Adam Mainz

Hard work day but someone has to do it

@adriann_liu There are 2 main modes. Eager and compiled

@MainzOnX i see you mention something about eager pytorch — is there a non-eager pytorch? would that do something along the lines of what the compile function does?

@moby763canary21 Thank you!

@MainzOnX banger article

@philipbankier 🫡🫡🫡

@MainzOnX Finally a clear breakdown, thanks! Looking forward to the series

@philipbankier Thanks man 🙏

Fantastic breakdown. If you learning about GPUs or TPUs then you need to follow @MainzOnX

@mr_r0b0t Going to be a fun series 😄

@MainzOnX 🏆🏆🏆🏆 very much looking forward to your series, this article is excellent. thanks for putting it together!

@jordanrule Awesome can’t wait!

@MainzOnX I'll publish a proof for you this afternoon, give me a bit in planning meetings all day.

@asmah2107 Well deserved !

@MainzOnX thanks for sharing ! really appreciate !

When you see memory as a key callout you know someone is on top of their game. Good solid starting list of concepts. Happy to see someone smart putting this out

@ishi__singhal It’s definitely a tough subject to get into

Finally truly understand kernels

Hello fresh TL, moots, and builders. It's a pleasure to congregate here with all you fine legends this Tuesday morn. Welcome back! I've gotten the opportunity to spend some time with Sol, and it has been an absolute joy to use. But did you know that @OpenAIDevs suggest prompting 5.6 Sol differently than other models in the past? The evolution of coding with AI has truly entered a new era. We started with autocomplete, tabbing to finish a line or two of code. The models got better, and we started delegating entire tasks. The models got better once again, and we're on to building specs, and then subagents to implement them end-to-end. But now, we're in the era of outcomes. Rather than obsessing over every detail of what the model does along its path, your attention is best served on describing the end result: outcomes, expectations, evidence, guardrails, and what it really means to be "done." It doesn't mean that you're completely letting go of the wheel, but it can lead to better outcomes if you simply allow the model the autonomy to explore on its own. To help you navigate this shift, I created a meta prompting skill specifically for GPT-5.6 based directly on OpenAI's 5.6 Sol Prompting guidelines. Links in comments.

@MainzOnX Really great article

@kayanivtweets No normal human should be able to do this

@MainzOnX This and that HF x Gemma experiment is genuinely crazy

No illegal. Straight to jail before it’s too late for the rest of us have nothing left to optimize

@naveenmarrii Yeah but this isn’t a new naming here either lol

Kernel is the most overloaded term. CNN kernel, GPU kernel, Linux kernel, kernel trick, Jupyter kernel, database kernel…

@Kevrsub Absolutely! More coming

@MainzOnX This was really well written. Thank you.

Going to be putting out some beginner / intermediate friendly ML systems, GPU and kernel articles this week and next. Have a list myself but if there is something you want to see feel free to put a comment below. If any get a good amount of people liking or wanting the same I add

@_arohan_ @marksaroufim Huh wasn’t expecting to find something I saw before you

I realized how much of a legend @marksaroufim is last week.

@JohnSnow361474 You can do this reasonably through Python DSLs ie triton. Torch usually handles all that for you

@MainzOnX I meant memory coalescing for an insane throughput boost, or is it just in c/c++?

@JohnSnow361474 What aligning? A+b indices? In PyTorch yes but in a custom kernel you have control of the indices yourself

@MainzOnX that's great, is aligning done automatically?

@jordanrule None but happy to hear yours!

@MainzOnX Do you have thoughts on the neural tangent kernel?

@yacinelearning Starting from the building blocks and will work my way lower level!

very neat beginner friendly intro to GPU kernel check it out!