Random Thoughts.....What Are You Thinking?

That is racial segregation.

if u mix ur orange juice and milk together u do u but i take mine separately

What the heck is "vibe-coding". I thought it was just writing code while being locked in and productive, but I keep seeing it used so many ways.

Vibe coding is when you ask a chatbt to create a script or program only giving the end results, then you do adjusts till you get satisfied with results.

A premise is that you really dont care what is code is in the end, so geenrally, if left alone, the vibe coded program is just something full of amends, that is tailor made for your use cases, and generally cant be evolved to another use case due to how amended it gets.

it's AI image generation but for code.

it's an anthropic marketing campaign

I finished reading the manga Boys Abyss.

I was expecting to be a hard to read mamga.

But it turned to be one of the most unfulfilling promises I ever read.

The ending was so genuinely bad, that Im pissed to have read it.

It is a well written story, no doubt. The author is pretty much a genius on putting hooks in every chapter, that makes you read dozens of chapters at once.

But that is it. His story is just trash. A good format with shitty content

Don't trash his vibe authoring


(c) TM patent pending

You mean it doesn't work?

it works, but you learn next to nothing and if it adds wrong code it can be hard to figure out why it doesn't work when you don't understand the difference between what it did right and what it did wrong.

It just doesn't work. Not if your goal is to do a good job.

If your goal is produce a lot of slop very quickly to pretend you've done work, yes, it works great for that.

It does ok with syntax if you don't know programming language at all. The logic, not so much.

Maybe something exists I'm unaware of, but I can only imagine the inconsistency between multiple people working on the same project. What a nightmare.

Agentic workflows that use the highest models and very high token burn do work to an extent. They'll typically produce code that does what was asked, especially when you're looking at tasks with a ton of training data like basic websites and CRUD operations. They do it by combining prompts with deterministic pathways(skills) and may ask the equivalent of hundreds or thousands of browser prompts to produce code.

In some cases, they're a good bit faster and cheaper than a real programmer. In other cases, they're slower and more expensive. Sometimes they'll make security conscious choices the amateur programmer would've missed. Sometimes they'll make huge glaring errors that could cost the deployer quite a lot of money. There've been quite a few instances of AI revealing API keys then people using them for compute and racking up tens of thousand dollar tabs.

They also typically produce unmaintainable code with no true owner. Without a person who knows how the code works, the only way to fix it is to hope AI can fix the mess it made. Reading through the volume of code a large agentic workflow produces in a way that allows you to troubleshoot it often takes longer than it would've taken to write it correctly in the first place.

Personally, I don't think businesses believe it's 'better than devs'. But, it doesn't have to be. If it can do 80% as well and their security team (or Claude Mythos) can prevent any financially devastating outcomes, then it's a cost cutting measure. Sure, it results in enshittification for the user (see: Windows the last couple years), but companies have never had any qualms about cutting other portions of the budget at the expense of customer experience. AI code is yet another way to do that.

There's also a big difference in AI application methods. For every person using agentic workflows to make something that mostly works, there are 50 dipshits asking ChatGPT to write something for them in a browser window. Those always end up broken in 5 ways the author doesn't understand and will never be fully functional.

Anyway, I consider the term 'vibe coding' to be a marketing campaign because it was astroturfed everywhere by the usual suspects (businessinsider, social media bots, the wall of indian influencers). The goal was to get everyone talking about it by highlighting an accessible term that would spread quickly.

This is still yet to be seen. I know Anthropic is selling a different story, but the noise is very high on anything remotely complex with every model/harness I've used. I'm coming to the conclusion that with complex targets the only real way to deal with this is have some sort of automated pipeline from AI finding to dynamic testing and automated triage.

Give most models/harnesses a web application and it can reasonably identify basic vulnerability types traced to input sources, stuff like SQL injection and command injections. Anything more complex, like issues relating to deserialization, and it starts to produce mixed results. Anything related to application logic is a even less consistent.

Anything more complex than that, it does profoundly worse unless heavily guided with prompts from static analysis. If you don't tell it directly where you want it to look, it devolves into scanning for silly things like strcpy/sprintf and doesn't find anything complex. The noise ratio also goes up a lot with these, to the point I'm finding something like <10% accuracy on issues identified via manual review. The issue comes about in these cases of, at what point are you just doing a bad job of static analysis to review the code just to make a determination on TP/FP/Unreachable compared to just...doing a better job at static analysis. You are looking at the code either way. The issue is the security community as a whole doesn't do this very well and has spent decades trying to eliminate it as a practice, but I have yet to see where any AI model produces the same quality of results (same with every automated tool for that matter)

IMO people are taking the Mythos stuff too much at face value, not recognizing that the gap between identifying a defect in the code and proving it exploitable, not even writing a working exploit, is huge.

I'm definitely seeing Mythos as another big PR blitz. Probably shaping up for a giant IPO.

I'll take your word on the rest of it, since you're clearly more versed in cybersec/devsecops. I probably should've bolded the 'if' there; I wasn't meaning to imply that it is actually capable of that. Just that if execs believe it can be done, heavily leveraging AI code can be seen similarly to many other cost cutting measures that reduce product quality or customer experience (enshittification).

Ya sorry, wasn't jumping on you there, as much as commenting on this very thing. The issue is how the general public perceives it and, as a results, executives. If those making budget/hiring decisions are seeing it as this thing that solves all these problems, then we have a big issue because what I've seen so far are these models create more work for triage, work/time better invested elsewhere.

Ironically, it seems hiring in this space has gone up because of it. It seems (anecdotally anyway) that a lot of companies I follow or have peers at, they are increasing hiring at the moment out of fear of their exposure and increased activity associated with the whole Mythos hysteria.

Didn't we go through this "execs think it's cheaper and that's the only thing that matters" thing with offshore dev teams?

I've worked with dozens of offshore teams. The code they generate is generally about as good as the slop shat out by AIs.

Good developers are worth their salaries. Companies don't always need good developers though. This was true before AI, it is true now.

Dang, Im agreeing with Dordo AI post for the first time ever, and I feel humilliated for that.

Actually no.

I dont agree with Dordo.

Maybe he lives in a first world country, then he can have outsourced employees from poor countries, but that, just for the dollar conversion rate, consider the low or medium wage from those countries as something huge.

Then, naturally, first world countries can have access to good devs from around the world.

But for us, who already live in a poor country, and have to deal with people that is what is left after the "best" already get better paid jobs, saying "about as good as the slop shat out by AIs." is an understatement.

I review outsource dev team work, and I would pick the AI code over what they generate every time of the day.

That's the same thing. Low skilled devs will have an even harder time finding work. The lower the skill, the more easily they can be replaced by tools.

How do you get more skilled when your skills are put prompts into AI. You don't. You just don't get hired.

I just make titties and run reports for my team's metrics.

using my powers for the good of the world.

idk ive used chatgpt before and i haven't really been seduced to use it for everything?

the only ai i use seemingly a lot is google gemini and that's not really by choice, is it

I've found some AI useful in generating reports, sure.

Github has "write release notes automatically from merge commits since last release". Useful.

Jira has automated reports for various things. Somewhat useful, not a manager don't care.

Nick makes vaguely entertaining AI titties - somewhat useful if you're into the heavily plastic look (I'm not).

As a more advanced search tool, I have found AI useful. It is just collating stuff from the internet. And has citations for its sources.

my stance has always been big > small > fake
even the flattest chest is better than silicon

Thats just not true..

I can make all shapes and sizes.


https://github.com/fulletlab/comfyui-anima-style-nodes
I found this yesterday and have been playing with it - holy crap man... This is very overpowered man... cuts down my lora usage like crazy...

When I click on a thread it automatically scrolls to the last post. Is that something new? its kind of annoying but maybe it was there already

Yo Spider-Noir is actually amazing.

Steamdeck jumping $250 that steam machine is gonna be two grand lol

didn't it used to jump straight to the last post instead of scrolling down? it is a bit annoying.