Gavin Uberti (@UbertiGavin) / X

Gavin Uberti

129 posts

Gavin Uberti

@UbertiGavin

CEO and Co-Founder @ Etched

San Jose, CA

Joined March 2022

Gavin Uberti
@UbertiGavin
Oct 12, 2024
It’s only reasoning if it’s from the Reasonique region of the human brain. Otherwise it’s just a sparkling stochastic parrot.
Aidan McLaughlin
@aidan_mclau
Oct 11, 2024
an efficient market would see apple's stock down 10% after publishing this
10K
Gavin Uberti
@UbertiGavin
Mar 21, 2024
The 21st century will be the most important century ever for humanity, thanks to the rapid advances in artificial intelligence. It makes no sense to sit on the sidelines in university. I'm excited to join the latest class of Thiel Fellows
Thiel Fellowship
@thielfellowship
Mar 21, 2024
Welcome to the Fellowship! businesswire.com/news/home/2024…
10K
Gavin Uberti
@UbertiGavin
Aug 28, 2024
Great thread walking through some principles of running LLMs on AI chips, written by an expert in the space
Dmytro Dzhulgakov
@dzhulgakov
Aug 27, 2024
Congrats to Cerebras on the impressive results! How SRAM-only ASICs like it stack up against GPUs? Spoiler: GPUs still rock for throughput, custom models, large models and prompts (common "prod" things). SRAM ASICs shine for pure generation. Long 🧵 x.com/CerebrasSystem…
5.4K
Gavin Uberti
@UbertiGavin
Oct 31, 2024
We’re proud to launch Oasis with @DecartAI, a video diffusion transformer that runs in real time. It’s a 500M param model that runs in real time on H100s, but our upcoming Sohu ASIC will be able to run 100B+ param models in real time.
Etched
@Etched
Oct 31, 2024
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!):
00:00
3.2K
Gavin Uberti
@UbertiGavin
Jun 25, 2024
Replying to @cHHillee
Hey Horace! Big fan of your work on Pytorch. In the blog post on our website (etched.com/memo), we outline exactly how our benchmark is calculated: > Benchmarks are for Llama-3 70B in FP8 precision: no sparsity, 8x model parallel, 2048 input/128 output lengths
6.7K
Gavin Uberti
@UbertiGavin
Nov 1, 2024
Replying to @teortaxesTex
I don’t think it’s a backhanded compliment - the model only stores the previous 24 frames in its memory. It’s an absolutely tiny model (500M parameters) with a tiny context window. We did this so it would run in real time on H100s that exist today.
779
Gavin Uberti
@UbertiGavin
Nov 2, 2024
I'm loving the absurd Oasis videos folks have been sharing. World simulators are WILD. A few of my favorites:
2.9K
Gavin Uberti
@UbertiGavin
Apr 19, 2024
Llama-3 400B is a dense model. From the Dwarkesh podcast with Zuckerberg:
5.5K
Gavin Uberti
@UbertiGavin
Nov 1, 2024
Replying to @nrehiew_
It’s open source, and is frame by frame diffusion (10 steps per frame). Check out the code yourself!
415
Gavin Uberti
@UbertiGavin
Jun 25, 2024
Replying to @cis_female
You could in theory stuff *way more* than 5x the FLOPS on the chip. It only takes around 10k transistors to build an FP16 fused-multiply-accumulate circuit that can run at 2 GHz (i.e. 4 GFLOPS)
1.2K
Gavin Uberti
@UbertiGavin
Apr 5, 2025
There is no data wall.
6K
Gavin Uberti
@UbertiGavin
Mar 23, 2024
Really glad to see folks finally changing the RoPE hyperparams
Marvin von Hagen
@marvinvonhagen
Mar 23, 2024
Mistral just announced at @SHACK15sf that they will release a new model today: Mistral 7B v0.2 Base Model - 32k instead of 8k context window - Rope Theta = 1e6 - No sliding window
4K
Gavin Uberti
@UbertiGavin
Feb 20, 2025
Using more FLOPs should make transformers smarter. DeepSeek R1 currently uses ~8 routed experts per token. So would selecting more (and possibly scaling by the router) improve performance? Come test it out at the Inference Time Compute Hackathon and win up to $60k in prizes!
Etched
@Etched
Feb 19, 2025
We're excited to partner with @cognition_labs @mercor_ai @CoreWeave and @AnthropicAI to host an inference-time compute hackathon, featuring >$60K in cash prizes and >1 exaflop of free compute.
2.7K
Gavin Uberti
@UbertiGavin
Feb 20, 2024
Replying to @swyx @felix_red_panda and @GroqInc
They split each layer on 8 chips, and Llama-2-70B has 64 layers, which gets you to 512 chips. The remaining rack of 64 is for other overhead (e.g. de-embedding).
814