tiny corp is accepting new interns. more info on tinygrad.org and tinygrad discord | comma.ai is accepting interns comma.ai/jobs#open-positions from $1150 buy -> comma 3X comma.ai/shop/comma-3x | best ADAS system in the world openpilot.comma.ai | from $999 comma.ai/shop/body the future of people arxiv.org/pdf/2305.20050.pdf | huggingface.co/teknium/OpenHermes-2.5-Mistral-7B | mistral.ai/news/announcing-mistral-7b/ | github.com/mistralai/mistral-src | github.com/openai/prm800k Pre-order tinybox buy.stripe.com/5kAaGL6lk9uX9nW144 more info on -> tinygrad.org github.com/tinygrad/tinygrad

@doge1931

6 ай бұрын

Google don't work the same any more brother. So when u tell someone to google something they may or may not get a relevant answer.

@NithinJune

6 ай бұрын

Did you put the twitch chat in the captions?? hmmm maybe don’t do that

@geohotarchive

5 ай бұрын

@@NithinJune you can disable the captions or change them to English (auto-generated) if you don't want to see the twitch chat.

George I'd like to congratulate you for starting a stream with the microphone turned on. I know it's been a long road to get to this point, and it's no doubt been challenging at times, but I just want to express how proud I am that you've finally done it.

these are the best chapter annotations I have ever seen, it's extremely valuable

@crowlsyong

6 ай бұрын

ya, my god this is amazing

@andybaldman

6 ай бұрын

The only useful thing AI can currently do.

Thanks for providing timestamps ❤️

@geohotarchive

6 ай бұрын

You’re welcome 😊

@ShpanMan

6 ай бұрын

Absolutely, it's awesome! With so many timestamps I suggest using the Jumpcut Chrome extension, easily manages the timestamps in a visual order and shortcut to go to the next one.

@Aiworld2025

6 ай бұрын

@@geohotarchivedouble thanks 🙇‍♂️

thank you guys for uploading!!!!

@geohotarchive

6 ай бұрын

Thanks to George for streaming :)

Excited for real time STT/TTS conversations without any latency.

What George is doing is way beyond me. Props to him

Hi George ! I am really enjoying your streams. Thanks for that and keep up the good work ! 😎👏💪

I just like watching you work. The look on your face when you are working and thinking is priceless!

this is the most on topic stream yet! well done

Well done, George. This is very helpful.

Hell yeah! Kickin' off the morning with George!

"Let's just read the paper. We don't need to watch some KZreadr read the paper." ... pretty much all that separates the 5% that actually do shit (e.g. geohot) from the 95% who don't.

Hello Geo! Awesome to see you again!

thx for uploading

@geohotarchive

6 ай бұрын

@spacemonster8954 thanks to George for streaming :)

Very considerate george, adding the AI safety feature

ngl, I loved watching this, but didn't have a clue wtf was happening. It was pretty freaky though bc I understood where this is all leading. Freaky times. Can't wait for more. Thx George, sending positive vibes.

For 7b that's crazy good.

Vibes were awesome yesterday. Keep those ones goin!

Who the fuck puts the chapters in these videos, I'm amzed you are doing a fantastic job!

@geohotarchive

6 ай бұрын

Glad you like them :)

I'm currently working on deep RL life simulation and bumped into your video. What an amazing coincidence! Good bless you and all the technooptimists all over the world!

@katehamilton7240

6 ай бұрын

I'm optimistic that AGI/Superintelligence is a transhumanist tech bro fantasy. It's also marketing hype, which deflects from real problems like AI companies stealing data and exploiting workers. There are computational problems that cannot EVER be solved regardless of computing power and runtime, and it's not possible to build a general reasoning model without ontological input. Worry more about AIs replacing human workers at all levels.

@mwlulud2995

6 ай бұрын

Can you explain what is the goal of your simulation ? Thanks

thank you for the extra effort to archive chat logs

@geohotarchive

6 ай бұрын

You're welcome!

YOU ARE MY FAV PROPHET

How do you tag chapters in your video? Do you use some tool?

@geohotarchive

6 ай бұрын

@manishsharma7 it's all manual work unfortunately.

The glitching is because you probably have True Tone enabled on Display settings. Try closing the lid on your Macbook, or disable True Tone.

Always a fun time!

Holy shit this 7B model is insanely good!! At least 3.5-turbo level wtf?!?!

@sucim

6 ай бұрын

@@audreyp7653Given the recent developments I'm not so sure sheer number of parameters will be the deciding factor going forward but I would still be happy to see open source 7T models

@larion2336

6 ай бұрын

It's not 3.5-turbo level (although, maybe at coding?). I've tried a bunch of these 7b models at creative writing task lately and they all fail in one way or another. It's probably the case that they're much better in some areas than others though. 3.5-turbo is a better writer for example (though it's heavily tainted by alignment cancer)

@sucim

6 ай бұрын

@@larion2336 It probably is the coding, I haven't gotten around to test it for general Q&A or creative writing

@larion2336

6 ай бұрын

@@sucim They are quite good at Q&A I'd say, and it is impressive how coherent 7b's have gotten in general. But yeah creative writing seems to show big gaps between different models. That said I just tried a 180b Falcon model on huggingface (free chat thing) and it really wasn't any better than the 7b model I was using today, probably worse even. So I think a lot of it with creative writing is datasetting issues, they want to constantly end the 'chat' and often write "The end." followed by author notes and the like. Hard to find a model that is properly trained to write a novel style story. I know when it came to NovelAI they got some really good results in small models and it was I think mostly from training on the right data.

🎯 Key Takeaways for quick navigation: Q* algorithm *for path planning* - The Q* algorithm can be used to plan a path, potentially outperforming current state-of-the-art methods. - Adapting a language model like Mistral for path planning has not been done before, but it has the potential to provide better solutions. Optimizing Mistral *loading speed* - George Hotz made significant improvements to Mistral's loading speed, making it much faster to use. Potential applications *of the Q* algorithm* - The Q* algorithm could be used in various applications, including voice assistants, autopilot systems, and neural interfaces. Chat templating *with Mistral* - Mistral can be used for chat templating, allowing users to create custom chat conversations. Mistral's impressive *performance* - Viewers were impressed with Mistral's capabilities during the stream, with some even calling it "AGI" (Artificial General Intelligence). Comparing Mistral *to other language models* - Samuel5755 found Mistral's system to be superior to J.A.R.V.I.S. Made with HARPA AI

@hunterkudo9832

5 ай бұрын

Nice summary.

How come his twitch chat is in the captions? How does that work?

@geohotarchive

6 ай бұрын

github.com/TheDrHax/Twitch-Chat-Downloader

Who would have thought that Goerge woult talk to an LLM for over an hour one year ago (when we were still amazed about Avocado chairs)?

The real reason you need 3 monitors - Redundancy

this is damn good ..too good

Thanks

4H batch - nice to have GHOT back

might be an updated firmware package on vendors website for the display

YES!

I love this so excited to watch the whole 4 hr video. I had to skip ahead due to excitement lol

Pro tip: turn on CC to see the comments from the live stream

The speculation is so intense that even Big Foot is getting jealous of Q*.

@user-dm2ig3mf3w

6 ай бұрын

Hope it‘s not Q-Anon… 😅

Are the chapter annotations automated?

@geohotarchive

6 ай бұрын

@justinpakzad7532 manual work.

thanks for advocating for positivity, small wins!

He should put in a bounty to automatically make chapters for his streams to save our boy here big time each upload.

@bhavikyadav2934

6 ай бұрын

or make an ai to do it

It needs to be given directions. I noticed different answers between “can you do the thing” and “ do the thing and think about it step by step, explain your reasoning” .

2:49:28 Right here... testing the math using Q* algo

Mmmmh something really strange on perf...It's decent on a 3090/ryzen 9 7950 and completely crap (like 1char/3sec) on a 3080/ ryzen 9 5900HX (laptop)....both on windows 11 ...any idea why ??

@pschoro

6 ай бұрын

my bad !! tinygrad was not installed on the laptop yet !! Wow night and day with and without it !!! OMG !!! Congrats for all of that George !!!

@pschoro

6 ай бұрын

FUN FACT...it can't resolve this simple problem !! I planted a tree that was 8cm tall. At the end of the first year it was 12cm tall, at the end of the second year it was 18cm tall and at the end of the third year it was 27cm tall. How tall will it be in the fourth year ?

tips on where to learn and comprehend deeply what you are exactly doing here? thanks

@banknote501

6 ай бұрын

Just start somewhere. Build a very small model in Python. Maybe view Karpathys videos. But most important: Do stuff yourself. You are not learning from reading.

@zino123

6 ай бұрын

@@banknote501 ty I'll try

Q* - Probably a variant of reinforcement learning thats all.

@happywednesday6741

6 ай бұрын

Looks like ChatGPT + arithmetic engine. But news it's broken advanced encryption are complete nonsense

@joelwillis2043

6 ай бұрын

@@happywednesday6741 It is exactly a variant of Q learning in reinforcement learning.

@and_I_am_Life_the_fixer_of_all

6 ай бұрын

@@joelwillis2043 Q+A*=Q*

@joelwillis2043

6 ай бұрын

@@and_I_am_Life_the_fixer_of_all Thanks for restating what I said.

@and_I_am_Life_the_fixer_of_all

6 ай бұрын

@@joelwillis2043couldn't resist when it seems to be so rare to find someone that gets it

I can't run it on my pentium 4, please Qwen-72B on a tinybox, with tinygrad

Why does Hotz use Macintosh?

If you unplug your monitor from the power outlet and plug it back in, that can fix it.

@crowlsyong

6 ай бұрын

I was gona say...like dude unplug plug it back in- with power and video connector

*misspells the beautiful soup package name* "Did i just get supply chain attacked"

We need to ask seals if the are okay with rocket lauches!! 🤣🤣

GEORGE!!! can you PLEASE! make a video regarding Voice Assistants, and Neural Networks??

It must be soooo frustrating (you know what I mean) 😂 subbed and liked

"reward reasoning rather than just final answer" - all old maths teachers say this was obviously the way to go :-)

@banknote501

6 ай бұрын

I'm an old Math teacher myself. Yes, this is how you teach maths, but your students will only learn the already known reasoning steps. We want AGI to discover the reasoning itself and improve those steps and discover new things. George is right, only the final outcome matters. ML failed before when it tried to implement image recognition by using fixed feature extraction algorithms. Only way to success was End-To-End, image in, classification out.

Carefull! Is all fun and games until the time travelers swat team appears

I'm showing my kids this video when they ask how AGI was created

seeing George struggling to remain positive is funny😂😂

is this video for someone who is beginner in ML space?

@banknote501

6 ай бұрын

Yes, sure. This is so easy, if you cannot understand everything that George is doing, you should not event start with ML.

@yendar9262

6 ай бұрын

@@banknote501dude 💀

[robots scraping at the door] -We need to stay positive XD

are you gonna release it?

@CustomDabber360

6 ай бұрын

I could sit here 4 hours and code it. But im not that smart dude

I can’t see mistral branch of tinygrad. I thought it was pushed...

@dobcho95

6 ай бұрын

It is in the master branch already - examples/coder .py

where can i get the conversation.py.

@fanxiang3183

6 ай бұрын

i just got that in the github repo later. thank a lot

10/10

Hotz good

remember guys .. positive positive positive 🍃

It occurred to me that the best way to handle falling stock prices is to blame it on an advancement so powerful that it scared the board.

defo need a low cut on the mic :)

Can u overlay the chat below the camera? There are tools for that I think. Using comments isn’t great since it only shows 1 message at a time 😊

@geohotarchive

6 ай бұрын

@Abhinavneelam ...more - show the description Transcript - Show transcript - English (United States) - Twitch Chat You can see the whole chat on the right side of the screen next to the video.

go George! Us Hackers Unite!

"If there's one thing I never wanna be, it's Drake" 💀

"positivity positivity positivity" "this is a PAPER!?!" the struggle is real.

Given you can argue with GPT4 and try to convince it of something that is incorrect and it stands its ground about facts... I would assume they've created a very large environment with consistent states in the SARS sense of the word 'state'. This env, or world is probably much much larger than the training data for GPT4 because the model can extrapolate one thing to another and create very good data. Much like artists and scientists can flesh out a rolling culture and shared knowledge by hybridizing every combination of truth, trope, attitude, and expression I would expect they used GPT4 to fill out every possible distribution of truth to extend the original dataset. Not in a way that created any new groundbreaking culture or knowledge, but that it could triangulate truth into any and all less filled out distributions. I mean, it can create very hard to write computer code that usually works, for God's sake. This means it can construct rigorous synthetic data in any area they need. Ilya said in an interview a few months ago that running out of data was not a problem because they could create synthetic data. What rewards and actions they are giving the agents in the world I don't know. But I assume the actions and rewards are consistent during reinforcement learning because the facts according to GPT4, which built the environment are consistent.

lol, training a from a bench dataset is peak ultrafit

nice haircut hotz

Don't force positivity upon urself if it doesn't come naturally. Let the negativity flow out first, empty the vessel, and once vessel empties, follow it up with a "but" and then postface it with a not so derisive comment - maybe even a constructive one. Over time, things will improve. Cheers.

Does George work at OpenAI?

@geohotarchive

6 ай бұрын

@Powdermonkey99 www.linkedin.com/in/george-hotz-b3866476

Q predicted this

Yeah i need to spread more positive vibes cau...ARRGH JENNIFER, WHERE IS MY FCKN BEAR

my bait was clicked

who the hell uses two space indent in python

Interested in Q Learning in depth? Yannic Kilcher explains it all in What is Q-Learning (back to basics) kzread.info/dash/bejne/oIN2z5aaiai_gpc.html

i didn't know that Justin Timberlake write code

George can do AI, but I can use Selenium to automate my job. Win some, lose some.

3:06 It knew you altered the code for human in the loop, so it is pretending not to know

I'll be so fucken stoked if tg can run mistrial models!

todo: what is this... lmfao

George is great, but in this didn't he just kinda hook up an at home auto gpt?

I wonder whether geohot thinks he's a better programmer than most because I'm streaming this in the background hoping that the good programmer fairy dust will get into my head by osmosis. I dabbled in ADAS way before it was cool

this is priceless, ty @geohotarchive

high ball energy- dude you need some weed and modelo

openai doesnt publish papers they publish brochures

I'd like to neurally interface with my toaster in the mornings. Can George help a brother out? Thanks.

George your biggest issue is impatience,, just slow down a "Little"

K E T C H U P

you're reading a paper from may...

Gotta admit Q* is a sick af name 😭

Hey! Tomato!

q*!=qualia

1:29:00 LMAO