Making Transformers & AI fun and easy to understand through 3D animations, demos, use cases & AI news, plus ML interview questions. 🚀 Founder @ https://nitter.gq/t.co/Spq8K9bYTX
Joined October 2026
- Tweets1.2K
- Following951
- Followers847
- Likes1.4K
Pinned Tweet
I make Transformers and AI fun and easy to understand, with 3D animations, demos, real use cases and AI news.
Every morning: one animated Evolution Grid, six upgrades side by side and when to use each. This one shows how attention lets every field of a card payment look at every other field.
If you build LLMs, large transaction models, or fraud or credit models, say hi. I reply to everyone. #MachineLearning #LLM
This must be best transformer history video in 3d. Making long video always get ignored in X but i still choose to make it because I believe it will become everyone in #AI. So please give a little encouragement. A free like, repost, and comments are highly appreciated!!!!!
Happy Sunday ☀️ Let's connect: AI, ML and LLM engineers, solo builders, AI leaders.
Today:
- ML interview Shorts: LR schedules, cross-validation, missing data
- breaking AI news + our take
16-min Sunday watch: the Transformer, attention to DeepSeek
youtu.be/pXlH5depzsw
Jerry Liu retweeted
Happy Sunday ☀️ Let's connect: AI, ML and LLM engineers, solo builders, AI leaders.
Today:
- ML interview Shorts: LR schedules, cross-validation, missing data
- breaking AI news + our take
16-min Sunday watch: the Transformer, attention to DeepSeek
youtu.be/pXlH5depzsw
Happy Sunday ☀️ Let's connect: AI, ML and LLM engineers, solo builders, AI leaders.
Today:
- ML interview Shorts: LR schedules, cross-validation, missing data
- breaking AI news + our take
16-min Sunday watch: the Transformer, attention to DeepSeek
youtu.be/pXlH5depzsw
Best place to start. A Transformer reads every word at once, and each word looks at all the others (attention) to decide what matters.
That one idea, from "Attention Is All You Need" (Vaswani et al., 2017), is the engine behind ChatGPT. Here it is in 3D 👇
The full series, step by step: youtube.com/playlist?list=PL…
youtube.com/playlist?list=PL… this includes everything inside transomer in detail.
This is YOCO (Sun et al., 2024) grown up: a causal encoder builds the cache, the decoder layers only read it, so prefill can stop at layer 25 of 49.
The twist: 4 full-attention layers each project their own KV from encoder states instead of sharing one cache.
The real test is needle retrieval at 1M.
Quick question for the #AI and #LLM crowd: which concept still feels like magic to you?
Attention? Embeddings? Backprop? KV cache? Something else?
The top answer becomes our next animation, and we'll reply to every answer with the one idea that made it click for us.
Let's #connect. Drop yours below 👇
🚨 BREAKING: An AI-made drama just hit 6.6 billion views in China.
《持械入宋》: 10 seasons, ~1,000 episodes under 2 minutes each, in just over 2 months. No. 1 on the AI drama chart of ByteDance's Hongguo app. Made by a team with almost no film experience (figures as reported in Chinese media).
Our take: when production gets cheap, the edge moves to story, iteration speed and distribution.
So who needs a studio now?
Wow. THANK YOU. 🙏🔥
0 → 773 followers in just 7 days. And so many of you are verified builders, engineers and researchers I deeply respect. I'm moved, and honestly fired up.
Animation is how I fell in love with AI. So here's the paper that changed everything, "Attention Is All You Need", in 3D. ✨
Every full lesson is free on YouTube 👉 youtube.com/@tfmlabai
Let's connect and grow our AI knowledge together, five minutes at a time, in the fragments of every day. Let's go! 🚀
How do you train one LLM across thousands of GPUs? 🚀
Six ways, by what each one splits:
• the data (data parallel)
• the weights + optimizer state (ZeRO / FSDP)
• each layer (tensor)
• the layer stack (pipeline)
• the experts (expert parallel)
• one long input (context)
Llama 3 405B combined four at once. All six in 3D 👇
Full Short 👉 youtu.be/clY6tiippeg
FSDP or tensor parallel: which do you reach for first?
Why does RAG retrieve the wrong chunk? 🔍
Getting a RAG pipeline to work is quick. Knowing why it pulled the wrong chunk takes far longer, because chunking, embeddings and ranking each fail in their own way.
@StevenHWhite put exactly that to us yesterday, so we built the film: each failure in 3D, how to debug it, the fixes, and when to skip RAG altogether.
Full film 👉 youtu.be/T03_YGj4s78
Hybrid search or a reranker: which would you add first?
How does AI turn text into a voice? 🎙️
Text becomes phonemes. Audio becomes tokens too: sound pieces from a dictionary of 1,024, stacked rough to fine for every 13 ms slice. A Transformer writes those pieces in order, like words, and a codec plays them back as a waveform.
@pablogimenezz asked us to show this, so here it is, in 3D. Thank you, Pablo!
Full Short 👉 youtu.be/Wr1yJyzPGuw
Codec tokens or spectrograms: which wins for natural speech?
