วีดีโอ

What are some alternate IT jobs that don’t require learning AI?

I am learning AI and ML, but I like to learn the basics, am I not an engineer, am I a researcher??

How about someone who wants to write a phd paper ?

I took part in a uni robotics project, and for a while I was the only one writing code that worked in reality, while more experienced coders wrote untested drafts based on their experience(it was an autonomous robot project). They also really resented me as they were supposed to be more “qualified”. Anyway, reason why I’m saying this is because after the experience I went onto learning mathematical python a bit more deeply, but you reminded me that I did learn a lot simply from doing things practically, and also reminded me that I don’t need to “wait” to finish all the abstract stuff before returning to the practical.

Sigmoid is not used much these days. LLMs often use the gaussian - GeLU and for classification task you often use ReLU.. Sigmoid is slow. And ReLU and GeLU are easy to explain! 😊 If one wants to know how to use, one needs to have basic understanding of the activation functions ans when to use what based on their characteristics. Only needs an explanation of what non-linearities are and how they are mapped (embedded) into features and eventually single floating point numbers. Its all soooo easy. But its made sooo complicated by an industry that feeds itself high salaries just because they speak buzzword bingo alö day 😅

Okay, sometimes we do go deep into the algo :) But most time you can use stock library algos.. just what comes with pytorch

Without a programming background you can let ChatGPT generate those lines of code 😊 And yes, you are right; CS is overcomplex. It is possible to do all that with simple code. It's just people intentionally make it complicated because they want it fancy... And right -- Git you don't need too, if you are a lone wolf coding. There is local file system history in IDEs and the OS.. like Time Machine

The prompt usually is not chosen by the LLM. The prompt is tokenized, positional encoded, then passed to the normalization, through the attention mechanism, where its semantic context helps refining the prediction of the next token with semantic focus, then again normalized after it has been masked and softmax -- finally passed into MLP (using GeLU for captuning non-limearities) and eventually results in a logits list mapping to the tokenizer vocabular, so every token gets a log prob assigned. Softmax converts to %, topK and temperature acts as sampling and selection. Finally the next token is predicted and that continues.. all tokens including the last predicted on get refeeded (auto regressive automata) until a special end token is predicted. The prompt is always fed into the LLM directly. There might be a non-AI system in fromt that passes the user prompt plus context (loaded via RAG workflow, vector db) in a prompt template -- but eventually the LLM processes a defined prompt, not one selected by an AI usually.

Because they have to find new use.