Excellent read. This consolidated and catalyzed my my spurious thoughts around personal information management. The input is generally markdown/pdf but over time highly useless for a single person. Thete would be value if it is passed through such a system over time.
A current 3rd year college student here. I really want LLMs to help me in learning but the success rate is 0.
They often can not generate relatively trivial code When they do, they can not explain that code. For example, I was trying to learn socket programing in C. Claude generated the code, but when I stared asking about stuff, it regressed hard. Also, often the code is more complex than it needs to be. When learning a topic, I want that topic, not the most common relevant code with all the spagheti used on github.
For other subjects, like dbms, computer network, when asking about concepts, you better double check, because they still make stuff up. I asked ChatGPT to solve prev year question for dbms, and it gave a long, answer which looked good on surface. But when I actually read through because I need to understand what it is doing, there were glaring flaws. When I point them out, it makes other mistakes.
So, LLMs struggle to generate concise to the point code. They can not explain that code. They regularly make stuff up. This is after trying Claude, ChatGPT and Gemini with their paid versions in various capacities.
My bottom line is, I should NEVER use a LLM to learn. There is no fine line here. I have tried again and again because tech bros keep preaching about sparks of AGI, making startup with 0 coding skills. They are either fools or genius.
LLMs are useful strictly if you already know what you are doing. That's when your productivity gains are achieved.
Brace yourself, people who are going to come to tell you that it was all your fault are here!
I got bullied at a conference (I was in the audience) because when the speaker asked me, I said AI is useless for my job.
My suspicion is that these kind of people basically just write very simple things over and over and they have 0 knowledge of theory or how computers work. Also their code is probably garbage but it sort-of works for the most common cases and they think that's completely normal for code.
I'm starting to suspect that people generally have poor experiences with LLMs due to bad prompting skills. I would need to see your chats with it in order to know if you're telling the truth.
One with ChatGPT about dbms questions and one with claude about socket programming.
Looking back are some questions a little stupid ? Yes. But affcourse they are! I am coming with zero knowledge trying to learn how the socket programming is happening here ? Which functions are begin pulled from which header files, etc.
In the end I just followed along with a random youtube video. When you say, you can get LLM to do anything, I agree. Now that I know how socket programming is happening, for next question in assignment about writing code for crc with socket programming, I asked it to generate code for socket programming, made the necessary changes, asked it generate seperate function for crc, integrated it manually and voila, assignment done.
But this is the execution phase, when I have the domain knowledge. During learning when the user asks stupid questions and the LLM's answer keep getting stupider, then using them is not practical.
Also Im surprised you even got a usable answer from your first question asking for a socket program if all you asked was the bold part. I'm a human (pretty sure at least) and had no idea how to answer the first bold question.
I had already established from previous chat that upon asking for server.c file, llm's answer was working correctly. Rest of the sentence is just me asking it to use and not use certain header files which it uses by default when you ask it to generate server.c file.Thats because from docs of <sys/socket.h>, I thought it had all relevant bindings for the socket programming to work correctly.
I had no idea what even the question was. I had chatgpt (4o) explain it to me, and solve it. I now know what candidate keys are, and that the question asks for AB and BC. I'd share the link, but chatgpt doesn't support sharing logs with images.
So you did not convince me that LLMs are not working (on the contrary), but I did learn something today! thanks for that.
I can get an LLM to do almost anything I want. Sometimes I need to add a lot of context. Sometimes I need to completely rewrite the prompt after realizing I wasn't communicating clearly. I almost always have to ask it to explain it's reasoning. You can't treat an LLM like a computer. You have to treat it like a weird brain.
The problem with these answers is that they are right but misleading in a way.
Glass is not a pure element so that temperature is the "production temperature" but as an amorphous material it ""melts"" in the way a plastic material ""melts"" and can be worked at temperature as low as 5-700c.
I feel like without a specification the answer is wrong by omission.
What "melts" means when you are not working with a pure element is pretty messy.
This came out in a discussion for a project with a friend too obsessed with GPT (we needed that second temperature and i was "this can't be right....it's too high")
Yes. This is funny when I know what is happening and I can "guide" the LLM to the right answer. I feel that is the only correct way to use LLMs and it is very productive. However, for learning, I don't know how anyone can rely on them when we know this happens.
I mean. Likely, yes, but if you have to spend the time to prompt correctly, I'd rather just spend that time learning the material I actually want to learn
I've been programming for 20 years and mostly JS for the last 10 years. Right now, I'm learning Go. I wrote a simple CLI tool to get data from several servers. Asked GPT-4o to generate some code, which worked fine at first. Then I asked it to rewrite the code with channels to make it async and it contained at least one major bug.
I don't dismiss it as completely useless, because it pointed me in the correct direction a couple times, but you have to double-check everything. In a way, it might help me learn stuff, because I have to read its output critically. From my perspective, the success rate is a bit above 0, but it's nowhere close to "magical" at all.
From my perspective (college student), it is true that there is godly amount of C code in the world which would remain true for a long long time. In similar sense there are also quite a lot of new java developers who maintain the millions of lines of java code.
However a lot of new infrastructure is being developed in rust. Infact it can be argued that the very reason it should be in rust is because it is critical. I think there would be great value if a person can efficiently thread between both rust and c rather than competing.
Java is not an old “maintain the code” language as you seem to imply. Despite not being the new hotness for about 25 years it’s an extremely useful and productive language and I assure you there are tons and tons of new things being written in it every day. The language is still evolving and has been getting great improvements. Yeah it has some warts, but it’s been running and keeping backwards compatibility for 30 years despite evolving.
> Infact it can be argued that the very reason it should be in rust is because it is critical. I think there would be great value if a person can efficiently thread between both rust and c rather than competing.
The entire graphics driver for Apple Silicon Macs in Linux, with zero memory bugs in the code despite being deployed to a large number of machines. including a fully conformant OpenGL implementation and most/all of Vulkan (I’ve lost track of the current state).
Fair enough. I have personally experienced pain with phantoms.
But then what would you advice to do ? There are nightmare patterns in rust and c++. Looks to me as pick your poison just that one is new and other has been around for fair bit.
I am sorry. I really don't see your point. Typescript is in a completely different league of languages than rust/c/c++ with vastly different use cases.
Also, the world still mostly speaks C99 when iterative improvements are constantly being made up to C23.
The point is that Typescript was able to replace JavaScript by adding to it. Converting a codebase from js to ts was simple. Going from C++ to Rust is not.
I think C23 is going in the right direction, but it needs to go further.
You examples seems to be for a dynamic environment where things are changing. Rust is naturally at slight odds in such a scenario
Instead of quick and dirty, it allows to embed the logic in the type system such that the compiler can help you. Getting off the ground is much harder but staying there is easier.
Maybe the compiler (and the syntax) isn't the right place to put these guarantees.
Surely there is some middle ground in design space between rust and ada on one side (fully in the compiler) and c (with sel4-style checking - proof checking on a post compilation artifact) on the other.
Note that the c in sel4 and ada have stronger safety and correctness guarantees than rust
> Rust can be written as simply as you want it to be written. Or, you can go crazy with generics or metaprogramming.
In theory you can do that, in practice you can't. Becoming a good $language programmer necessarily involves reading other people's code, for example when diagnosing issues in misbehaving libraries. You have to be familiar with every common language feature that other developers use, otherwise you won't be able to do your job properly.
I'm curious about the c with sel4-style checking that was mentioned. Does sel4 have a style checker and/or proof checker that you can run on your c code? Or do they just have some style guide you have to read and memorize?
> Does sel4 have a style checker and/or proof checker that you can run on your c code?
The specification and proofs are open-source [0], but I suspect that they are tailored for seL4-related uses and probably aren't well-suited for "general-purpose" C code. I think using their proofs would also necessitate writing a specification for your own code which is probably going to be an ordeal in and of itself.
They do have a style guide [1] as well, but that's just a small part of the full verification process (described in [2]).
It is great that the ecosystem is growing, but there is still time for things to trickle down.
As college student I have invested time into rust because I see value in the language but there are absolutely zero entry level jobs or internships. To get my foot in, my only real options are C/C++.
Although in rust, function like macros make it super hard to trace code. I like them when I am writing the code and hate then when I have to read others macros.
The first piped query language I used was Nushell's implementation of wide-column tables. PRQL offers almost similar approach which I have loved dearly. It also maps to different SQL dialects. There is also proposal to work on type system: https://github.com/PRQL/prql/issues/381.
Google has now proposed a syntax inspired by these approaches. However, I am afraid how well it would be adopted. As someone new to SQL, nearly every DB seem to provide its own SQL dialect which becomes cumbersome very quickly.
Whereas PRQL feels something like Apache Arrow which can map to other dialects.
A recommendation from a person you know takes into account not just their knowledge of your preferences, but also how much and in what way they like/care about you, and conversely, your taking of the recommendation is colored by your rapport with the recommender. All that is something a recommender system has no access to.
Or, more bluntly: you aren't going to mate with a For You page, so it doesn't have the same evolutionary cheat code to your preferences as other people have.
