Nathan Odle
@mov_axbx
Followers
3,699
Following
322
Media
866
Statuses
7,440
Just acc. Preferably on gravel. I race rally cars, do computer things, and engineer stuff. pronounced “oh, dull”.
United States
Joined December 2008
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
ブロック
• 455277 Tweets
西田さん
• 288450 Tweets
Twitterくん
• 276603 Tweets
西田敏行さん
• 263545 Tweets
俳優さん
• 184519 Tweets
生成AI
• 148689 Tweets
Bluesky
• 131702 Tweets
AI学習
• 125948 Tweets
ブルスカ
• 118340 Tweets
ツイッター
• 117498 Tweets
Grok
• 105932 Tweets
スーパームーン
• 83798 Tweets
イーロン
• 77912 Tweets
インスタ
• 73277 Tweets
イラスト投稿
• 72798 Tweets
ぐりとぐら
• 65248 Tweets
pixiv
• 51775 Tweets
블루스카이
• 43324 Tweets
タイッツー
• 43058 Tweets
#INDvsNZ
• 37876 Tweets
Misskey
• 36073 Tweets
他のSNS
• 35863 Tweets
OUROAD UPCOMING PROJECT
• 32149 Tweets
池中玄太80キロ
• 30588 Tweets
ドクターX
• 25113 Tweets
無断転載
• 22324 Tweets
YIBO VOGUE FORCES OF FASHION
• 20877 Tweets
相互さん
• 15933 Tweets
絵師さん
• 12107 Tweets
Last Seen Profiles
Pinned Tweet
You have all been very patient, so here you go.
7x4090 Writeup.
And don't you dare shame my frontend skills.
Now go buy some GPUs and get to it.
78
111
1K
@vedugudev
Don’t know if you’re going off family name or what but dude lives there, sells stuff from there, ships from there, good enough for me
That’s the American Way, I’m fully aware of Japanese sentiment toward outsiders but I’m invoking the Transitive Property of Cultural Nomenclature
7
10
2K
OK so aside from games and some things like CAD and Adobe, there’s nothing much keeping people on Windows right?
That’s a really shallow moat.
IT should just push the Linux Desktop ball across the line once and for all.
366
39
1K
If you require any stuff from Japan you should go buy it from this guy. Video games, turntables/hifi, etc
6
24
976
Can someone tell me again why we keep our code and models in separate places
50
4
817
Please enjoy this SGI Indigo2 workstation from 1996 running llama2.c by
@karpathy
. 1.4 tokens/sec with the 15M TinyStories model!
Just a little porting for the big endian IRIX machine, all in an evening’s work.
What would ML researchers have thought 28 years ago seeing this?!
36
139
770
So here's my understanding of one of the groq tricks (the paper is really easy to read):
Since a neural network's computational graph is known at compile time, they also know at compile time how data will flow between computational units.
As a result, they can do away with
25
51
718
Ok ok I’m working on 7x4090 AI rig writeup tonight
Hope you find it worth the wait
54
28
662
@arpitingle
They know where all the cars are going when everyone wakes up for work and how fast they all drive so they can get rid of traffic lights and turn lanes by telling everyone when to leave the house
14
17
489
Waiting on a GPU and an electrician but this thing is about ready.
I’ve gotten a lot of questions about this AMD Epyc 7x4090 build, thinking of doing a Spaces on parts selection, etc if you guys think you’d be interested
45
20
454
@luke_metro
That’s not a bad thing. I love when I have an idea and find out someone else had it then thought it was significant enough to use resources on, means I had a decent idea
2
2
339
@deepfates
Startup idea: we make a site that hosts static websites for free, for people from cities around the globe to use and enjoy
We can call it GeoCities
8
9
272
You ever get a new car or the like and sometimes you just stop and look, taking in the raw essence of the thing
Yeah
26
9
252
@hlntnr
Helen, you and I almost certainly wouldn’t agree on a lot of things AI but I’m sorry for the way folks are treating you in these comments.
In an age where we’re working to wield some of the most powerful tech in human history, it’s important to behave as civilized beings.
10
2
250
You guys are GPU guys, you wouldn’t be interested in something like this. I don’t know why I’m even posting this pic
13
9
238
Someone asked me to ELI5 the
@GroqInc
deterministic network thing so here's your car analogy:
They know where all the cars are going when everyone wakes up for work (when they compile) and how fast they all drive (compute latency) so they can get rid of traffic lights
5
11
219
Summarizing 24 hours of Linux Desktop thoughts:
0. Forget everything you know about desktop Linux now. Yes, it sucks.
1. I wonder how much of the world realizes macOS didn’t just spawn into existence. It was built from NeXTSTEP in <3 years by professionals.
2. Hardware
41
9
214
One of my base assumptions in life is that given enough time I can learn and do just about anything
Pretty sure nobody went out of their way to tell me this, I just figured it out on my own
This is actually a big quality of life upgrade
12
10
203
@franklingraves
The more ironic take is that llama2 kicked off the open source LLM movement which is starting to compete on many fronts with OpenAI. MSFT owns GitHub and has a stake in OpenAI, so they enabled the competition by not complying with the takedown request
3
3
200
People might look upon religion more kindly if they think of it as social technology that has been reasonably successful at getting people to engage in long-term thinking.
This also provides a useful framework for judging the value of an individual religion to society, rather
8
14
196
I'm sometimes reminded how few people cross between the CS/EE/MechE worlds.
I do this for fun, and am not as competent as some who focus on just one field, but the combo spec has some advantages.
1/n
24
9
196
“What’s your AI compute stack look like?”
Um, it’s a stack…of computers I guess
13
7
184
@varun_mathur
You are assuming that board wants OpenAI to continue to exist. You may instead be telling them what they are hoping to hear.
2
3
175
Apple has Googleitis
Google got fat and lazy off ad revenue and can’t ship
Apple got fat and lazy off stuff like 30% App Store commission
It’s a free market and you shouldn’t punish winning, but these necrotic companies are bad for American progress
7
5
170
@techhsiren
Recognize that a lot more decisions are reversible than you think
Perfect is the enemy of the good
10
4
140
@ID_AA_Carmack
I asked a group chat full of semi types this very thing a few months ago. Consensus seemed to be that there weren’t any advantages but I remain stubbornly unconvinced that a zero vibration, high vacuum environment isn’t useful in some way.
15
3
129
Race wasn’t our normal success but we did come out with a new sponsor
*real pic not dingboarded lol
11
3
119
We have achieved Triton Inference Server on the 7x4090 internally
What a pain, but it cooks
6
8
116
@yacineMTB
Clearly there are those in the tech industry without the self-awareness to realize how silly they look in a black turtleneck. It should have been retired after Jobs, like a jersey number
4
2
113
Me: I want to pitch an OS startup
Everyone: we have an OS startup at home
The OS startup at home
10
1
111
Quit calling them 10x. They’re 1x and the rest of yous are 0.1x.
6
4
112
@AutismCapital
She's a fraudster and a criminal but I don't like people picking on her appearance. Not saying that's what this post is but plenty out there that are. Plenty to say about her character without devolving to playground insults. Let's be civilized
13
0
106
Especially considering that with AI coding, creating desktop apps for Linux that are comparable or superior to Windows is pretty straightforward.
14
0
109
Another vintage UNIX workstation running LLMs with llama2.c
This time it's SPARC: a Sun Ultra 45 from 2006 running Solaris 10.
4.55 tok/s on TinyStories 15M, not bad.
I reused the simple big-endian port I did for the SGI Indigo2 with no mods, should I submit a PR
@karpathy
5
11
105
Made the finals at the
@scale_AI
Hackathon with my voice assistant with knowledge retrieval.
Met folks like
@goodside
,
@karpathy
,
@bradlightcap
, etc and had fun fumbling through a challenging live audio demo. Humbled to be there, thanks to all! Congrats to the winners!
#AI
6
5
104
7xA770 is now 8xA770
Had to reflash retimer card to bifurcate Slot 2 to x8x8
Now I can run 8x tensor parallel instead of 4x <eyes llama3.1 70B>
Ignore jank cabling, will have to engineer a more permanent mount
8
4
107
ZIRP is gone, people are into C again, they're singing the national anthem, companies are bootstrapping
I think the pendulum just hit max U, look out
4
9
100
@Zack_Bioptic
For what I bought from him the prices were reasonable and his service was very good. In addition to the quick shipping and note, he messaged me immediately to make a personal contact.
I also get that price is important but it isn't always everything. I contrast this
1
0
94
0
0
95
@dabradybunch
@dennyishung
I mean, you realize Taipei is in Taiwan and there’s no CCP in Taiwan, right? I don’t see any hypocrisy.
0
0
92
7x4090 doing ~10 tok/s with a Q8_0 Llama 3 70B GGUF quant under llama.cpp, it's a start. No groq, but not terrible for a worst-case scenario
19
2
90
More thoughts on Linux Desktop:
I get it. You aren’t getting power plants and years of enterprise Excel macros off Windows. But who cares? Apple doesn’t.
The target audience is really anyone that could get along with Mac.
Get ASUS, Lenovo, HP, Dell, etc together and fund new
11
5
92
7xA770 giving off some cyberpunk vibes
Waiting on a few risers then it’s time to ride
13
4
87
So, I installed Quake 3 Arena here in 2023. From the original CD, circa 1999. I went into multiplayer and the thing connected to an online server where I was able to play Q3A. WTF
@ID_AA_Carmack
is this matchmaking running on a machine in your basement or what 😆
6
3
86
Prediction: NVidia order cancellations will be a thing in ~18-24 months
Not trying to convince anyone, you’re free to bookmark and call me out if I’m wrong
13
2
84
OK WOW.
Last I checked in on doing AI things with the 7xA770 machine it was a bit of a struggle. Docs were not great, lots of issues, etc etc etc.
But now, check it out. With the release of IPEX-LLM, Intel's framework for running AI on CPU, iGPU, and dGPU things look waaaay
6
7
77
Our friend
@yacineMTB
has talked about turning gaussian splats into CAD models but there's another (not mutually exclusive) way to make CAD better.
Why don't we have a Quest 3 or AVP app that lets you talk through creating a CAD model?
"Let's start with a 40 by 80 by 20mm
11
3
76
Neuromorphic RasPi-like device just dropped, what are you building with it
10
1
73
@var_epsilon
If you don't play with stuff you don't really learn. I posted yesterday about how as someone not in the industry I'm oblivious to a lot of things people in tech know about but sometimes I'm the one looking at them like, dude how do you not know this.
2
3
70
This take is 100% correct. Even if Yann and zuck are generally in favor of open source, their fiduciary duties mean there has to be a business case.
I'm OK with that, and also glad that the business interests are aligned with what the rest of us nerds want.
zuck is not a "good guy" or doing it to "support open source". it because it's a good business decision to drop a llama 400b nuke on closed labs
i am +ve about zuck and -ve about closedai not because of good vs evil, but because zucks good outcome is aligned with mine and yours
79
122
2K
4
4
69
New contest / experiment
Reply to or quote/repost this post to enter drawing for 1 free year of Dingboard (~$155 value)
Perhaps
@yacineMTB
can reply to confirm not fake. Drawing tomorrow
40
15
66
This reply to Karpathy is, in some ways, the most epic post I have ever read
Open source guys.
Offering Karpathy.
H100s
9
1
229
5
1
67
Making people believe AGI is right around the corner is an easier way to make money than actually building it
6
6
66
Well
@Icalbert101
as promised I got ahold of the
@PB_Swiss_Tools
ratcheting screwdriver to compare against the
@LinusTech
. Used them to reassemble the SGI Onyx and Challenge machines I was working on.
As much as I’d like to uphold the honor of my Swiss ancestors I have to hand
7
2
175
On the 7x4090 writeup I'll have purchase links for the gear I used
Will provide normal links to all items but may add affiliate links with the proceeds to go to some good cause
If so, which cause? The only non-profit that comes to mind at the moment isn't hurting for cash
5
0
63
Claude, and Suno and I would like to add to the rich cultural experience that is the Internet with this smash hit
7
10
64
Who wants an Arc A770 GPU?
In order to enter my drawing all you have to do is reply to this with one of the following:
1. Sci fi book recommendation (I like robots)
2. Your best recipe
3. A ‘Best kept secret’ place in a major city
Must live where Amazon ships.
@YoungPhlo_
72
3
61
@yacineMTB
Yeah, for the most part open sourcing stuff is not altruistic.
It's called removing your adversary's advantage because LLMs are a tool for you and a product for them
2
2
59
clone new repo
run script
notice it's downloading conda
^C^C^C
sudo adduser condatrash etc
logout
login as condatrash and run script
4
1
60
Lots of people asking in comments on my recent post how they can “study MechE”.
I get it, there’s a degree in MechE so that means you can study it. And yeah, go get some books on heat transfer, dynamics, etc. Basically just go look at the standard MechE curriculum.
But.
2
6
60
In the span of a few hours I participated in convos with people on X about fluid dynamics, restaurant design, heat transfer, AI, DNA, and gas station ops.
This place is a generalist's dream come true
4
5
60
To top it off, learn something about manufacturing itself. BOMs, supply chain, provisioning, etc.
You call yourself full-stack? What I just described, friends, is the real stack.
4/4
2
2
58
Heads up for cracked CUDA / game dev nerds.
You could make a mint simply by building good, fast software to work with 3D scans and point clouds.
Also ML for turning dense point clouds / meshes into accurate and sensical parametric models.
Lots of work with image to model but
6
1
59
How to succeed in business:
1. Don’t do anything illegal
2. Don’t overborrow / overspend
3. Tell the truth, while getting the word out
4. Take care of the customer
I’m sure people will want to add their own take in comments but it all pretty much boils down to this.
7
8
59
After close to 24 hours of screwing around and learning about magical HPC stuff, a Phi-1.5 LoRA finetune is actually running on this 7xA770 cluster via DDP
Baby steps, working on faster/better
Confirmed GPUs are working. It's training
9
5
56
Peak Apple was a long time ago. They’re riding that walled garden horse like Google does ad revenue.
Good thing is, SpaceX, Anduril, etc show it’s possible for scrappy upstarts to compete. Especially when incumbents can’t afford to hire up all the talent
ZIRP is decel
- Vision Pro Fail
- Mediocre M3 performance boost
- "new" iPhone with literally zero added features compared to iPhone 14 Pro (no one cares about Titanium)
- Tim Cook like how do I save stock price? Lets announce a $110B buyback
We're either at peak Apple or peak Apple is behind
301
139
3K
2
1
57