Are Large Language Models Really AI?

The next stage of this process is AI robot teachers that they will send out to the 3rd world, to teach these kids who they are training to come work in western nations.



It's all coming together!

Absolutely, it's degusting and so obviously a personnel cost savings measure. For the majority of companies the single largest expense is peoples compensation and benefits. Reducing that red line by 5~20% is a non-trivial amount and companies will do anything to justify it.

I subscribed for ChatGPT earlier this year, yeah it has been working moderately well for me so far.

It can generate images about right. It generates some useful scripts, it answer some questions moderately well.

But I have been hearing, and now I just watched a video declaring: Claude puts OpenAI in shame.

Aside having a better product overall, Antrophic seems a more sustainable company than OpenAI.

Right now Im very inclined to switch to Claude.

We will see

I've been using Claude Code for some very heavily guided tasks and I've found it more effective than the others. It still needs a lot more guidance and handholding (at least for my use case) than what they would have you think, but so far it's been more effective than Gemini, Copilot, and ChatGPT esp at handling large or complex repositories.

Claude seems the better of the bunch for coding, it still needs a ton of modification and implementing it directly into prod is just asking for long term issues. Get the modules and functions you need out of it, provide descriptive names / labels to everything, comment the code appropriately then you can build a production application out of that.

This case was mainly about analyzing some code and creating a file for testing some stuff. I didn't feel like reading the whole spec and figuring out how to hit some code coverage areas if I could automate it. I was able to get it 90% of the way there and fill in the blanks without having to craft it all by hand. I knew the code really well though and could say "don't do this" and "structure it this way based on this" rather than just tell it to make the file. It did a good job and cut out a lot of tedious, annoying debugging, but also picked up on some stuff I might have missed if I had done it by hand. I still had to fix the end result but it saved me enough time to be worth it.

It's also a very well documented format and the code is pretty linear.

When writing code, my approach is generally to handle coding with it at a function level, small enough chunks that I can read and interpret or modify myself, then integrate them into a larger arch I'm working on. Anything else results in more work IMO.

This has been my experience as well. It's really good at providing a function or method that does a specific thing for you, which you can then incorporate into a larger project.

This is why I say AI isn't going to "replace" developers / IT people. Photoshop didn't "replace" graphics artists. Instead both tools augmented those professionals allowing them to reach higher levels of individual performance. Human judgement is something that can't be replaced by an algorithm, every attempt has failed miserably.

You two are definitely in a minority in that way of thinking. And it requires being able to say "I just need a function to do X". Someone without the skill and knowledge to know what they want in the first place can't even ask that question.

Vast majority of people's use of AI is "write all my code for me so I don't have to".

Then get someone else to fix it for you when it inevitably breaks because wtf wants to read AI code.

If AI was a person it would be in the upper 10% of "easy to read code".

(doesn't mean its making great decisions about the content or w/e else) but AI i've worked with does nicely format and even comment the code.

(might even be in the upper 1%, its probably better than I am)

Nice formatting and comments don't make readable code. You can format a turd to look like the Mona Lisa but it's still just a damn turd.

Code that makes sense is easy to read. I've yet to see any AI generated code that I would say qualifies.

No we're not, we're just the Engineers not the Executives, and even then what most Executives think and say are two radically different things.

Public companies lean heavily on share value as an executive performance measure. Right now lots of people are buying into the AI Hype Bubble, and the data shows that any product or paragraph with the letters "AI" in them causes share prices to go up.

Just to illustrate this effect, I present the Oral-B Genius X Rechargeable Electric Toothbrush With AI.



I am not exaggerating, this literally made the meme. No rational person can thing "AI" has anything to do with electric toothbrush's, yet a company found a way to stick those two letters on a product to raise share price.

Industry insiders know most "AI" stories and paragraphs are just hype and nonsense trying to sell a product before the bubble bursts. Everyone is locking down and getting ready to weather the coming storm, that's why we see so many layoffs occurring. Just remember all the crypto hype before the crash, followed by the recovery, followed by more measured hype.

Alright, you must work with better engineers then.

99% of the engineers I work with try to get AI to do 100% of their work so they don't have to.

There is a lot of people earning big money with AI, and that is a fact.

People criticizes AI slop, but AI slop sells like cocaine. Just see how many views that UNLISTED video from Nickelny got in few minutes.

I think AI is like that WS that was buffed by SE, and suddenly the community perceive that job as the OP of the season. And they jump in the bandwagon.

Life is like that. Bandwagons exist for a reason. Smart people know when to hop in and when to jump out.

The world belongs to people who can adapt to many circunstances.

Those can profit be when the bubble inflates, or when the bubble burst.

And I respect that.

If AI ends up being all they want it to be, the absolute chaos that huge numbers of people being unemployable will be a serious problem.

The wild talk of peace time being over, conscription changes everywhere and politicians almost begging for wars in all western nations is fairly worrying for the conspiracy minded.

Anyone who knows the mechanics of what it really is doesn't believe 90% of the hype that is spewed. It's a data processing algorithm, that's all.

Now it's a very good data processing algorithm, capable of processing ridiculously large sets of data and pulling out patterns that were previously very hard to spot. This has important implications in any field that involves data analysis, especially things like materials science and chemistry. Instead of having to check hundreds of different combinations of a chemical hoping to find the one that works the way you want it to, AI can chunk through the data and inform you which combinations are most likely to provide the result you are looking for. The larger and more curated the data you feed into it, the more useful the output is.

Take the coding discussion from above, what it's really doing is using those key words (called tokens) and chunking through it's data set, known as the "model" to find the code combinations that most likely correlate with those key words. Some key word combinations will have very strong correlation, so you get a good answer, while other key word correlations will have very loose correlation, which produces a guess, what is called an "AI hallucination".

AI Bro's don't understand the underlying mechanics and to them it seems "magical", they end up anthropomorphizing it and project their own human thoughts and beliefs onto an algorithm.

This is why all the AI companies are hovering up as much data as possible, all while bunkering down getting ready to weather the upcoming storm. Whomever survives will have a very large competitive edge over the rest and become the next Alphabet, Amazon or Microsoft. There are very useful and important applications to this data analysis capability, and like all things it's going to take one or two decades before it really comes to fruition.

I didn't realize Halvar was still blogging, but I ran across his blog looking at some BinDiff stuff and realized he was still active, which is awesome. At any rate, he posted this, which I thought was interesting in the context of this thread: https://addxorrol.blogspot.com/2025/07/a-non-anthropomorphized-view-of-llms.html

This has been a shitshow of a week for the security industry. The Anthropic stuff has been front and center, with few seeing the *** by omission going on. I am lucky our execs see through this and listen to the engineers who explain it, because a lot don't and there's a lot of people taking what they say at face value. It's coming from every direction and the timing of it coming after RSA is no coincidence (RSA is a CISO/vendor schmoozefest where a bunch of non-engineers run around trying to sell ***to managers who don't know better, it's ground zero for most *** in this space and has been for decades). In the end, what they believe works is what matters, not how it actually works, and that's what scares me the most.

Its like "smart" was a few years back. There was a why is this thing smart blog or several that got picked up by news aggregators.

This has been done for over a decade now without AI. IDK if AI can do it faster but I do know it will burn LOTS more electricity doing it.

I expect to see AI sewing machines any day now. It will be about as useful for that as it is for toothbrushes.

That's what salesmen do.

I had a job where I had to unsell people on what they thought they wanted and resell them on what they needed.

Yeah I had to break it down to our people that it was a giant publicity stunt to drive sales on their code review products.

Read that blog post and it's really really good at breaking it down. It looks magical because it's been trained on human data, then parrots the data with the highest correlation to what we asked it to parrot.

"AI" is faster, much faster, orders of magnitude faster. Like I said it's a data processing algorithm.

Data science has had that method for decades now, but attempting to create usable code for it resulted in needing insane levels of computing power and time. The "hard" part of AI is not the data search / lookup function but the initial parsing and indexing of the model. ChatGPT 3.0, one of the simpler models, has 178 billion parameters inside it. To build this required crunching around 500 billion unique data elements called tokens.

Trying to do that with regular compute would of taken a ridiculous amount of time, and so nobody really bothered to do it at scale. That is until a technique was discovered to use nVidia's CUDA language on a highly parallel coprocessor that we use to render 3D graphics. A nVidia 3080 Ti has over 10,000 "cores" and over 900GB/s memory bandwidth. Those cores aren't very good at general purpose logic (if / then, compare / etc..) but are extremely good at doing math on 16 and 32 bit numbers. Using CUDA (later OpenCL) and suddenly the data processing for "learning" became possibly in a reasonable amount of time.

And thus the race kicked off to build / train the best model by crunching even more data. Whomever can build / train the better model can release the better product, and the more GPU compute / memory / storage you have access to the faster you can build that model and the faster you can get out in front of your competitors.


To highlight the difference, before AI they used computers to model chemical compounds. The import part here is they had to model them individually and hope the output was good. Lots of power and you can model a bunch at once, but when complex compounds have hundreds of thousands if not millions of permutations, finding the "right" permutation comes down to luck. With AI, you spend a ton of time / energy building the model by feeding it curated data of every reaction and property of every component of that compound you are working on. Afterwards you put in the tokens for the output properties you want and the AI will then use correlation to determine the most likely permutations. Then you can focus your time on modeling the most likely and reduce research time by efficiency.

The fun story of this week.

Now this is what I call AI slop!

https://www.theexploit.co/articles/cybersecuritys-hottest-new-go-to-market-strategy-refusing-to-launch

So, I spend the last hour trying to catch up with SDD.

SDD is a very interesting and promising methodology that step up over raw vibe coding.

BUT

The frameworks derived from it feels like legit AI slop.

I watched 2 people using AIOX, that is the brazillian implementation of SDD. One of those 2 was someone without any development background. The other one knew something.

The non-dev user using AIOX was a pure disaster. SDD-based frameworks used by non-dev users isnt objectively better than plain vibe coding. That person tried a pure prompt using AIOX, and the end product was poor. The she brought out of thin air a spec.md, the most important artifact of SDD, tossed on the CLI and redid the work. Personally I thought it was marginally better than vibe coding attempt, and it had a major UX issue: it was designed to make a engine sound every mouse scrolling, so the front end generated by AIOX was basically a failure: the site was noisy as hell, with those frigging engines screaming while you navigate in the page... But it was good for some laughs.

The second person didnt do much, just installed the CLI and showed one of the agents of AIOX.

I suppose what those implementations of SDD brings to the table is just a couple of skill.md ready to use. But this also is futile: they just made overcomplicated to develop, because you now need to manager 7~9 different agents, each with a shitload of commands to customize.

The purpose of AIOX, and other SDD-implementations, is to emulate a real dev team using agents. But I think this is no good.

A better use of SDD would be using a generalist dev team, that know the basics of all development process, and use SDD to implement the specificities.

It makes no sense trying to use agents just because. AIOX had a scrum master agent! Why the hell would I want a scrum master agent to develop with AI???

Just dropping this here, because the first trick is really golden.

Thanks to the guy for letting me know before I even started to get my paws dirty.



EDIT from future Panta: after watching the full video, now I wondered if I should go to Claude.

This consumption model is quite aggressive. I dont think I can optimally deal with such a constraint after extensivelly using Copilot and company paid license.

I feel like I can only put our my A-game if I can freely use the AI to refactor, rearrange, fix and redo things. Now if I need to have in mind to be economic with my prompts, I dont think I can perform reasonably.

I will take an in depth look at Copilot model before having my decision. AFAIK, Copilot doesnt use tokens for pricing. It is close product where there is a premium "request", that feels like a non-token consumption model, so I can work with large amount of context and artifacts to improve the output.

If my assumptions are correct, either I will stick with Copilot, or maybe I will use it in tandem, using Claude for chatbot instead of chatGPT + architecture tasks, and leave all the heavy lifting to Copilot.

The biggest worry of AI I see is that it's going to make vast numbers of people completely useless to society (most white collar workers for example), and by society I mean the elites that rule over us and pay for everything.

You are kind of seeing human life having less and less value year on year, the MAID stuff and other similar things show that. That would have been completely unacceptable 10 years ago, but they are going all in on it and don't give a crap. They are even "logging people off" for depression now in some places.

The elites also fear millions of specifically more intelligent (white collar!) people being out of work, and potentially them rioting and causing problems due to them no longer having a purpose. That may lead to them wanting to find ways to thin the heard out of those more intelligent people (low IQ people are less of a problem, they are far easier to control due to little organizational ability).

There are 60-300 million people (depending on nation) and not even 1% of that number in any country are police and military to subdue them if they start going off on one enmasse, they not only do not want but could never hope to handle it if people started kicking off in large numbers. You can absolutely guarantee they are talking about this stuff in the government think tanks.

That's when the leaders start wanting to start fabricated wars with forced conscription of men (the only group of their own nations population they are really afraid of), so all the problematic men get thrown into the meat grinder and are no longer a threat to them.

Surely that won't happen though cause all our elites have been proven to be good people who would love to look after us in the AI age, and our politicians care about us deeply and are not corrupt at all.

Elon said everything will be ok.

We don't need AI for that. Vast numbers of people are already completely useless to society.

I actually wrote this 2h ago, but I gave up.

Radys post is too much of a red flag for me to join in

This simply isn't true. It's the equivalent of saying photoshop replaced graphics artists or google replaced secretaries. AI is just a data processing algorithm, a tool. For some people it can enable them to be more productive, for others it does nothing. The only places it "replaces" people is when someone was brute forcing a data problem by throwing masses of humans at it, and in that situation they were not throwing expensive US college graduates but rather cheap Indian college graduates. Right now tech corporations are using "AI" as an excuse to further downsize staff and cut expensive operational costs in preparation for the upcoming bubble popping.


The AI evangelicals are trying to postpone that collapse by further hyping it with ever wilder claims. The more they can hype it the more they can attract investor capital, while paying themselves from that capital.

It seems to me like AI is the new RTO. They are blaming layoffs on all the roles AI is going to replace, but the reality is a lot more complicated.

There was a huge hiring boom in 2020-2021, we interviewed someone downstream of this and his salary expectation was obscene because everything was inflated, like 2-3x what any of us made at that point. A lot of companies have R&D projects or things that haven't worked out, when things go TU in the world, they are the first to go. There's also just a lot of fat to be trimmed in big companies, on average any team of 5 people has 2 people who don't do anything or produce any meaningful work.

Companies also know they can pull plugs, see what breaks, then rehire as needed. This happens a lot at large companies and they often go in waves with hiring FTEs, firing FTEs, hiring contractors, firing contractors. FTEs cost a lot because they have more rights, insurance, PTO, etc, contractors don't and they can get rid of people, bring in contractors to replace them, and eliminate overhead in accounting and HR knowing they'll work themselves to the bone to get renewed. There has also been a huge growth in offshoring work.

some roles may be optimized by AI to the point they may cut junior staff or those who are on the fence. Others may see more impact, like some marketing fields, but these are far more niche cases. That marketing nonsense from Anthropic everyone kept repeating is just that, marketing nonsense. If any other industry in the world put out that they were going to do all of that, no one would believe them, but people take it all at face value without applying any analysis for whatever reason. Debate this with the average person and the only evidence they can provide of this white collar apocalypse is data from AI companies making claims.

The bigger concern is if you work for a company who buys into the hype and doesn't understand what it takes to do the work with AI. They'll end up putting layoffs in place because they don't know better (see pull plugs and see what breaks), but I don't think these are the folks doing thousand+ person layoffs, rather small companies. I'm far more concerned about my job because someone thinks the work can be replaced by AI rather than it actually being doable

Downsizing by 10~20,000 people looks bad on a corporate press release, which publicly traded companies are required to do. It causes stock prices to slide as investors see that as an indication of expected hard times. By "AI washing" the layoffs they can put a positive spin on it and make it look like a productivity increase which would maintain the stock price or even cause it to go up.



The government stimulus checks had people go out and spend money, which creates demand signals. Those demand signals indicated market growth and companies responded by getting into a bidding war to steal as much talent as possible to meet that demand. That demand has proven to be short lived and now those companies are carrying excess talent that they paid premium prices on and now need to dump it.