Gavin Uberti Profile
Gavin Uberti

@UbertiGavin

Followers
3K
Following
496
Statuses
115

Building model-specific AI chips @ Etched

Menlo Park, CA
Joined March 2022
Don't wanna be here? Send us removal request.
@UbertiGavin
Gavin Uberti
3 months
@juberti @OpenAI Congratulations to Sam for nabbing a world-class engineer!
0
0
8
@UbertiGavin
Gavin Uberti
3 months
@Tanishq97836660 Very cool!
1
0
0
@UbertiGavin
Gavin Uberti
3 months
@krystof_mitka Very well written! I love this sort of research
0
0
2
@UbertiGavin
Gavin Uberti
3 months
Has anyone been able to get to a death screen in Oasis? Has anyone made it into the end?
@Etched
Etched
3 months
Introducing Oasis: the first playable AI-generated game. We partnered with @DecartAI to build a real-time, interactive world model that runs >10x faster on Sohu. We're open-sourcing the model architecture, weights, and research. Here's how it works (and a demo you can play!):
1
0
8
@UbertiGavin
Gavin Uberti
3 months
This is a pretty neat trick:
@IsaacKing314
Isaac King 🔍
3 months
AI minecraft fact: If you walk backwards, it almost always keeps you walking level, which then forces it to generate totally flat terrain out in front of you.
0
0
2
@UbertiGavin
Gavin Uberti
3 months
@mattshumer_ @Etched It’s not running on Etched hardware yet - it’s running on H100s, so we had to keep it a very small proof of concept (500M params) with only a one second context length. Once Sohu is released, we will be able to run a 100B+ version of this model, still in real time!
1
1
9
@UbertiGavin
Gavin Uberti
3 months
@teortaxesTex Memory is generally not the issue for video models - it’s almost always interconnect bandwidth and FLOPs
0
0
1
@UbertiGavin
Gavin Uberti
3 months
@teortaxesTex It’ll take both! It took a ton of previous research to even get to this point (s/o to diffusion forcing, this model currently does 10 passes per frame) but there are other interesting techniques that bring this down to ~2
0
0
4
@UbertiGavin
Gavin Uberti
3 months
@Teknium1 It's a tiny (500M param) proof of concept, akin to BERT's 340M parameters in 2018. But new ASICs will enable 10s of GBs of KV cache and 100B+ param models to run in real time - stay tuned for more!
2
0
11
@UbertiGavin
Gavin Uberti
3 months
@teortaxesTex But new hardware like Sohu will enable huge, 100B param models with 10s of gigabytes of KV cache in their context windows. The promise of diffusion world models needs object permanence (and much more) - we’re hard at work building the hardware to make it possible.
2
0
5
@UbertiGavin
Gavin Uberti
3 months
@nrehiew_ It’s open source, and is frame by frame diffusion (10 steps per frame). Check out the code yourself!
1
0
15
@UbertiGavin
Gavin Uberti
3 months
Our ASIC Sohu will enable the exploration of entirely AI-generated worlds
@elonmusk
Elon Musk
3 months
@Etched @DecartAI Wow, this is happening fast
4
3
66
@UbertiGavin
Gavin Uberti
3 months
@daftura @Etched @DecartAI Check out GameNGen - You can’t play it though, sadly
1
1
3
@UbertiGavin
Gavin Uberti
3 months
@0000CCS @bons_vincent @Etched @DecartAI This is a tiny model, just 500M parameters. But giant world simulators, where you give it a prompt and it generates an interactive videogame, are going to be the future. It'll require new hardware, which we are building!
2
0
10