How LLMs actually work! Recently gave a 2-hour long talk on this. 42 slides, all yours 👇

Presentation: (left and right arrows on desktop, left and right swipe on mobile)

Attended it live, loved it! @paraschopra

glad you liked it

Just down with all slides, hands down never saw this type of intuitive explanation. Must read!!

solid resource — most LLM explainers either stay too surface-level or dive too deep too fast. covering it in 42 slides is a good format for building the intuition before getting into the details. saving this for when people ask me where to start.

Thank you for sharing

Any recoding of that 2 hour session available?

no, it was a closed session

No problem the slides are too good!

which ai tool did you use to make these slides , was it gamma + napkin along with claude or something else ?

just claude code + my feedback/guidance

as claude provide connectors with gamma and excelidraw , which is nice (at as of today using those connectors it does the tasks little better) but it feels like in few months claude itself will do most of task and there will be no need for these tools.

If meaning in large language models emerges through the self-organization of patterns across billions of training examples, does it follow that training on a language with a larger vocabulary (more distinct words) would produce better results assuming equally sized datasets are available for each language?

yes

Thank you. This is made so simple to follow.

What's the most surprising thing you learned about LLMs while preparing your talk?

@paraschopra Thank you for the wonderful presentation. Would it be possible to share the recording?

42 slides on how LLMs work is genuinely more useful than half the paid courses out there. saving this for a weekend deep dive

Fabulous stuff!

why are you guys not building foundational llms

Thanks for sharing this... will help us to understand...

well done, nice and straightfoward!

People can refer papers ig. Thanks for the fancy animations though

Gr8 one paras , very simple presentation but covered inside of llms beautifully, really loved it.,is there any recorded video of ur talk?

no

Thanks Paras for sharing this. Was really helpful

Any chance that a recording will be available ?

Thank you.

i'm glad you're teaching this. but once people see the slides they'll realize there are <br> like... 3 things you actually need to know.

Great slides, but 42 of them might oversimplify the hardest parts. Most people leave these talks thinking they get LLMs, but still can't predict when they'll fail.

Very concise and easy to grasp. Thanks for educating us.

Alpha

Thank you so much

the intuition-building pieces — why attention works, what embeddings actually represent — are what most technical explanations skip. 42 slides forces the right kind of ruthlessness. saving this for the next time someone asks me to explain transformers.