Ansh Mehra

(Thread)

How to Think Like the Architect of this Universe This Universe has consistently created species that become better at consuming, processing and remembering information. It has achieved this through insane levels of documentation, hidden in the DNA and epigenetics of each and every one of us. If DNA is a book, epigenetics is the collection of notes in the side margins. It allows nature to document changes without messing up the underlying code. 2 billion years ago, we started life from basic cells like prokaryotes who used protein signalling to remember information. It took us 1.8 million years to see early mammals who developed proper nervous systems. In last 20 years we’ve made LLMs that can process and analyse information 100x faster than any species ever alive. There are many repeating patterns I see in how nature created us and how we are creating LLMs. Nature writes in DNA, we write in documents. Evolution uses a slow brute force search for best solutions, we use Gradient Descents with trillions of adjustments to find intelligent results. However, as Andrej says, we are building ghosts ,not animals. Universe made us through evolution while we are making LLMs through imitation. I wonder if we should even make animals at this point. Maybe just making ghosts is the safer approach for us. Maybe the limitations of LLMs and training results are a temporary but deliberate constraint set by evolution because it wants us to stay relevant for a longer time. My recommendation would be to run this prompt through any LLM of your choice to further deepen your understanding: "Please read my notes and help me understand how to apply this knowledge to my job and daily responsibilities. Help me strengthen my first principles and keep quizzing me until you are confident about my understanding. Keep giving me ideas on how to apply this knowledge for practical benefits"

Everyone is learning prompts, but very few are learning how to think. That’s the real gap. Agentic AI is not about writing better prompts, it’s about breaking down work into systems. If you can’t define a workflow, structure a process, and write clear SOPs, AI won’t help you much. Because AI doesn’t fix bad thinking, it just amplifies it.

But how exactly do humans run supervised training? How do humans act like a parent for this childlike AI? To reveal that, we will have to understand how Parameters and Weights work in LLM SFTs. I’ll explain them in the next post.

There's one problem though. If you let this loop run without any limits, the model starts fooling the system, similar to how students try to fool their teachers during exam time. The model figures out patterns that usually get high scores from the Reward Model, even if those patterns are not helpful (like writing long but wrong answers, hoping to impress your Physics Teacher for grace marks) However, scientists have figured out how to solve this issue with penalties as well.

Let's make AI simple for everyone. Remember how life was when you were 2 years old...

My recommendation would be to run this prompt through any LLM of your choice to further deepen your understanding: “Please read my notes and help me understand how to apply this knowledge to my job and daily responsibilities. Help me strengthen my first principles and keep quizzing me until you are confident about my understanding. Keep giving me ideas on how to apply this knowledge for practical benefits”

Code, math, non-English languages - all tokenize differently, which is why models can behave differently across languages. But after we make tokens, we watch a beast of a technology take control - a Large Language Model. Check out the next post to learn about LLMs in a simple way.

Do not build Agentic AI Workflows unless you understand AI Tokens. Not understanding tokens can cost you and your company a lot of money. (a thread)