Jorge Castañón, Ph.D.’s Post

🚀 Curious about what's under the hood of a GPT? @moebio did a clever visualization of the inner workings of an LLM. Check it out here: https://hubs.ly/Q02qp-tv0 🔍🧠

🚀 Curious about what's under the hood of a GPT? @moebio did a clever visualization of the inner workings of an LLM. Check it out here: https://hubs.la/Q02qq9Jc0 🔍🧠

This is a brilliant visualization of how a LLM AI works.

How fast can you produce a working game using #genAI? It took me less than a minute to produce minesweeper. Everything is on this page. Time taken = [write prompt] + [wait for GPT output] + [check output] + [copy & paste code] + [run code] = 28 + 11 + 15 + 3 + 2 = 59 seconds 🤯

LLM Visualization This is actually pretty amazing! It helps to visualize the core components of LLMs like nano-gpt and GPT-3. https://bbycroft.net/llm

GPT API users, it's so easy to switch to GPT-4o, and it gives some impressive benefits. A 16 second video talking about them! https://lnkd.in/e_m9JNsh

GPT-4 is no longer the best LLM in the world, Claude-3-opus has overthrown GPT-4-1106 Link to the leaderboard: https://lnkd.in/dCpXQXZp #lmsys

Karpathy’s Zero to Hero series ——————————————— Build GPT2 (124M) from zero to end The video start with empty file and end up with a GPT-2 (124M) model: - first it builds the GPT-2 network - then it optimizes it to train very fast - then it sets up the training run optimization and hyperparameters by referencing GPT-2 and GPT-3 papers - then it brings up model evaluation, and - then run actual training overnight - Finally it looks through the results. The "overnight" run even gets very close to the GPT-3 (124M) model. Video: https://lnkd.in/d4w4z-td GitHub repo: https://lnkd.in/dhu6H5Dd

LLMs do not "think". The example in the picture is a try on GPT-4o from yesterday. I used a prompt that is reported as a fail for GPT3. The new model can surf the web and got much more parameters and training. Still in its final judgements it shows that it inherently has no intuition on the nonsense and semantic meaning of the inquiry. Because of this, it takes a lot of twists and turns to make LLMs mimicking real humans. In this article I am discussing the advanced pre-work it needs to get a workable digital twin of your customers. https://lnkd.in/eWfhkNPN #digitaltwin #customerinsights #causalai

Yet another thing to think about around prompting within different LLMs, even within the same family. This study evaluated the effect different formats had (markdown, plain, YAML, and JSON) to the outputs. GPT-3.5 performed better with JSON, while GPT-4 preferred markdown. Study: https://lnkd.in/ghQ2NPYn #llm #genai