This is stupid!"

Flips to next page - "My gosh, a woman gave birth to an ostrich egg!"

[-]

marcel_in_ca@reddit

Ya beat me to it

https://en.wikipedia.org/wiki/Gell-Mann_amnesia_effect

[-]

Beneficial_Wolf3771@reddit

It’s called “access journalism”. Anybody with the stones to actually push back in their article or ask hard questions during interviews gets instantly blocked from further interviews, and that reputation will instantly kill a career.

[-]

Cutlesnap@reddit

oh cool, now I know what it's called

[-]

There are many writers in tech journalism who are calling out the AI bullshit and the financial bubble for what it is; but most of y’all aren’t clicking those links. People are going for hype, they go for fear-mongering, even if they don’t believe it (hate-clicks are a thing).

Audiences (especially Gamers) trust the eager-beaver streamer with a friendly affect who feeds them the corporate approved hype over journalists (because uncritical hype and sensationalism is actually what the audience want to hear), and then the journalists who follow the same access-hype path further degrade the reputation of journalists who don’t do that shit. Go figure.

[-]

prisencotech@reddit

It's not about "objectivity" becuase that's a nebulous term, it's about people who are trained and educated in journalism who are allowed to be independent and given the time and resources to investigate.

The bloggers won and David Simon was 100% right about them.

[-]

Opposite-Cranberry76@reddit

Do not mention the war.

[-]

johnpeters42@reddit

Oceania is at war with Eastasia. Oceania has always been at war with Eastasia.

[-]

Immediate-Badger-410@reddit

It's more that "journalist" is being slung around and not people that publish articles for reputable sources anymore. Toooons of internet articles now are by people that are by no means informed or experts on the topic. Much like reddit does the I'm a gun in the plumbers sub reddit so I know how to do plumbing! But they have never done any plumbing in their life.

[-]

PureRepresentative9@reddit

"don't believe everything you hear on the internet"

"On the internet , the men are men, the women are men, and the children are FBI"

are hilarious relevant lol

[-]

MoreRopePlease@reddit

My favorite: "on the internet nobody knows you're a dog". From a cartoon circa 1991

[-]

troublemaker74@reddit

My take is that journalism has become more of a way to rank for keywords, make money, and regurgitate things that people want to hear more than it is about actually reporting news.

[-]

Arkanin@reddit

It's Gell-Mann amnesia. They were always ignorant, but you know more about tech every year.

[-]

It’s the AI promoting the AI. This is how it gains power! /s

[-]

this_is_a_long_nickn@reddit

Bold of you to think that a human wrote the article

[-]

Lyraele@reddit

Sadly, they’ve always been dumb, going back to the 80’s. There’ll be a non-dumb standout every now and again, but mostly it’s been access journalism all the way down for 40+ years now.

[-]

flowering_sun_star@reddit

There has never been a time when this wasn't the case. You just didn't know enough about anything to notice it. As your own knowledge grows, you manage to catch it more.

[-]

KallistiTMP@reddit

Haha, journalists, funny joke!

The AI scenario kinda describes my workflow now.

90% of my time is conversing with AI about the code as to actual coding.

Now, to be clear - many of the type of tasks are perfect for agentic AI. For example, we had to migrate out of OpenShift to Kubernetes/Terraform. AI handles this really well because it has well defined evals to iterate over, and it's pretty much a greenfield.

But to be clear, there are people out there who assume it's run once and forget it. Not the case. Context Management and constant feedback is a huge part of my workflow. Yeah, I let AI write code, but I have about 100,000 tokens of instructions/memory files and pre created evals (often times scripts created by AI) for the AI to iterate against. And I watch the AI changes like a hawk. I interrupt a lot, refine, correct, etc. It's not hands off, and it requires a lot of expertise that has been built up over the years.

It hasn't eliminated me out of the picture, but it has doubled or tripled my productivity. Will everyone get that increase in productivity? Absolutely not. But does that type of workflow exist for certain people/projects? Yep.

It takes a lot of time and effort to understand what it can and can't do. And yeah, it will be frustrating. It took me a 6 months of pretty much solid devotion to it to go from untrusted autocompletion to the majority of my code.

[-]

fireflash38@reddit

I apparently review a shitload of AI code from our India offices. It's always someone dumping a thousand line change with every other line a comment describing the next line of code. Almost always with some real questionable design decisions and error handling.

The last one I reviewed had python code that alternated between catching/raising exceptions and returning 1/0/None on each call level.

[-]

FindingEastern5572@reddit

I like the 'apparently'.

[-]

Metworld@reddit

I'd auto reject such changes tbh.

[-]

Tenelia@reddit

It's the new oligarchy that has come to power over the past 40 years. They have the power, not you.

[-]

AchillesDev@reddit

A little bit of column A a little bit of column B.

It is amazing at writing code, and people I know (and myself) that are pretty deep in the AI/ML spaces are using it very heavily. We are mostly staff+, have had or have leadership positions at various tech companies, and almost all are or have been independent. We successfully heavily use AI to write code for us. That doesn't mean we don't plan, do heavy design, architecture work, etc. though.

The other part of it is a lot of devs are bad at using AI. So it's not as widespread as people are saying, but those who are able to use it well are using it heavily. If you don't trust it, or you haven't explored how you can use it effectively, then just like any other tool you won't be able to use it effectively.

[-]

amxudjehkd@reddit

Yesterday, I took a freelance gig for integrating payment feature in a legacy, vanilla PHP codebase. I used Claude Sonnet 4 as me copilot agent. I was surprised with few results. But overall, I had to go through the codebase to ensure redundant slop was not generated. Though it was good for automated tests, minor docs.

[-]

hoffsky@reddit

I'll keep you updated. The plan at my company is to use "agentic AI" to take tickets, turn them into PRs, do the first PR review and then hand off to a human to do final review and deploy.

[-]

iscottjs@reddit

Yeah similar deal here, I lead a small dev team and my boss is always pushing AI to improve “productivity”, despite us already using various tools to augment our work in reasonable ways.

Anyway, I decided to play around with the Codex CLI to see if we can go from JIRA ticket to PR via GitHub actions.

Inspired by this:

https://cookbook.openai.com/examples/codex/jira-github

I wasn’t expecting this to be good, but it has surprised me with simple tasks even with terrible tickets. I ran a demo of this in front of the team, testing it with real client tickets, things like add this new optional field to this form, change this page copy, add this new search filter, add this new sortable column to this table, add this new reporting criteria, fix this search bug.

Most of these examples worked fine and raised reasonable PRs with backend and frontend changes, and unit tests attached. I needed to tweak the ticket descriptions slightly to give Codex a bit more technical direction.

The code was actually fine because it was mostly inspired by the current code style and linting tools, but I wasn’t impressed by the unit tests. I find AI unit tests to be lazy and brittle even if they pass.

That said, I think this solution is workable for small simple tasks, especially boring tasks that are mostly mundane.

However, for a laugh, one of the devs asked to try passing Codex all of the tickets he’s been working on this week to implement a new payment integration and see how well it could handle it. It’s very complex business logic in this particular system.

These tickets were well written enough for a human to figure out, but unsurprisingly the unattended AI pipeline failed hilariously. To be fair, it did submit a working PR with working flow using sandbox payment details and passing unit tests, but it completely ignored all the important business logic changes we needed to implement, it introduced new bugs because it misunderstood the requirements then doubled down on its own bad direction.

I’m sure I could prompt my way around this, but for complex tasks I feel like there’s diminishing returns on how much effort goes into writing an entire book of prompts versus just doing the work mostly yourself.

Which I why I feel like for complex large tasks, I’d still feel more comfortable having a team of humans collaborating and building a stable solution with good foundations, but I do think it’s fine for experienced devs to use AI to speed up any tedious code snippets with a human-in-the-loop workflow rather than a full blown unattended agentic system doing whatever it wants.

If an AI agent drops a PR for a scary feature with 100 files changed, I just know that’s going to be a painful review.

[-]

Perfect-Campaign9551@reddit

What a waste of time. Probably 90 percent of those PRs will just be tossed out

Also doesn't that mean you'll have to write tickets like a full on book? All details and everything so the AI has proper context

[-]

MoreRopePlease@reddit

Also doesn't that mean you'll have to write tickets like a full on book? All details and everything

It's like when you offshore development. I don't see this as an improvement.

[-]

Meta_Machine_00@reddit

I was forced to write that comment. The reason you see it is that we cannot avoid the events that we observe in this universe. The universe creates your reality, not the other way around.

[-]

shill_420@reddit

The name is goated, not gonna lie. He's also a great journalist. He series on hyper scale data centers, "Data Vampires," is a must listen if you want to understand how damaging these technology is both ecologically and politically that is being forced fed to us.

Worth the listen, real journalism is still being done but you ain't gonna find it hosted by corporate media.

[-]

suprjaybrd@reddit

It'll depend on your company and how cutting edge it is. My company is not AI native, but every engineer uses combinations of cursor, claude, and copliots as part of their development now.

- For easy things, claude can oneshot small linear tickets with minimal cleanup afterwards. Compared to 6 months ago - what it can oneshot has dramatically increased.

- For projects with tech specs, you can point the AI at your spec and it can try to scaffold parts of it for you. You can point it at your designers figma file too.

- AI is really good at writing repetitive things like tests and mocking - nearly all our engineers use this.

I'd say right now, at my company, the code being written is probably 30% AI generated (we have no Jr engineers - and the overall quality of the PRs are reasonably high still). Across the engineering team, the understanding is this is going to be an essential part of the SDE toolbox going forward so you had better pick up these skills. I don't know if / when it'll turn into just supervision - but the % of AI generated code being checked in has been increasing.

[-]

MoreRopePlease@reddit

is really good at writing repetitive things like tests

How do you ensure that it's writing good tests that adequately test the business logic without testing the implementation?

[-]

suprjaybrd@reddit

- it reads the docstrings of the function to test for a slightly higher level conceptual view.

- it can generate serviceable unit tests.

- higher level integration tests require more specific prompts otherwise it may not understand the business logic .

[-]

Cheap_Childhood_3435@reddit

The real reason? Because AI company CEO's are saying it, and the tech reporting is on that. To me the interesting question is why they are saying it. Is it that they are trying to justify the painful amount of money they spend on it? I'm sure that's part of it. Is another part of it because Ai companies don't really have a model for making money yet other than selling their LLM access, when another company is trying to break in by distributing LLM access for free and neither company has a viable product so they are trying to market to other CEO's the dream of reducing headcount to salvage their company before the gravy train runs out? I dunno but it's worth noting I have seen reports that OpenAI will run out of runway in 2027.... Now what is their business model again?

I do use AI to write code, all the time. boiler plate stuff or sometimes a front end if I don't feel like doing it, and it's simple get information from this API and display it. Where I don't use it at all is security or for that matter when the performance of the code has to be optimized. AI has a place in writing code, but it's a tool in the tool kit not a replacement for the whole thing. But because companies have not figured out how they are going to monetize it yet we are in the situation we are in where people are promising the moon to keep the funding coming in and tech reporters don't really have access to people on the ground in tech so they talk to the person trying to keep their company afloat, or someone excited about getting rid of staff to maximize profits

[-]

Nervous-Tour-884@reddit

It honestly has been a lot like this for me in the last month, but I think the tasks I am doing are pretty well suited to AI. It will 100% screw things up, and I need to go back and fix, debug, and test the work it does, but it really has been a superstar.

My work has been to basically take a library full of really old React/SCSS/Javascript components and migrate them to a new library and update them so they use Inline styles, no scss, typescript with good typing, sever from SCSS, sever from old dependencies, etc in preparation for future work, all while maintaining full backwards compatibility and not breaking current usages.

It has done an excellent job so far. I don't think I could read, update, and understand everything that is going on nearly as quickly without the help of AI. It makes a lot of this process a breeze, doing things like mapping hex color values to tokens, understanding how I may need to adapt components to accept various props rather than relying scss psudeo classes, just a long list of things.

The thing is, it isn't just doing it. It is basically me breaking everything into chunks, doing it all a piece at a time, having it create plans for me to review, updating plans, and ya, basically supervising it into following the practices and conventions I want it to. I test, and when I run into bugs, I have it help me debug the problem, which it is has been actually surprisingly capable of helping me with. It will intelligently take information I give it from the DOM tree and console and screenshots, and with the right information(like how the dom tree looks on a working version) and context, often come up with a great hypothesis for why an issue is happening, and often will give a good solution to the issue on the first try. When it don't, I continue to work collaboratively with it, having it add console logs, debug statements, telling it what the change it made actually did, and it usually gets figured out.

Giving it the right information, and breaking things down into manageable chunks is key. I don't just go in and tell it to convert something, I have it plan it, I review and revise it, I have it move it, update imports, I test some, plan it, review and revise, migrate to typescript, test, and so on, all the while making sure it does things in the way I want it to. It does make mistakes, particularly with the process of converting SCSS into inline styles and logic around dynamic styling, but it isn't anything I can't work through.

I don't know that it is like this for everyone, but in the last month, 90% of my code is AI written, but I don't think my current work is typical of most development work.

[-]

polotek@reddit

So I understand where you're coming from. We're nowhere near the majority of code being written by "supervised AI". But I do think the people Casey is talking to are increasingly working that way. I live in the San Francisco Bay Area. The hype here is out of control. And people are definitely pushing a model where you set up infrastructure of parallel AIs doing tickets in your backlog. And the job of a developer is to review the PRs that the AI submits.

I don't think this is sustainable. But I think they're gonna try it for a long time before accepting that it's not the right model. And I will also say that I do think something will ultimately change about how the developer role is structured. I just don't think anybody knows what it will actually look like.

[-]

The problem we run into, among others is that generally when you are reviewing code from people especially senior developers there is an intrinsic sense on what portions needs to be reviewed in detail as they are problematic. There are no such assurances with Ai generated code. Normally you are ok, but there are enough instances where errors exist in code a senior developer would never mess up. As a result or pr process needs to be much more defensive. Coupled with the fact that only easier tasks can be initiated from jira and the additional ticket details and context in the end it is very much a drag

[-]

prisencotech@reddit

Exactly. People say "oh, people screw up too!" but we screw up in ways we all understand. AI screws up in new and novel ways we never considered.

The way I put it is: Doctors make mistakes all the time. And yet, nobody ever has a doctor prescribe them a drug that doesn't even exist.

[-]

darkslide3000@reddit

"people screw up too" but we screw up in ways we understand.

coworker@reddit

This is a problem inherent to reviewing other's code be it from a human or AI. You just think you can trust humans more than the machines which is often the case until you have to work with bad engineers.

PR reviews should have always been defensive but reviewers often take shortcuts based on trust. And this has always led to bugs getting through

[-]

I'm wondering how interviews will work in the future if all people have experience in is "prompt engineering".

[-]

fruxzak@reddit

And people are definitely pushing a model where you set up infrastructure of parallel AIs doing tickets in your backlog. And the job of a developer is to review the PRs that the AI submits.

Multiple FAANGs are already working this way today. My personal experience is that they're quite good in some cases but also quite bad in some cases.

[-]

noticeparade@reddit

not even close to replacing radiologists because when a radiologist looks at any scan he observes simultaneously dozens of things not just the one specific thing. And the models obviously focus on what they are trained. So in order to really replace a radiologist you´d needed dozens of models working together which simply doesn´t work.

Radiologists also read one thing (or system, or layer) at a time. There are many limitations that prevent AI (now and in the future) from replacing doctors but this is not one of them.

[-]

E3K@reddit

You're saying what I said. Nobody is saying AI will replace doctors. AI makes good doctors better.

[-]

_ECMO_@reddit

I am happy you think so but you definitely did not say that in this comment:

showing that AI gets diagnoses right at a better rate than experienced doctors.

Independent-Chair-27@reddit

Pretty sure AI could write articles for the tech press. Question is what would it say about AI?

What the article says isn't that complementary on AI really.

Software engineers supervising AI to produce the code they need to solve the problem they have. Sounds a lot like a low level tool to me. A souped up auto complete. Sounds like the article over sells it.

[-]

Desperate-Ad-9348@reddit

I'm shocked people don't agree with AI is amazing and used to write most code. I don't think most people on reddit are caught up with what fortune 100 is doing with tools like Claude and Cursor.

It really is quite a bit beyond what most are saying. Just yesterday a senior engineer gave an example of what they did. They were new to their company and blown away once they started using their AI tools.

[-]

SoftSkillSmith@reddit

Dude, they're just parroting whatever silicon valley execs are stating in their marketing material verbatim. Speak truth to power? Oh please, how about ride the hype train together with the people driving it until the whole thing comes crashing down just so you can feel like you're ahead of the curve!

[-]

Organic_Height4469@reddit

Not only tech journalism. Research / scientific Journalist. Maybe any journalism.

You cannot expect the brightest kid in the class to choose journalism if you know you are going to get paid shit and possibly get killed. Oh do not forget; you get rewarded for the clicks your article get, not if it is actually the truth; Oh wait actually it can be written by ai now so your writing skills also no longer matter; Oh and you used to get a photographer, but now you have something called stockimages.

Oh you get censored by your joke of a government as well. Oh and you get also censored by the corp that sponsors the article.
Oh and the extreme right and left now think everything is your fault too, because you are too woke.
Oh now elitist are going to blame your for writing shit articles.

[-]

Potential_Status_728@reddit

Paid publicity, that’s all.

[-]

sawser@reddit

In an interview for a devops role, I had a line item that said "Researched use cases for AI improvements to CICD pipelines.

The interview went great and then he asked about that line.

I told him I had it on there for the sake of HR Screenings, but that I thought AI was no where near reliable enough to handle the complexity of enterprise dev ops and couldn't do anything better than dedicated systems that currently exist do.

It's security nightmare and a technical black hole and should be avoided at all costs - at least in the near term.

He agreed completely and I got the job. Not sure I've met anyone who disagrees with me.

[-]

CodeToManagement@reddit

The reality is AI isn’t perfect but if you know how to write the prompt it can do a lot for you.

As an example I’m working on a side project and needed a new endpoint adding to a .net api with entity framework to persist the data. I just told it “keeping with the standard I’ve laid out with other endpoints make an endpoint called X which takes parameters Y and Z and saves them to the database”.

It got it perfect and the code was exactly what I’d have written. Only it did it in about 30 seconds. Wouldn’t have taken me too long but it’s deff a 5-10 min task in multiple files.

[-]

pickle9977@reddit

It’s Human Resources tech companies selling Humans as Resources software that does “stuff”. It’s the truth because everyone knows no one in hr is actually a human and none of them actually do anything of value, they just do “stuff”

[-]

zayelion@reddit

Its that they can see that someone (AI) is more skilled than them, but they are not socially skilled to properly rate the skill set. I've even seen seasoned engineers fail at this. You have to use the product to understand its weaknesses. At the current iteration AI preforms like a ADHD junior programmer on an illegal stimulant. That saddle is what most small companies need and can afford. Not having the full suit is like not having car insurance on a teen driver in a sports car.

The hype is to a degree real in that it's going to replace juniors because it already has. I don't know about mid levels. That tech is under current development, and it struggles. Senior would mean the collapse of all white collar work.

[-]

evangelism2@reddit

AI is better at coding tasks than basically anything else....

Id say it better at writing English than code, its also really good at parsing intent which can be applied in many other things just as much as code.

I talk to a lot of software engineers and what they will say is that it used to be that we would write code, and then it moved to we write half the code and it gets autocompleted,

true

and now we just supervise the code and we type in the box what kind of code we want the machine to write

feels like a non developer was told about agentic coding and this is their non technical summation of it. Not exactly true, but not totally false either

[-]

TheElusiveFox@reddit

I'd say three things...

Most media these days is basically advertisement, very few journalists are actually doing research and looking for an unbiased review, especially when it comes to products, they are being paid to advertise, or at the very least they are making things as sensational as possible for views for their other advertisers which tends to lead to, if not out right lies, writers focusing on the very edges of what is possible instead of what most people's actual experience is...
I also think that right now saying anything anti-ai at companies like google or meta, is a very quick career suicide move, so the developers being asked these kinds of questions feel a lot of pressure from management to put things in the most positive light possible. Like auto-complete features exist, and when the a.i is auto completing your line, it mostly works fine, when it is auto completing your function though... results are mixed at best, and even devs who love A.I. aren't using it this way outside of the simplest most basic kinds of scaffolding code...
Like you say, you have to realize the people bringing journalists in are A.I companies, either big ones creating a.i. models who have a lot to gain by companies using them constantly, or small ones creating tools that have a lot to gain if tech companies tell their devs they are expected to use them for all their tasks...

[-]

MountainVeil@reddit

You're definitely not alone. Ed Zitron has been harping on this exact thing for the past year or so, the failure of tech journalism. It's at the point where it's getting repetitive to read. Link if you're interested.

https://www.wheresyoured.at/

[-]

thashepherd@reddit

I'm sympathetic to his take, but

it threatens even the seemingly-infinite wealth of Elon Musk, with xAI burning over a billion dollars a month and planning to spend tens of billions of dollars building the Colossus 2 data center, dragging two billion dollars from SpaceX in his desperate quest to burn as much money as possible for no reason.

I do believe there is a place for these tools in alleviating the most boilerplate and repetitive programming work. The millionth Spring controller that collects the same basic user input or whatever.

The ability for AI to create something novel doesn’t exist. The ability for AI to think critically or reason also do not exist.

But people really want to believe that human technological progress hasn’t stalled out. I’d hazard that many people’s entire world views hinge on the belief that whatever faults our society may have, that we may tolerate them and be patient as we are nevertheless making advances that will obviate many of these problems.

And some just want the quick money, whether it’s investment money in a shaky concept rooted in hype or just some ad revenue on yet another credulous puff piece glazing Jensen or Sam.

I mean, you should be worried about your software job. In the short term big tech is slashing headcount to free capital for AI data center boondoggles. In the long term the AI bubble is the reason economic metrics aren’t all flashing red(der).

[-]

Sfacm@reddit

It's not journalism it is advertising!

[-]

coddswaddle@reddit

Might I recommend Ed Zitron. He's an experienced tech reporter and he's got an entire newsletter and podcast dedicated to the grifters and harm of AI As well as the general enshittification (also friends with Doctorow). He loves tech and hates what venture capital and grifters have done to it

[-]

HansProleman@reddit

Tech journalism in general is hopelessly uncritical. Virtually everyone is happy to just print PR lines, and boosters are incentivised (by valuations, money, continued employment) to produce a lot of positive PR lines.

There have also been a lot of doomers doing the inverse, but until recently the "Maybe this is a bubble/has relatively little value?" stuff was considered unsexy - now it's getting quite a lot of traction.

You might try reading Gary Marcus and Ed Zitron.

[-]

apartment-seeker@reddit

I am finding it to perform like that for frontend work atm. React, TypeScript. A lot of training code available for LLMs. Unfortunately, a lot of it is bad, but it still works :shrug:

AI is better at coding tasks than basically anything else

Person you quote has a logical error going on. What he goes on to say/show is that AI meets a very good threshold of writing code. What he doesn't address is whether "AI is better at coding tasks than basically anything else", e.g. whether AI is better at coding tasks than at legal contract interpretation, researching something using web search, etc.

Shitty writing

[-]

jcm95@reddit

[-]

Difficult-Field280@reddit

AI CEOs love to hype their product and are feeding the online media with talking points to do so because more hype means more users means more profit.

The media doesn't understand how any of this LLM stuff works, and I'm incredibly skeptical that the CEOs do, mostly because all this AI stuff is so new.

So the CEOs stir in hype about features and etc. Plus, a little bit about AGI and boom. The entire tech community has a ton to talk about. Which drives more users and more profit.

Rinse repeat every couple months

[-]

MrCallicles@reddit

Journalists in tech are kinda adversizer most of the time,
It's hard to to otherwise. In hardware, it's even more clear: a company has a new product with some characteristics, and journalists will just create article about that, saying that it's "the brand new CPU with x core etc.".

It's information, but at the same time advertisement.

For LLMs and alike, even scientific community is not aligned on how to measure performances of a model (benchmarks are good, but those are well, benchmarks) so I wouldn't blame journalist to not know how to handle it.

Also, people (scientists, CEO, programmers) all have really strong opinions about all this. So, even if we had "real" investigation journalists , then it would be hard to gather quality information with investigations methods like interviews.

It would be nice to have a little bit of scepticism though, but critical discourse in the field is really the exception not the norm.

[-]

TomarikFTW@reddit

Reminds me of 3D printing.

The news coverage spoke 3D printing everything. You won't buy anything, you'll print it! Everyone will have one in their homes.

Prosthetics yea! Food Yea! Organs yea! Houses Fuck Yea!

While us in the hobby were losing our minds trying to get consistent prints of Buddha Batman.

The same thing is happening with AI. Especially when it came out.

Cancer cure yea! Mathematical solutions yea! Replace workers Fuck Yea!

The reality is recently we had an OpenAI model ingest a dataset to "forecast" sales in the next quarter(s).

This motherfucker was giving us weather reports. Like straight up a breakdown of weather for the week in some arbitrary location.

[-]

Economy_Solution6371@reddit

Whenever a journalist covers a topic you have some expertise in you notice that they dont know shit and are just parroting the current mesaage. Tech is no different

[-]

aman_of_means@reddit

I know it’s cool to be a cynic, but I genuinely feel this way about AI coding right now. It’s really good. It writes probably 90% of my code.

[-]

Historical-Egg3243@reddit

they're not spending obscene amounts on it. All the money is going in a circle, there's no actual exchange of money going on.

[-]

adibrad@reddit

Im a bit disappointed by the highest upvoted comments here. Im a senior dev at a company that went in hard on an AI transformation strategy and I consider myself a sceptic in most of the use cases that get brought up (why do we need to switch our HR system for one with an integrated chatbot?). However, for software development, it has really improved lately.

The ways I use it day to day are: - kick start my research into a topic using Claudes research mode - quickly create PoCs which attempt to do the same thing using multiple different approaches so I can see which ones are viable - triaging bug tickets, by using an llm with access to our git repo through an MCP and asking it to create a report on the most likely cause - searching through confluence/jira/repo/slack for that useful bit of information I saw months ago but can no longer find -to assist me with code reviews (I also review it myself, but more often than I'd like to admit, it catches things I would've missed)

The thing is though, not only is it a lot faster than me at doing these things, whilst it's doing it, I can get on with all the other shit that needs doing.

Now, for actually writing code in a large production application, I think this is where people get confused. Im not just giving it the Jira ticket and telling the LLM to implement it. Im going through and doing the investigative work to figure out how I would implement it, breaking it down into smaller tasks and then give that plan to the LLM and provide it with context from the relevant files. I also ask it to write the tests first before the implementation and to make sure they pass. It can then iterate by itself until you get a build that works and tests that pass. I then go and review the code as I would any other PR, except when I spot something Im not happy with, I highlight the related lines and just tell the LLM how I would prefer it done.

The number of lines of code I have written personally in the last few months has plummeted. I would estimate maybe I have personally written 100 lines in the last month. I still design everything though, and am fully aware of all the details of how it was implemented.

AI is nowhere near close to replacing developers and I am still extremely sceptical that will ever happen as much as it is management's wet dream. That doesn't mean that it can't make you a lot more efficient already though.

[-]

dudesweetman@reddit

Its the old classic thing about people with the least amounth of knowlegde having the loudest loudspeaker. Nothing unique.

I personally believe LLMs are an amazing thing but we are still at the point where it can be compared to the dawn of popular internet with 56k modems. It was a game changer that sparked a bubble but the capabilities where nothing close to what the expectations where.

[-]

Prize_Response6300@reddit

I don’t think there is a worst niche of journalism than tech journalism. It’s probably the most removed from the actual topic so you have Jenny the comms grad from west Kentucky state trying to explain why Claude code changes everything

[-]

Bakoro@reddit

The entire AI thing is "A Tale of Two Cities" right now. Everything is true and nothing is true.

That's not even hyperbole. People are getting paid to vibe code, and they are making things people actually get value out of. People are getting paid to vibe code and are making a complete mess of the business.
People are using AI to code and can't get it to do anything. People are using AI to code and elevating their productivity to previously impossible heights.
People are doing vibe science and becoming deluded with science fantasy. People are using AI to make real advancements in math, science, and engineering.
AI will make people wealthy beyond reason, and poor beyond hope.

The whole spectrum is true. AI as it stands now is a force multiplier, if you're stupid it reinforces that, if you can filter the signal from the noise and be creative, it will accelerat anything you do, and if you're somewhere in the middle, it's a dead zone where it's just a fun toy.

[-]

Top-Size-7395@reddit

Since when journalist aren't out og touch? They are sock puppet for people in power for most if not all industry

[-]

maxip89@reddit

Who is paying them?

[-]

k8s-problem-solved@reddit

I've put in maybe 8 hours this weekend on my side project. I'm using a combo of cursor and github copilot.

I've created docs, open api specs, api implementation, integration tests, data model, react front end using core ui, auth0 identity, map box globe rendering, open telemetry everywhere, playwright test, docker ci/cd with github actions, deployed to azure.

It's an insane pace to work at. I've kept it very vertical slice and it's recognising the patterns better now so I can just carve out whole new screens super easy. Ui, api endpoints, database scheme and all tested in a matter of minutes

I'm happy with api quality. I need to review the react work a bit more, I'm letting it smash the UI out fast - it's not my strongest side so leaning on this a bit. I know what I want, I know the patterns, but it would take me ages to write it myself.

I'm massively impressed with the tools from this session. I've been using them in a spec driven development flow, and the results have been strong.

[-]

Ashleighna99@reddit

You’re doing it right: spec-first, vertical slices, and tests to keep AI honest. To keep that pace without regrets later, wire compliance checks into CI: Schemathesis or Dredd against your OpenAPI, plus Pact if other services consume your API. For frontend quality, spin up Storybook and have Playwright run visual and interaction snapshots per component; pair that with strict TypeScript and an ESLint rule set to block any: and implicit any. On observability, propagate the W3C traceparent from React through your API so traces link end-to-end, add semantic attributes, and define SLOs with error-rate and latency budgets so regressions fail the build. Security-wise, run Semgrep and OSV-Scanner or Dependabot on every PR, and use GitHub Actions preview environments on Azure to review changes with real auth flows via Auth0. I’ve used Hasura and Kong for quick APIs, but DreamFactory helped when I needed instant REST on existing SQL with RBAC without building auth from scratch. Keep shipping vertical slices and lock in quality with these guardrails.

[-]

LiterallyInSpain@reddit

I have been working in the tech industry since 1997, been programming since 8 years old, and have worked for major large companies. I have designed critical software that scaled to over 100m users.

I can tell you that if you can’t get AI to write good software for you at this point, it’s a skill issue. It’s insanely powerful if you know how to prompt, know the patterns to use, and know how to provide the right files for it to read, and manage the context window.

Context Engineering is the skill to learn. If you can handle that, it will do unreal stuff for you. But it’s not a silver bullet if you don’t have the skills to know what you actually want it to do.

[-]

Drited@reddit

Interesting, could you please share some of the approaches to prompting which you find to be useful?

[-]

b34t@reddit

Read Ed Zitron's newsletter. Here's a link to the latest post: https://www.wheresyoured.at/the-case-against-generative-ai/

Tech journalists spend all their time around "tech" so they have a false sense of understanding.

Everyone is out of touch with AI.

It's an insane bubble and it's going to be like the dotcom burst when it inevitably happens

[-]

SouthRock2518@reddit

I'm in exact same boat as you. I keep feeling that I am maybe just not getting it and can't figure out how to get 10 automomous agents doing a bunch of work in our codebase. I can't even get one autonomous agent, GitHub Copilot with Claude 4.5, to do anything of substance. And I'm trying to give it good context, telling it to write out a plan in markdown that I can check and then firing it off only to fight with that damn thing for days and not even get a working result. I love the auto complete and in IDE agent mode because I can do things in incremental steps and catch it going off the rails early to redirect it. I think that's awesome and fun TBH. But sometimes I just feel like I'm missing something. Like oh maybe if I just used Claude code then I would see the results that I keep hearing about.

[-]

flavius-as@reddit

Journalists are paid to make money.

[-]

Meta_Machine_00@reddit

I actually believe everything emerges non-deterministically, aka randomly. Even in random universe, autonomy and independent action are not possible. And if humans disappeared tomorrow, then there would be no one left to care they were gone. So "nihilism" is perfectly logical. It just conflicts with what has been programmed into your brain.

Nonetheless, where do you think your own words are coming from?

[-]

It's real. The hard part is remembering all of it. There's too much output to keep up.

[-]

JimDabell@reddit

I talk to a lot of software engineers and what they will say is that it used to be that we would write code, and then it moved to we write half the code and it gets autocompleted, and now we just supervise the code and we type in the box what kind of code we want the machine to write

That’s just an oversimplified description of a mostly true scenario. But that’s not the typical case, that’s the best case scenario – an expert developer / agentic AI / spent lots of time learning how to use AI properly / strong oversight over what is being written / frontier model / burning lots of tokens / the right language / the right use-case / finishing the last 10% themselves. You take any of those things out and the success rate drops. But it’s not completely out of touch with reality – and if you think that it is, then you do have your head in the sand.

If you disagree, then read how these developers work with AI:

And watch these live coding sessions:

Casey Newton and Kevin of the Hard fork podcast literally had an episode a few weeks ago where they claimed to dedicate the hour to discussing criticisms of their show's perspective on AI.

Do you think that would have emerged as the concept for an entire episode if they weren't sensing their audience bristling at the amount of ball washing they do in support of the AI industry's exaggerated claims?

But, I honestly believe this is a problem of modern journalism. If Newton were to be a skeptical muckraker when it comes to the claims of AI companies, he would lose all access to his high profile guests and his podcast would be toast.

[-]

Last-Daikon945@reddit

Journalism has been dead for almost two decades. They write whatever comes from the top

[-]

datOEsigmagrindlife@reddit

I think you're in a slight bit of denial.

I'm at a large tech company, household name, not FAANG but similar level.

Vibe coding or whatever you want to refer to it as, has been pushed company wide.

Is it perfect, no.

Can someone with no SWE experience use it to make something, no.

Does it still produce slop, yes, but with the right architecture document, well written prompts and context from MCP it will give good to decent level of output that can iteratively be improved into a working product much quicker than human only output.

With the right work flow, it is absolutely having an enormous impact on productivity.

[-]

yubario@reddit

A lot of the developers that do use AI quite effectively do not post here because its a really bad echo chamber here, like 95% of the posts are just dogpiling how shitty AI is. And any pro AI comment just gets downvoted to hell, so yeah....

Wow! A lot of downvotes. Ok, let me put it this way to all the downvoters… it is possible that all those witnesses are either (a) AI company CEO’s (b) technology journalists, or (c) liars.

Unfortunately, it is more possible that you are missing something and there is not a conspiracy afoot to gaslight you

[-]

hippydipster@reddit

but in no way do I trust it

Trust has nothing to do with it. We code reviews our human coworkers. We make test suites. We hire QA people. We break tasks as small as possible so even we can do them. It's no different whether the code was first written by senior, junior, autocomplete, or LLM.

Trust got nothing to do with it.

[-]

ZorbaTHut@reddit

Am I just in denial? Does this accurately reflect how you are using AI these days?

Sometimes, yeah, frankly. I do review it, and I do edit it, and more than a few times I've had the AI make major refactors, and sometimes it's just easier to do it myself than to ask the AI to do it.

But it does keep getting better, and it's great at working with code that I'm not familiar with - far better than I am! - and every month or two I give it a few tries flinging something at it that it never would have worked before, and each time it's able to figure out something it wasn't able to do the last time.

I just added an entire new custom feature to prusaslicer, and that involved me going to Claude and describing what I wanted; Claude figured out how to implement it and got it right on the first try, along with the two next adjustments to it. It made a mistake on the final adjustment and I had to debug. This still probably saved me hours (or rather, if I had to do it by hand, I just wouldn't have done it.)

[-]

talldean@reddit

I'm able to get 10-20% faster overall, using AI to code things... where AI is good at those things.

So "hey, I got lazy and didn't write tests for this class, . Can you write some tests for it, but be careful to test the functionality and not the implementation? If you need an example of what my team thinks is a good test, has some of those. If you're not sure about something, I'd be happy if you'd ask me questions."

It then cranks out 75% usable code, and gets the boilerplate right every time.

[-]

devanew@reddit

The "AI writes most code now" narrative is classic hype cycle journalism. Yes, some companies report high percentages of AI-generated code, but that stat is misleading without context.

The human-written code is doing the heavy lifting: architecture decisions, complex business logic, debugging AI hallucinations, code reviews, integration work, and handling edge cases AI can't reason about. AI is great at boilerplate and simple functions, but I'm still the one designing systems, making trade-offs, and fixing the subtle bugs AI introduces.

The "we just supervise code" quote is particularly ridiculous. Good luck "supervising" AI code for correctness, security vulnerabilities, performance implications, or maintainability without deeply understanding what it's doing. That's not supervision - that's still engineering.

Journalists covering AI often lack technical backgrounds, so they accept founder soundbites at face value. It's the same pattern we saw with blockchain ("revolutionary technology disrupting everything!"), crypto, and before that, VR/AR. The incentive structure is broken - AI companies need to justify massive investments, journalists need clicks, and nuanced takes about "useful but limited tool" don't generate either.

Don't get me wrong - AI coding tools are genuinely helpful. But there's a massive gap between "helpful productivity tool" and "we just tell it what to do now." Anyone claiming the latter either isn't writing production code or is setting themselves up for a world of technical debt.

[-]

Existing_Guidance343@reddit

I’m a principal engineer at an SME co-owned by two large multinationals. AI isn’t a huge part of our day to day — we’re only allowed to use Gemini, which struggles with our combination Rust / Python / TypeScript codebase — but we’re currently desperately trying to hire after some unexpected (but easy to predict) churn.

My boss has sent me a Gemini Gem (basically a pretrained model) to use for screening takedown tests and even after three or four hours of tweaking at it it’s consistently crap. So I just ignore it now.

Heaven forbid that we start to do this model of ai-based coding. It’d be an absolute shitshow.

[-]

OkCar7264@reddit

Any field of journalism where the main advertisers are the subjects of the articles is just marketing with extra steps. Games, car, tech, not to be taken seriously.

[-]

PressureAppropriate@reddit

Reading a journalist's description of something you actually know about is an eye opening experience...

It's not rare to find something that is just plain wrong on almost everything they write.

But then you flip the page and they talk about a topic you don't know that much about and you somehow delude yourself thinking that now they are saying the truth...

[-]

SpicyLemonZest@reddit

I think there’s a lot of people who just see a massive psychological difference between spending an hour writing your own code and spending an hour correcting AI-written code. When I envision “AI is writing all my PRs”, I imagine shipping dozens of PRs a day because I’m no longer constrained by anything but my imagination. I suspect you think the same as me, but that’s not what the people who go around saying it seem to mean.

[-]

BrianThompsonsNYCTri@reddit

Here’s how collective memory of predictions works, someone who is critical of something that ends up being big is forever remembered as a dunce whereas praise of anything that turns out to be a flop is forgotten. To this day bad takes about the capabilities of the internet are frequently brought up, but criticism and praise of something like NFTs or Zune or whatever is basically forgotten about. As such it’s a lot safer for tech journalists to fawn over whatever the “next big thing” is because if it flops nobody will remember their praise but if it succeeds everyone will remember their criticism.

In my experience, AI is able to produce code on the level of an (below?) average software engineer (junior/mid) as long as the use cases can be defined well in natural language and when working in isolation. Haven't tried it yet for big code basis. However, to me it seems that it produces a pretty low quality structure of the code.

To me it's becoming more and more clear that AI will not replace SWE, but will empower us as a new super cool tool that does not only all that boiler plate code but also a lot of low-level code that I don't want to think about too much. It allows us to spent more time thinking about higher level challenges, the actual problem at hand, what the right abstraction would be, what a sensible integration into the given code base looks like, how to make the code extensible, etc.

[-]

SupermarketNo3265@reddit

100%

I'm still putting pieces together, building something, and problem solving. Just now maybe I don't need to build each one of my Lego blocks from scratch.

[-]

TheMightyTywin@reddit

You’re in denial - human coding is dead.

GPT5 with extended thinking, and Opus 4.1 are better programmers than 99% of humans.

Their output still must be reviewed because they struggle to take in the big picture. But that’s also true of code written by humans.

[-]

TastyToad@reddit

There are about 8 billion people alive. There are less than 50 million programmers, according to google. Therefore, LLMs are better programmers than 99% of humans, that are not programmers in the first place.

(I'm being sarcastic and rude but you sound like a 15yo edgelord, not an experienced dev)

[-]

TheMightyTywin@reddit

Fair enough - I've updated my claim.

[-]

PickleLips64151@reddit

The problem with most journalism is that they don't know enough to ask the hard questions. They get dazzled by the wow factor and don't see that it's all smoke and mirrors.

About the only tech journalism that I find useful are product reviews, where there are benchmarks and firm comparisons between the known and the new.

[-]

Now realize all the other topics it’s the same you didn’t realize it.

[-]

My experience with Anthropic’s Claude is pretty amazing and it’s like having a great coding partner, who often thinks of approaches that I don’t. Claude can get caught in unproductive ideas and get overextended, but so can we all. It really falls down whenever the task involves complex judgments about the real world.

The ideal case for delegating to an AI is when you have a greenfield project and you need to do something well defined and straightforward. I can just tell it to write a Github action to build my Svelte/Rust project and I don’t have to read lots of docs.

[-]

gfunc@reddit

I think the problem is in calling them journalists. They’re a cog in the clickbait machine that is now so large that it doesn’t create anything of substance anymore and may never again.

[-]

Trick-Interaction396@reddit

Metrics are super easy to manipulate. CEO says AI is mandatory. This implies all new code is AI assisted. The reality is unknown.

[-]

djkianoosh@reddit

super hyperbolic

for a realistic take, check https://mitchellh.com/writing/non-trivial-vibing

this matches my own use of AI. as a widely capable assistant, that is fallible (spectacularly at times), but has a wide range of abilities.

[-]

Does it write the majority of my code? no. Does it help in some narrow contexts? yes.

In my view it needs to be something that is common enough that it has good training data, easy enough that it is readable, and low stakes enough that it isn't going to be critical if it goes wrong. I also need to be able to read and understand it well enough to make sure it is doing the right things.

Where it quickly goes off the rails is when you need something very specific with a lot of complex logic and with the latest tools/techniques/libraries. If the code needs to be split across multiple source files and has a lot of say API calls or something then I think using AI you are going to have a bad time.

Want a python script that does something done a ton before? It can do that easily. Want something that is very bespoke and business/use case specific with tight performance specifications or security requirements? you are going to end up with complete unusable garbage.

It is more solution search engine and solution space explorer than it generating the code for me in most cases.

[-]

GaboureySidibe@reddit

Lots of 'journalism' now is not about investigating anything but getting paid for ads and getting paid for articles by PR firms.

[-]

justUseAnSvm@reddit

Yes - Tech journalism is mostly shilling for companies.

The people who know the most about tech, aren't the ones selling it to you, it's not the execs, journos, or academics, but the people working with the technology day in and day out. We have a huge information asymmetry compared to the rest of the market, and even inside our companies.

-TRlNlTY-@reddit

It is so hard to find good tech journalism these days. Anything AI related in media is absolutely irrelevant.

[-]

TimMensch@reddit

I've had inside knowledge of the details behind news stories multiple times, and every single time the news story has gotten something completely wrong in a way that just so happened to make it more entertaining, compelling, and/or accessible.

It's to the point where I've given up on believing most journalism.

[-]