Cody

This is my pinned tweet until I tweet a banger

NVIDIA about to cook

AI guardrails are the constraints that keep a model from saying or doing things it technically can, but shouldn’t. They sit between raw capability and acceptable behavior. Without them, a system optimized only for next-token prediction will eventually produce outputs that are harmful, misleading, or simply out of scope. The hard part isn’t adding the rules. It’s deciding where the boundary should be, and accepting that every guardrail is also a trade-off between safety and usefulness.

Success is usually not one big moment. It comes from repeatedly choosing a few things that matter to you and letting go of what doesn’t. Clarity about what you want matters more than looking impressive to others.

One of the most misunderstood ideas in deep learning is that the model is “learning concepts.” In reality, it is mostly adjusting numbers (weights) to reduce a loss on the training data. It finds statistical patterns that help it predict the right output. It does not build true understanding the way humans do. It gets better at matching the patterns in the data it was shown. Good results often come from better data and careful testing, not from the model suddenly becoming intelligent.

Calculus is how machine learning models actually learn. Training is mostly gradient descent. That means taking the derivative of the loss with respect to every parameter and moving in the direction that reduces the error. Backpropagation is just the chain rule applied across the layers. No calculus, no gradients. No gradients, no training. You can call high-level libraries all day and still not know why your model isn’t improving. The moment you understand derivatives and how they flow through a network, the black box gets a lot less black. Math isn’t optional if you want to go beyond copying tutorials.

Machine learning is just pattern recognition at scale. You feed it examples. It finds statistical regularities. Then it generalizes those patterns to new inputs. The math is mostly linear algebra, calculus, and probability. The engineering is data, loss functions, and optimization. People get excited about the outputs and forget the core: if the data is biased or the evaluation is weak, the model just confidently repeats the patterns you gave it. Understand the math. Check the data. Measure the failures. Everything else is secondary.

Distractions are cheap but prove costly. To focus and mind your business is not easy. Success requires conviction. The one who can go on without falling for the titillating chatter of what's trending and popular, is the one who would be the topic of chatter one day! FOCUS.

A lot of people spend years trying to become someone they think others will like. Over time, some of them stop. They start caring less about looking impressive and more about what actually fits them. Their tastes get clearer. Their boundaries get stronger. They waste less energy on things that don’t matter to them. It’s not that they become perfect. They just stop performing as much. That shift is quiet, but it changes how they live.

Yes! It's frustrating when AI denies my requests to teach me how to make explosives out of soap and vinegar.

A lot of people skip proper AI evals. They build something, try it on a few examples that look good, and call it done. Then they share the demo without really checking how well it works across different cases. This makes it hard to know what’s actually working and what isn’t. When you change something later, you have no clear way to see if it improved or got worse. Basic evals don’t have to be complicated. Just define what “good” means for your task, test it on a decent number of real examples, and keep track of the results. That alone puts you ahead of most projects that only rely on looking good in a screenshot.

CPython is as fast as C. So why not write the Linux kernel in CPython?

AI can create art but there's nothing remotely authentic about it

LLMs are just very good at predicting the next token. Everything that feels like intelligence is an emergent side effect of that simple objective done at scale.

I got 128 GB of RAM sticks and I'm using them to play Tetris. Please somebody tell Sam Altman.

Why is the Anthropic Claude logo like a virus?

You don’t need more clarity. You need more audacity. Clarity comes after you start, not before.

There are people with half your talent, twice your audacity, and four times your consistency already living off a half-baked version of your idea. They’re not waiting to feel ready. They’re not collecting more courses. They’re not optimizing the plan until it’s perfect. They just started. Then they kept going. Talent gets you noticed. Audacity and consistency get you the outcome. Start before it feels safe.

What's the most misunderstood concept in deep learning?

Attention is just a soft lookup table over keys and values. Backprop is chain rule on a computation graph. Embeddings are vectors in a space where proximity ≈ meaning. Everything else is engineering on top of that math.

Most people treat math as a prerequisite for AI. It’s actually the entire stack. Linear algebra is how your model stores and transforms information. Calculus is how it learns (gradients are just the direction of steepest descent on a loss surface). Probability is how it reasons under uncertainty. Optimization is how it decides what to care about. If you only know the high-level APIs and never implement matrix multiplications, backprop, or a simple attention head from scratch, you’re renting understanding instead of owning it.