Bookmarks

llm.c by hand ✍️ C meets Transformer

This combination is perhaps as low as we can get to explain how the Transformer works.

Special thanks to @Andrej Karpathy

--

Part 1: gpt2_forward

Karpathy's llm.c implements the transformer forward step as the following sequence.

(skip)layernorm_forward
1. matmul_forward
2. attention_forward
3. matmul_forward
(skip) residual_forward
(skip) layernorm_forward
4. matmul_forward
5. gelu_forward
6. matmul_forward
(skip)residual_forward

--

Part 2: matmul_forward

B: Batches (of Tokens)
T: Tokens
C: Input Channels
OC: Output Channels

--

C programming and matrix multiplication are two of the most important topics. But it is often difficult to get people excited by these topics.

I hope this exercise can help people see further into the LLM black box, and appreciate C programming and matrix multiplication by hand. 😀

---
100% original, made by hand ✍️
Join 56K+ readers of my newsletter:

90% of these consumer apps are full of shit, but it doesn’t take away from the fact that this dog barking translator app should be making $200k/mo already

I did one of these a few days ago, but these app’s are so clueless with their marketing so here we go again:

There are 68 million households in the US that own AT LEAST one dog…

Not even Asia or Europe

Literally just America.

Do you realise how many customers that is? The amount of people who would use an app like dog translator? Most humans love their dogs more than their family

This dog translator app needs to be posting TikTok slideshows, but with an incredibly strong (and slightly unethical) twist. Something that touches the viewers emotion. Something slightly sad

First slide:

“I’ve had Jeffery for almost 12 years, and today I had to say goodbye…”

Second slide:

“We had so many memories together, but by far the best ones were when we had conversations through the silly little dog translator app during our evening walks”

Third slide:

Insert screenshots of dog translations

This strategy works well because it hits the emotions of those who own dogs themselves. They dread the day their “furry friend” passes away. It sounds slightly horrible but super effective

Subconsciously, viewers will want to check out dog translator. They’d love to experience the same feeling they watched in the videos with their own animals, even if the app hardly worked

You can take these winning formats and turn them into stories on Reddit. There are so many pet Reddit groups where emotional “story times” like this with CTA to a specific app go crazy

If you’re an app owner, two more points I want to make:

1) If you’re going to build affiliate teams that drive traffic to your brand, don’t listen to the people that tell you it’s a numbers game. It’s not. You need 15-20 affiliates max that can ACTUALLY deliver

2) Never pay any marketing team based on the views they generate. This is not how you stay profitable. You’re asking to burn cash. Pay per sale or sign up only

If you’re a brand and you want me to connect you with proven affiliates, you can hit me up. You will be charged a fee. It’s not free. I’m not a charity

If you’re broke, I recommend creating an in-house education program and training fresh affiliates yourself

Run the numbers

HOW TO BUILD A BIG CONSUMER B2C MOBILE APP IN 2026

1. start with a single recurring behavior people already document, like meals, sleep, workouts, studying, dating, or routines.

2. anchor the app to one question users already ask themselves daily, like “am I doing this right?”

3. narrow it to one audience, like college students tracking meals, busy parents tracking sleep, or single people tracking dates.

4. design the product so the core value appears visually in under five seconds

5. build the demo before the full product and let the demo define the feature set

6. keep videos between 20–40 seconds so curiosity builds without dragging

7. default to faceless formats like screen recordings, slideshows, or b-roll with captions (easier to do, can do founder led if that's your thing too)

8. create multiple hooks around the same demo instead of multiple demos.

9. use comparison formats like before/after, expectation vs reality, or me vs me

10. write hooks the way people text friends: short, casual, and specific

11. treat pauses, rewatches, and saves as the strongest signals of interest (this is v important)

12. read comments as public product research that reveals confusion, desire, and identity

13. paste comments into Claude Code and cluster them into concrete product changes

14. use AI inside the app to interpret inputs and surface one clear answer.

15. add a short onboarding quiz so users feel the output is made for them

16. deliver a result worth screenshotting within the first session. add CTAs to share. track the % of people who share and iterate to increase this.

17. place the paywall immediately after the first moment of clarity

18. ship small visible improvements weekly so users feel momentum

19. iterate in public so content doubles as changelog and proof

20. measure virality through shares and installs per view, not follower count

21. design outputs users want to send to friends without explanation

22. own multiple posting accounts early to test hooks in parallel (and to own a network of accounts kinda like your modern day media network)

23. give each account one format, one hook style, and one audience segment

24. scale formats that produce consistent installs instead of chasing one-off spikes

25. turn early power users into creators by resharing their posts

26. use slideshows as mini case studies that encourage rewatches and saves

27. build lightweight community loops like streaks, challenges, or progress updates. could be premium features too but start free.

28. keep the product NARROW so clarity stays high and competition stays thin

29. convert organic installs into profit early so growth funds itself

30. reinvest cash flow into more creators, more accounts, and faster iteration. Understand LTV/CAC extremely well, certain creators will make sense/others not so much. Use AI for outeach.

31. Build a portfolio of small apps once the loop works so dividends compound quietly. find more app ideas at @ideabrowser to get creative juices flowing

32. congrats, you’ve learned the most valuable skill in 2026...shipping small apps fast, reading the internet’s signals, and compounding what works.

33. you set yourself up nicely for 2027 and beyond