All innovation and the gains from industrialization is due to abstraction. When I first got into programming Python, this was the standard take for people using Python. The concern was that Python abstracted away too much from "real" languages like C++ and that you weren't going to learn proper computer science if you didn't know C++. Many of my viewers were also students who would tell me how their professors taught in C++ because Python was too high level in their opinion. The thing is, when C++ came out, it too faced this exact standard take. It was accused of being too high level compared to C and assembly. C++ even faced the same claims that Python faces about how the abstraction would negatively impact performance and that it was too much bloat. You already know what I'm about to say about Assembly... but it too was accused of being too high level by the TRUE machine coders. We can look back at all of this and laugh, but it's important to recognize that for the entire course of history, not just for computer science, all innovation has been abstraction. That's been the key the entire time, but the next layer of abstraction has been accused of being not real and fear mongered by the local purists of the time...but eventually becomes the thing that the same purist types point to later as the "good ol times."

@Sentdex AI is much more than the abstraction you referred to in your response. Never in history have we had tools which ideate && build almost entirely. His point is that the tacit knowledge developed in the process of coming up with both the ideas and implementation is getting lost.

@Sentdex I promise, the comparison of Python’s abstraction vs. C++ was absolutely not the “exact same” as comparing C++ to C

@Sentdex yeah but the problem is that python has really shitty abstractions that leak a lot of the problems of C (for example, how objects are passed mutably but not numbers, which is inconsistent). If you're going to abstract, you should at least build something better.

@Sentdex do you have source(s) for assembly also being considered too high level? would love to read

@Sentdex i'm confused about the fear mongering here. it doesn't resonate with my lived experience. i don't see many noobs cranking out high quality software with ai all of a sudden

@Sentdex Avoid excessive abstraction; prioritize understanding underlying mechanisms.

@Sentdex I'll just leave this here...

@Sentdex The difference being that all those abstractions were deterministic. Ai isn't which is my greatest annoyance with it and the reason its not that usefull

@Sentdex LLMs are not abstraction the same way Python is. LLMs are probability black bix machines which is inconsistent in its output. LLM are generators and a productivity tool if used properly.

@Sentdex But the thing is they were right. If all you know about how a computer operates I through the abstraction level of python, or even C, then you don't really know how a computer works. That is ok for a programmer, but don't consider yourself a computer scientist.

@Sentdex agreed. abstraction doesn’t remove understanding, it shifts where humans spend their attention.

I don’t get people who who complain about LLMs like this. You’re still responsible for your education, your knowledge and eventually your career. I watched plenty of people cheat in college 15 years ago in my CS classes. They would put notes up thier sleeve, notes on thier palm and google the algorithms for projects. The industry weeded out all of them. Some couldn’t even get ANY tech job and became mangers in the service industry. At the end of the day it doesn’t matter to you. Those that are curious, hardworking and want figure things out will rise well above those who don’t. They always do.

@Sentdex C++ abstractions ARE too far removed from the computer; it’s basically broken. C is an excellent. Python is ridiculous, and extremely wasteful of the hardware (between 5k-10k X is waste). We all would have been much better off if AI was happening in C. JAI looks promising.

@Sentdex Haha very well said

I agree abstractions are good, based on the use case. I understand that Python was an abstraction, C++ was an abstraction and hell C was an abstraction. Python is extremely useful abstraction for ML research and analysis, and absolutely useless when it comes to systems programming. This is why in companies like Waymo, you have researchers who develop ML algorithms in Python like languages, and then you have compiler engineers who primarily write this in C++. In NVIDIA's NVFP4 competition, the top optimizers are constantly in CUDA and PTX, rather than abstractions like CuTile or Triton. These languages are essentially tools based on the use case. Unless I am actually planning to learn, I will never write Transformers from scratch with C++, I'd use PyTorch abstraction instead. But like Karpathy did in nanochat, use Rust for tokenization if I actually care about efficiency and performance. LLMs are extremely useful, depending on the use case. The concern is when these students who are passing on all their thinking tokens to LLMs. Programming and programming languages have always been tools for problem solving. Here someone would critically think about a problem, come up with a solution, experiment with the available tools, and iterate. If one is not able to critically think about a problem, they will not be able to critique any LLM. Then we are just left with LLMs critiquing themselves. If one were thinking about competition, this is great - there is high job security if all your competition is doing is using the latest and greatest LLM. AI security will be a blooming field, just like cybersecurity basically mirroring the delicate ecosystem but for the entire software stack. Apologies for the brain dump, I just thought I'd write something down when I read your post. Your from scratch series were really useful when I started ML.

@Sentdex C++ and Python serve different needs. For something where performance is important ( a lot of critical systems we use daily) you cannot use python and you are forced to understand things at a lower level.

@Sentdex Wait so learning programming langauge to work on a llm isnt the correct way

@Sentdex You missed the point.

@Sentdex But but but they all were at least deterministic...

@Sentdex When you're after the highest performance you need to understand the silicon you're using and optimize your code for that. Most developers could go their whole career and never need this level of performance

@Sentdex Every generation thinks the next level of abstraction is too much. Until it becomes the new baseline

@Sentdex Correct. ADDING a layer of abstraction! Now you have to zoom out and take a look at the reaches of the new system.

@Sentdex Isaac Newton summed it up when he said that he could see further by standing in the shoulders of giants. You can abstract away the problems that have already been solved

@Sentdex purists in the comments

@Sentdex Abstraction is like power steering for your brain, you stop wrestling the machine and start shipping ideas, but forget the gears and you panic when it squeaks.

@Sentdex abstraction is just humans admitting we’re lazy and then winning anyway

@Sentdex It is a higher abstraction but like the past,those who know (at least some of)the system from the ground up while ALSO using the latest higher level tools will succeed. If a programmer is upset how LLMs changed programming, they missed the assignment (I've been coding since 2001)

@Sentdex x.com/josecanciani/s…

@Sentdex C should have been the stopping point, you literally can solve any higher level problem while still being human readable. Every "abstraction" following C was a complete waste of time and invited imbeciles to ruin software as we know it