Konstantinos Faliagkas’ Post

How does ChatGPT work? It might feel like magic, but it's far from it! Here's an engineer-to-engineer explanation from Evan Morikawa who heads up the Applied group in OpenAI (that includes ChatGPT): When you ask ChatGPT a question, several steps happen: 1. Input. ChatGPT takes your text from the text input (simple, I know!) 2. Tokenization. It then chunks it into tokens. A token roughly maps to a couple of unicode characters. You can think of it as a word. (Note that tokenization doesn’t necessarily mean splitting text into words; tokens can be subsets of words as well.) 3. Create embeddings. An embedding is a multi-dimensional representation of a token. ChatGPT explicitly trains some of their models to explicitly allow the capture of semantic meanings and relationships between words or phrases, and embeddings help with this. The system then turns each token into a vector of numbers. These are called embeddings. 4. Multiply embeddings by model weights. These embeddings by hundreds of billions of model weights (yes, really!) This is where most of the "magic" is: ChatGPT's weight matrix gives it the ability to work as well as it does. There's a lot of pertaining and fine-tuning for ChatGPT to produce this massive matrix. Note that this operation is an expensive one, compute-wise, and it's where specialized hardware like GPUs make a massive performance difference versus using e.g. CPUs. 5. Sample a prediction. At the end of this multiplication, the vector of numbers represents the probability of the next most likely token. The next most likely token are the next few characters that spit out of ChatGPT. ChatGPT, in the end, is a probability machine that gets really, really good at always predicting the next best token/word, to the point that it feels like it understands what you ask. Once you know how it works, it's easier to understand its limitations (e.g. it won't be able to come up with things not in its training data - not without hallucination at least!) as well as how to get around these limitations (e.g. supplying additional context or training). It's also easier to understand why e.g. asking ChatGPT to summarize a text requires a LOT more compute than asking it a question (summarizing a text means a lot longer prompt, and the longer prompt means far more calculations; but asking a question "just" uses the existing weight matrix) For more details and illustrations, see this week's free The Pragmatic Engineer deepdive on how ChatGPT works and scaling challenges (and hardware limitations) the team overcame: https://lnkd.in/eDb8d_T5 (Thanks, Evan, for all the details!)

We're really excited about our next steps on the Wealthyhood journey and are now looking for exceptional investment professionals to join our team! Here are the details: 👉🏼 Investment Operations (Consultant) 📑 Part-time contract (with full-time potential) 📍 Athens, Greece 🇬🇷 ✔ Necessary requirement: Type-B certification by the Hellenic Capital Markets Committee (Πιστοποίηση Β, Επιτροπή Κεφαλαιαγοράς) If you believe you're a fit, just drop me a message. If you happen to know someone who meets the description, please share or simply comment for your network! 💙 JD on LinkedIn below. 👇🏼 #investing #investments #operations #fintech #startups #trading