Are Large Language Models Really AI?

Everywhere will be India in 100 years.

The UK, EU, US, Canada are all doing/done trade deals that allow unlimited Indian workers to come in.

The level big corp will goto to import a serf class is crazy.

Before they were happy just building factories in india to undercut western workers, now they just import them all to work here instead.

Algorithms have been used to drive engagement for a few decades now. They are powerful enough to produce addiction.

Who needs AI? Or meaningful for that mater? Addiction is what counts.

Earlier I received a video recommendation about a guy criticizing AI gen in arts, but obviously the guy was an emocionado, and started to shoot his white knight bullets in the AI.

The entire video didnt bring a single new thing that wasnt mentioned by people criticizing AI. I feel like we reached the plateau of new arguments in each side. At this point, it is not a discussion anymore. It is just tribalism.

The only “new” thing I can bring when someone start this discussion near me is: “lets talk about this again in 100 years”.

If you art or game generated by AI is criticized mostly by prejudice, maybe in the future people will have another posture about AI, and your refused work can be accepted them.

It isnt your haters or your buddies the judges.

Time is.

Just wait, and we will see.

Earlier I watched a video about a guy saying how CS big names have been adhering to vibe coding.

His example were footages of Linus Torvald using AI, and a blog post from the creator of ruby on rails defending Ai usage.

Then he mentioned the supposed big one in terms of coding, the Claude Opus 4.5.

I checked some footage of people using it. Feels like it not just accept the prompt, but willingly ask for more information before spitting the code.

Also, it lays out all the steps to achieve the expected product.

The end result is still subpar, but having an AI writing a viable walkthrough for a development is a big thing, at least it has a value as a benchmark and/or help with cold start.

I still feel awkward using text to generate products. Maybe if we could draw stuffs like architecture, and it could go from there, I would feel more confortable in using AI

Due to layoffs I got some C++ code bases dumped in my lap (I would never claim to know C++) I've been using that model to support them, it probably gets 80-90% correct out of the gate, then I slam my head against the wall to get the rest of it working.

So yeah its helpful and do a lot of the tedious stuff, i wouldn't want anyone clueless using it because the moment it does something wrong they'll have no where to go.

It does a decent job refactoring code too.

Does a great job documenting code.

While the lyrics were written by a human, the voicing is AI.

It has 7.1M views so far.

AI is real

(but the music is trash, but people is consuming this trash)

It's lead to a new career field, Vibe Coding Cleanup Specialist.

https://www.linkedin.com/posts/deveshbhardwajj_ai-startups-engineering-activity-7373218403068788736-VXlT/

https://www.thirdrocktechkno.com/blog/vibe-coding-cleanup-specialist-everything-you-need-to-know/

(Original article behind a login prompt but reddit commenters posting relative parts)
https://www.reddit.com/r/technology/comments/1ne8w42/the_software_engineers_paid_to_fix_vibe_coded/

Service provider for Vice Coding Cleanup Services.

https://softteco.com/vibe-coding-cleanup-services


This is 100% legit now. Executive staff put "Use AI to enhance workflow productivity" as goals for senior management, who then had no choice but to push it. This creates a mountain of buggy code, so they then turned around and either converted senior developers into "vibe coding cleanup specialists" or hired outside firms to do it. The result is zero real productivity increases but you can write down imaginary ones on paper backed up by metrics.

Late reply but GenAI is no different from photoshop or other creator tool. When photoshop and it's myriad of filters because mainstream did graphics artists disappear and be replaced by normal users? No they didn't, normal users can use it to enhance photos but graphics designers can take it (or similar tools) and go to even greater heights.

Right now is the equivalent of that time period in the mid 2000's when desktop photo and video editing tools with all those filters became easy to use. Eventually the dust will settle.

That's just "development" these days.

Fml.

Already repeated "*** off with that ***" on being asked to review AI code a few times.

The other day I asked chatGPT: I want to project a environment monitoring app, what do I need.

Honestly, I liked the mini project it gave me: it had the basic components necessary, plus the software I could use for each responsability.

In the end, I feel like the problem with AI coding is the same problem with human coding: lack of requirements or better explanations.

We would complain that if we ask "write something app", and it drops a swiss cheese in terms of security. But if we ask a human "write something app", the same swiss cheese would be given, but taking longer.

Why? Because you cant fit an entire project in a single statement. Not machine nor human can guess every single requirement, every single architecture and come up with the best solution.

But if you take the AI step by step, asking small pieces of codes with proper instructions and requirements, most likely it will give something acceptable.

This is beating a dead horse, but I can see in the future people needing to learn more about system designs than just the newest junk garbage language that people come up just because they dont want to add ";" in the end of line.

And if we spend more time discussing system design than trying to figure out what lib to use, what pattern to use, then overall the software development grows in maturity.

This is because it copy pasted someone else's project off stackoverflow or github.

I'm dead serious, nearly all the "code" that the various LLM's were trained on came from stackoverflow, which is the exact same place we all end up after googleing "create python REST service to import jpg files". We would then look over the thread at the various suggestions, copy paste what we liked out of it, modify it and then it becomes "our" code. AI is just guessing at what it thinks you want from it's own inventory of already scanned code patterns.

This is why it absolutely needs refining by an experienced developer, those were all suggestions to an original post that looked kinda like what you asked for. The highest rated suggestion / accepted answer is likely to be a good solution but it's only a Prof of Concept or Prototype until you've modified it for the required use case.

Embrace it instead of fighting it. Then in a year have your job title changed and then you can get a very sizable pay raise. Seriously companies are now throwing large piles of cash at anyone who can prove they know how to cleanup the mess those LLM's vomit. Every executive pushed for "20~80% productivity increase" from AI, and those managers were forced into using it to generate piles of garbage that they do not have the time or skill to make safe.

No. I get paid enough to not care about a pay rise if it means working with trash.

Those models not only copy things but they also can process some magic of generating a coherent and accurate answer most of time (lately).

My company has an instance of AI model integrated with Sharepoint.

And it is impressive how it can scan all content produced in those pages and give good corporative answers (things you will never find on internet and even inside company, you will only learn after talking to a lot of people in those far departments.

In my eyes, those corporative chatGPTs are a game changer in terms of spreading corporate knowledge. You just write some random article, drop in some sharepoint, then the LLM will find it, understand it and provide the answer if you ask something related to it.

What impresses me is how loosely related a question can regarding the topic. It can infer the related article, even if it is only one, and answer.

No they copy things as a sequence, that is how the LLM learning process works. For stackoverflow it's just the entire code section for the highest rated responses with key words from the original post. It disassembles your prompt into tokens, then returns the values with the highest correlation to the tokens you requested. When doing a complex prompt it will append sections of code together into a clump that is syntax accurate. Stack overflow has already admitted it sold access to all it's posts to every coding LLM out there, it's the #1 place people for coding / scripting solutions to basic questions.

Importantly there is ZERO concept of context or circumstance.


This is an illusion. You are not special nor is your request special or particularly unique. The probability of someone else on this planet having asked the same or similar question as you is 100%. You make a request for something that is three components, it will find the most likely match for each of those components then put them together. It's no different then if you had done a google search for each one, then copied the most likely answer from stack exchange (or occasionally reddit or some dudes coding blog) and repeated for the other two and put the results together.

Because LLM's do this fast we think they are somehow magic, but it's just token chains inside a ridiculously large matrix.

This thumbnail is just... awesome...

It feels like there is a bias here. Which one you would say is the best out of those 3 frm the thumbnail alone?

I just think funny because it feels like chatGPT made a PS1 game and Claude made a Minecraft extension...

Just don't ask grok.

Grok recently was put into contracts to use with the US military… somehow.

One assumes that its porn adjacent image capabilities had nothing to do with this but that they will be appreciated by the troops.