the tiny corp (@__tinygrad_

the tiny corp

4,515 posts

the tiny corp

@__tinygrad__

We make tinygrad; sell tinybox for the GPU middle class. Our mission is to commoditize the petaflop.

San Diego

Joined June 2023

@__tinygrad__

See the tiny corp’s full profile

Open X

Pinned
the tiny corp
@__tinygrad__
May 12
All products are now available to order! $45k for tinybox green with 5090s, and we have two variants of the tinybox pro v2, one with 5090s and one with RTX6000 Blackwells (that's a whopping 768 GB of RAM)
the tiny corp
@__tinygrad__
Jul 10
comma's model takes 0.14ms to run on the eGPU! We are still wasting a lot of time enqueuing though, fix coming with HCQ2 where we compile all command queue dispatches to C.
the tiny corp reposted
Elliot Arledge
@elliotarledge
Jul 10
GLM 5.2 FP8 with the native MTP at c=1 on 8xMI300X averaged 137 tok/s and peaked at 183 tok/s in a tough LRU coding problem! thanks @__tinygrad__ for the GPUs
00:00
the tiny corp reposted
Ramez Naam
@ramez
Jul 9
A week ago many were saying that the AI race was over. That it was down to just Anthropic and OpenAI. Or that Anthropic had taken a permanent lead. Now we have solid models from OAI, xAI, Meta, and open weight from GLM, with line of sight to Gemini 3.5 Pro, ChatGPT6, and more
the tiny corp
@__tinygrad__
Jul 8
We haven't focused much on Python wall time in tinygrad. Does someone with a good LLM loop want to try? Who thinks they can get 10 seconds? Speed up the wall time of `time STEPS=10 PYTHONPATH="." python3 examples/hlb_cifar10.py` without persistent caching
the tiny corp
@__tinygrad__
Jul 8
We need to make a video like this of the tinybox line.
comma
@comma_ai
Jul 8
Designed in America, Built in America. comma four is $100 off for Independence Day 🇺🇸
00:00
the tiny corp reposted
Harald Schäfer
@___Harald___
Jul 7
Anthropic is now a banned vendor at @comma_ai . I recommend other companies do the same. People should consider how reliant they are on cloud token vendors. If you can't brainstorm, code or be productive without them, you should be scared. Regain your sovereignty or you're ngmi.
the tiny corp
@__tinygrad__
Jul 7
We'll all look back in five years and remember when the enshittification cycle got too fast. Any company that relies on Anthropic is so obviously foolish. Google and Microsoft knew how to slow play it.
Claude
@claudeai
Jul 7
We're extending access to Claude Fable 5 on all paid plans through July 12.
the tiny corp
@__tinygrad__
Jul 7
In a few years, between hardware and algorithmic improvements, this will all look as laughable as governments restricting encryption to 40-bits.
Jukan
@jukan05
Jul 7
CHINA CONSIDERS RESTRICTING OVERSEAS ACCESS TO CUTTING-EDGE AI MODELS China’s Ministry of Commerce has led meetings over the past month with major AI companies, including Alibaba, ByteDance, and Z.ai, to discuss measures that would restrict overseas access to
the tiny corp
@__tinygrad__
Jul 7
You start out adding one bit of nonsense. It's just a little nonsense you say. But then you have 3 workarounds elsewhere in the codebase for that one bit of nonsense. Then you add 17 little hacks to deal with those 3 workarounds. And soon, your whole repo is hacks and nonsense.
the tiny corp
@__tinygrad__
Jul 6
I cannot believe how good GLM 5.2 is. Several weeks in now and it's mostly all I have been using. It doesn't have the alignment issue of cloud AI, it's much more clear what it can do and can't because it doesn't care about you taking another $$$ spin at the token slot machine.
the tiny corp reposted
Paul Gustafson
@paulpgustafson
Jul 5
Tinygrad finally got around to writing up a clean spec, looks a lot more promising --
tinygrad/spec/tinyspec.pdf at master · tinygrad/tinygrad
From github.com
the tiny corp
@__tinygrad__
Jul 3
We are here when you are ready @tenstorrent. We love how you build in public and have buy it now buttons on your website (unlike certain other fake chip companies). Let's get you a software stack to match!
Joe Fioti
@joefioti
Jul 3
I really want to like TT (cool hardware!) but every time they make announcements like this I go look and its always hand-optimized special-cased nonsense that runs *just* this model at like batch_size=1. No one buys hardware to run a single model, you guys need a proper compiler!
the tiny corp
@__tinygrad__
Jul 2
why hello sir, can I interest you in an exabox?