galen

62 posts

galen banner
galen

galen

@G413N

cofounder https://t.co/ijW1mX8mRr

wandering Katılım Mart 2023
87 Takip Edilen696 Takipçiler
Sabitlenmiş Tweet
galen
galen@G413N·
computer use is too important to relegate to post-training. this has been many months in the making, I'm super proud of what we've achieved as a team and excited to scale!
Standard Intelligence@si_pbc

Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.

English
7
11
173
13.6K
galen
galen@G413N·
i'm excited for private models without capability tradeoffs, workshop is building this
Workshop Labs@WorkshopLabs

Open weights isn't open training. @AddieF38654 on our team wrote up her experience trying to post-train a 1T parameter MoE model using the existing open source infra. Let's find out how many monkey-patches it takes to post-train an open-weights model. A thread🧵

English
0
0
8
614
galen
galen@G413N·
@tszzl and we have so much room to scale
English
1
0
33
1.6K
galen
galen@G413N·
@JunliWang2021 @si_pbc loved reading the paper when it came out! lots of similarities to our process
English
0
0
8
182
Junli Wang
Junli Wang@JunliWang2021·
Great work by @si_pbc! Teaching computer-using agents on 11 million hours of video data is an unbelievable achievement. We touched on this direction with VideoAgentTrek (thanks for the cite!), but the scale and efficiency they've introduced here are genuinely next-level. Exciting times for the community!
Standard Intelligence@si_pbc

Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.

English
3
3
42
4.7K
Standard Intelligence
Standard Intelligence@si_pbc·
Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.
GIF
English
186
402
3.9K
1.1M
galen
galen@G413N·
@Oli82817545 os world measures language models, they have eg a "number of actions" which doesn't make sense if you're outputting an action 30 times per second this is very early research, we'll release something once we scale and I expect we'll need harder benchmarks very soon
English
1
0
9
813
Oli
Oli@Oli82817545·
@G413N looks really cool but iam a bit sceptical whats the os world score? also when will it be available as open model or an api?
English
1
0
2
193
aidan
aidan@aidanmantine·
@si_pbc truly everything is computer
English
3
0
79
3.4K
galen
galen@G413N·
general intuition is really something special, it's been amazing to watch Pim go in an entirely new direction as a founder and blow it away on execution, the culture there is incredible and they're doing great work, honored to have made a difference :)
Pim de Witte@PimDeWitte

Very excited for the SI team - fun fact, General Intuition likely would not have existed without Galen and his early mentorship as I was getting started in the field after @lachygroom introduced us. Having mostly traditional researchers in my network, and nobody who was self-taught like Galen, it was great seeing people paving their own path and being so far ahead of the curve. Follow this team!

English
0
0
35
3.5K
gelleproductions
gelleproductions@glproductions·
@si_pbc Is there any plan to ever release this model to customers for computer use? Or is this just a research project?
English
1
0
4
1.2K
galen
galen@G413N·
we’re assembling a 30PB storage cluster in downtown sf. got custom engraved drives for people helping. dm if you’d like to drop by
galen tweet media
English
1
1
41
4.7K
galen
galen@G413N·
@Maciek51285880 @si_pbc hmm silence from the generation code sounds like something is bugged, is this on the default prompts provided?
English
0
0
1
41
Maciek
Maciek@Maciek51285880·
@si_pbc Guys I've tried everything to run this model. Both Ubuntu + RTX 4090 and Windows. No luck at all. From inference i get non completing results (silence or cracks) and from server/client i got total silence. Does anybody got any results from the code on gh? 😦
English
1
0
2
140
Standard Intelligence
Standard Intelligence@si_pbc·
At Standard Intelligence we’ve been researching scalable cross-modality learning. We’re excited to share some early results in the form of 𝗵𝗲𝗿𝘁𝘇-𝗱𝗲𝘃, an open-source, first-of-its-kind base model for full-duplex conversational audio. 1/
English
52
132
831
177.2K
galen
galen@G413N·
@samsja19 (and it's an uneven 6-mount rack for 4 gpus)
English
1
0
1
72
samsja
samsja@samsja19·
@G413N Nice, curious about the spacing between the gpus, wouldn't be better to be evenly spaced between them ?
English
2
0
0
147
galen
galen@G413N·
built a 4x4090 space heater recently, took abt a week of debugging to get it running nicely. Thread to add public knowledge---
galen tweet media
English
50
53
1.1K
192.7K
galen
galen@G413N·
@samsja19 well we needed space for the Keeper of Melatonin
galen tweet media
English
1
0
1
104
galen
galen@G413N·
@inerati @rfleury I just remapped ctrl+s to commit and push and now it makes sense :)
galen tweet media
English
2
0
3
69
galen
galen@G413N·
psa in pytorch 2.3 the is_causal flag is no longer just a type hint. It's now necessary to avoid a silent kernel default to MemEff attention because Flash won't take any mask as input.
galen tweet media
English
0
0
4
755
/
/@gazorp5·
@G413N what was changed to fix the infiniband issue?
English
1
0
0
50

Keşfet

@tszzl @JunliWang2021 @si_pbc @hktsre @_neelr_ @Oli82817545 @aidanmantine @jannikschilling