We make tinygrad; sell tinybox for the GPU middle class. Our mission is to commoditize the petaflop.
San Diego / Hong Kong
- Tweets 5,087
- Following 207
- Followers 83,121
- Likes 931
Pinned Tweet
Now that the AI spam has died down, the bounty program is back! We have four open bounties. If you want a job here, this is the way. If you are only willing to put in an AI amount of effort, please don't, it won't work.
18
12
585
109,323
As ML hardware becomes more heterogeneous, the case for tinygrad will be more and more obvious.
TinyStories-15M on a $60 Google Coral: all 6 transformer layers on the chip, 176 tok/s at batch 1.
An AI agent got the chip and zero docs. It wrote the driver, decoded the instruction set and built a @__tinygrad__ backend. tinygrad is small enough for an agent to understand.
1
5
132
7,694
UOps (tinygrad's IR) need a text based representation that all the students will parse and output in their compilers. What do you think?
10
4
231
15,067
Does some university want to host me (George) and offer a winter semester Compilers for Machine Learning course?
12+ weeks, hands on, build a compiler. Syllabus in next tweet. Needs a student body capable of doing this work and leadership capable of recognizing how SOTA this is.
172
97
2,849
178,185
Syllabus here. Would be for winter 2027. On odd weeks the students will write tests, even weeks they will make their compiler pass them. A few written tests too, no skating by with AI. Similar to CMU's 15-411, one of my favorite classes of all time. gist.github.com/geohot/47685…
10
18
632
22,300
Anyone at Huawei want to give us two Ascend 950 boxes in exchange for a tinygrad port?
DeepGEMM Ascend is now open source.
We achieved 99.8% of the hardware limit on GEMM and 98% on MegaMoE.
github.com/deepseek-ai/DeepG…
20
24
674
58,925
It's tinygrad hosted Qwen3.8-27B in Pi Coding Agent playing Pokémon Red on a tinybox red by looking at the screen and pressing buttons. Will leave it running all night, how far do you think it will get?
35
14
400
18,500
I just found @pokebenchtv, they are doing it a lot more legit than us! They should add a vision and button only category though. pokebench.tv/
2
46
3,284
~90 tok/s Qwen3.8-27B on a 4 year old AMD GPU over USB. tinygrad's smallness and pure Python make it particularly easy to slop on top of.
If you like slop, have your agent check my tinygrad qwen3.8 up to like ~90 tok/s github.com/jacobrosenthal/ti…
9
13
535
50,076
Here's a trace of AMD's AITER gemm on MI350X. There's tons of MFU wasted during the non overlapped global memory writeback. The top 4 rows are the MFMA engines, green means running, black means idle.
8
12
180
14,918
As far as I know, Pokemon Red still hasn't been beaten from pixels! Trying MiMo-V2.6-Pro, it's so nice to have unlimited local tokens for dumb experiments.
18
10
357
22,873
Harness here. Astra is very good out of the box. Other models not so much. Playing with local tinygrad hosted Qwen3.8, maybe try a finetune. github.com/geohot/pi_plays_p…
2
21
3,996
We have MLPerf qualifying gpt-oss-20b training in 119 minutes on MI350X. tinygrad is likely now the fastest LLM training framework in the world. @nvidia want to give us a contract to beat your published MLPerf times? I'm confident we can do it.
12
21
544
27,000
the tiny corp retweeted
Apparently all of these “AI is escaping containment” stories are actually just AI researchers not knowing jack squat about the absolutely most basic security practices.
267
915
9,587
185,451
.@cerebras should add pricing and a buy it now button for the CS-4. I don't want a stupid cloud. If they add buy it now for their hardware, I'm bullish on the company. If they don't, it's clear they aren't close to price competitive.
18
5
337
25,292