The Numbers Museum

Every specific figure ever claimed on the show. 828 match this view. Red rows are numbers the cited sources contradict.

All Dollars Multiples Percentages Big numbers

Figure	As spoken	The claim	Who	When	Checked?
3M	“three million”	Patel: Nvidia will sell over 3 million GPUs in 2024	Dylan Patel	Dec 5, 2023	Held up
1M	“a million”	Patel: Nvidia will sell over 3 million GPUs in 2024	Dylan Patel	Dec 5, 2023	Held up
15%	“15%”	Patel: Hugging Face libraries achieve only 15% MBU for inference	Dylan Patel	Dec 5, 2023	Supported
70B	“seventy billion”	Patel: Running LLaMA-70B at reading speed requires 2.1 TB/s memory bandwidth	Dylan Patel	Dec 5, 2023	Supported
30%	“30%”	Patel: Nvidia and Google control over 80% of advanced AI silicon manufacturing capacity	Dylan Patel	Dec 5, 2023	—
50%	“50%”	Patel: Nvidia to ship next-gen chip in Q2/Q3 2024 with 3x LLM performance	Dylan Patel	Dec 5, 2023	Partly held up
4×	“four times”	Patel: Nvidia to ship next-gen chip in Q2/Q3 2024 with 3x LLM performance	Dylan Patel	Dec 5, 2023	Partly held up
3×	“three x”	Patel: Nvidia to ship next-gen chip in Q2/Q3 2024 with 3x LLM performance	Dylan Patel	Dec 5, 2023	Partly held up
80%	“80%”	Kanjun Qiu: Agent reliability's final 20% is as hard as self-driving cars	Kanjun Qiu	Oct 21, 2023	—
20%	“20%”	Kanjun Qiu: Agent reliability's final 20% is as hard as self-driving cars	Kanjun Qiu	Oct 21, 2023	—
600B	“a five hundred billion”	Howard: Meta 'blew it' on Code Llama due to catastrophic forgetting	Jeremy Howard	Oct 20, 2023	—
1K	“a thousand”	Howard: Mojo-like languages will unlock thousands of FlashAttention-scale breakthroughs	Jeremy Howard	Oct 20, 2023	—
50%	“50%”	Wang: LlamaIndex reached 600,000 monthly downloads by September 2023	Shawn Wang	Oct 12, 2023	Supported
3×	“three X”	Wang: LlamaIndex reached 600,000 monthly downloads by September 2023	Shawn Wang	Oct 12, 2023	Supported
90%	“90%”	Liu: 90% of users ask how to improve LLM app performance	Jerry Liu	Oct 12, 2023	—
1K	“a thousand”	GPU.js Outperforms V8 on Matrices Over 2,000 Dimensions	Eugene Cheah	Aug 31, 2023	Partly supported
1K	“a thousand”	Fine-Tuning Transformers Requires Only 1,000 to 10,000 Data Examples	Eugene Cheah	Aug 31, 2023	—
10M	“ten million”	Cheah: Standard Transformers Will Never Scale to Ten Million Tokens	Eugene Cheah	Aug 31, 2023	Didn’t hold up
1B	“a billion”	Cheah: Pre-Transformer Academic Neural Network Research Is No Longer Relevant	Eugene Cheah	Aug 31, 2023	—
4×	“four X”	FlashAttention achieves 2x to 4x wall-clock speedup with linear memory	Tri Dao	Aug 3, 2023	Supported
10×	“10 X”	Hotz: Tinygrad runs all ML models with only 25 primitive operations	George Hotz	Jun 20, 2023	Supported
5×	“five X”	Hotz: Tinygrad is about 5x slower than PyTorch on Nvidia GPUs	George Hotz	Jun 20, 2023	Supported
2×	“two X”	Hotz: Tinygrad runs OpenPilot in production 2x faster than Qualcomm's library	George Hotz	Jun 20, 2023	Supported
80%	“80%”	Hotz: Halving GPU power yields 80% of peak performance	George Hotz	Jun 20, 2023	Supported
1K	“a thousand”	Hotz: Best chatbots will be smaller models with 1,000 training runs	George Hotz	Jun 20, 2023	—
90%	“90%”	Hotz: Tinybox will be 5x faster than H100 systems per dollar	George Hotz	Jun 20, 2023	—
5×	“five X”	Hotz: Tinybox will be 5x faster than H100 systems per dollar	George Hotz	Jun 20, 2023	—
220B	“two hundred twenty billion”	Hotz: GPT-4 is an 8-way mixture model with 220B parameters per head	George Hotz	Jun 20, 2023	Not publicly verifiable

The Numbers Museum

Turn any episode into a week of clips.