George Hotz | Programming | can tinygrad whisper? | AMD Rant | Giving up on AMD | Speech to Text
Date of stream 3 Jun 2023.
from $1499 buy https://comma.ai/shop/comma-three
Live-stream chat added as Subtitles/CC – English (Twitch Chat) – three-dot menu icon – Show transcript
Source files:
– https://github.com/geohot/tinygrad
Follow for notifications:
– https://twitch.tv/georgehotz
Support George:
– https://twitch.tv/subs/georgehotz
Pre-order tinybox:
– https://buy.stripe.com/5kAaGL6lk9uX9nW144 (https://tinygrad.org/)
Chapters:
00:00:00 intro
00:01:20 motherboard
00:02:06 we giving up on AMD
00:02:45 AMD drivers, bad AMD patch
00:04:35 Intel vs AMD driver patches
00:07:17 3 editions of tinybox
00:07:53 AMD driver will never be good
00:08:32 kernel panic not acceptable, contact at AMD, capture mindshare
00:09:55 should I buy AMD or Intel for ML
00:12:05 top priority for gpu division at AMD
00:12:55 Intel stable kernel driver, ROCm
00:13:25 did not expect running AMD demo app in loop to crash the system
00:13:40 Lisa Su respond to email, Jensen Huang
00:14:10 would fix AMD, no docs, short AMD
00:16:35 AMD does not care about their driver
00:17:37 Pat Gelsinger trying to turn around Intel
00:18:25 plan what to do, get tinygrad on mlperf
00:19:13 3090 vs 4090, NVLink, price/perf/power, jetson orin
00:20:25 how to fix AMD graphic division, callture
00:21:20 Lisa Su do you want to compete against Nvidia
00:22:05 meditation, Linus adding AMD GPU driver in kernel
00:25:54 user space, driver need to be stable, internship at AMD
00:27:23 bad culture, mediocre work not accepted, middle management
00:28:25 pytorch team, facebook, fb product stupid, twitter bad development environment
00:29:25 fire middle management tomorrow, Yann LeCun, Geoffrey Hinton
00:30:12 whisper openai, ai safety, open ai bad practices, Sam Altman
00:31:22 brand vs performance advertising, do not work at open ai Andrej Karpathy
00:34:00 whisper.py, locked tinygrad bounties, Twitter needed help, trust Elon, neurolink
00:36:00 AMD kernel driver 3 people team, hundred middle management, comma IPO
00:38:45 finding the whisper model
00:44:18 apology for Nvidia, how much we love Nvidia
00:45:50 no safety checker, no alignment, don’t talk to the government
00:54:00 writing whisper.py
00:59:00 AI Drake, Lil Mabu, Lil Windex, music generator
01:14:40 attention blocks
01:25:40 class Whisper, bach, chiptune
01:37:20 load, prep audio
01:41:37 goal by the end of the weekend, talking and words on screen
01:43:00 friendship with computers, 2030 girlfriend company, ai safety
01:47:50 Sam Altman vibe, normies thoughts on ai
01:49:00 Ford Foundation, inequality, power
01:52:23 Industrial Society and Its Future
02:03:00 life of homeless, oligarchy and not the oligarchy, us gov budget
02:07:30 pizza, west needs god, jina, conspiracy theories, NGOs
02:14:00 getting cancelled, military, taxation
02:17:15 colonies on mars, moon, fusion, hyperloops, homeless
02:20:20 challenging social contract, better soda, eugenics
02:25:00 tinygrad welcomes good pull requests, good genes
02:30:15 based department, being on plane guided by morons
02:38:00 reading the bible, cultural catholic, quran
02:45:45 pizza, power over nature, people
02:50:55 England, losing all the money
02:52:55 why Elon got so rich, worked hard, smart, did what it takes
02:54:30 targeted campaign attack against Elon, Kanye West, Biden, Podcasts
03:05:00 the primagen, dr disrespect, Martin Shkreli
03:11:00 ADHD, tweakers, alcoholics
03:17:00 sine wave, mozart
03:21:00 qkv_attention
03:39:30 TextDecoder
03:46:45 languages
03:54:20 tiktoken
03:59:35 good contributors in tinygrad
04:00:50 using comma three whisper llama
04:02:20 attention block
04:03:20 disktensor speed on mac
04:06:45 dataset audio
04:11:30 visualizing sound
04:12:45 load_state_dict
04:18:00 AssertionError can’t expand
04:19:30 code working, love transformers
04:21:10 commit whispering
04:23:30 python capture audio
04:31:10 hello twitch can you hear me
04:32:35 how great is whisper
04:43:20 pyaudio buffer size
04:47:00 can’t pickle local object
04:48:45 real time speech test, changing offset
04:54:35 true dat, working, laggy
04:55:20 key value cache
04:57:45 testing and finding bugs
05:04:44 not working anymore
05:10:00 it works
05:11:25 ai safety
05:12:40 piping whisper into llama, output llama into say, sexy gan
05:13:40 fixing bugs
05:24:40 multiple times decoding
05:31:00 need to merge speedups for realtime
05:31:40 thank you for watching
05:32:00 prison hierarchy
Official George Hotz communication channels:
– https://geohot.com
– https://twitter.com/realGeorgeHotz
– https://instagram.com/georgehotz
– https://tinygrad.org
– https://geohot.github.io/blog
– https://twitch.tv/georgehotz
– https://github.com/geohot
We archive George Hotz and comma.ai videos for fun.
Follow for notifications:
– https://twitter.com/geohotarchive
Thank you for reading and using the SHOW MORE button.
We hope you enjoy watching George’s videos as much as we do.
See you at the next video.