@bernhardsson profile picture
Erik Bernhardsson @bernhardsson

Building everyone's favorite AI infrastructure platform @modal

Joined February 2013
  • Posts 13,352
  • Following 3,861
  • Followers 55,538
675 Photos and videos
One way to think about @modal in the limit is that we're an aggregator of compute on a global scale and resource manager for that. Basically a very large global computer.
Charles 🎉 Frye @ ICML
@charles_irl
10h
x.com/i/article/207395394671…
1
2
38
4,572
My cat is celebrating America’s 250th birthday in the most American way: taking risks
12
3
327
14,355
A bit funny how FTX continues to make money from all the weird holdings. They bought a cube of tungsten for $200k in 2021 and tungsten prices are up 10x since then.
54
97
6,993
655,713
It should be more widely discussed how annoying it is that “Tier 1” means top notch for almost everything except for data centers where it has the opposite meaning
10
1
134
14,306
Erik Bernhardsson reposted
SGLang
@sgl_project
Jun 23
Huge congrats to @modal on Modal Auto Endpoints! 🎉 Owning your inference means owning the code that runs it, and we love seeing that philosophy in action🫡Honored that SGLang powers part of the stack, alongside the work on DFlash speculative decoding 🧡
Modal
@modal
Jun 23
It is not too late to _actually_ own your inference. Introducing: Modal Auto Endpoints.
0:52
9
10
67
7,195
Managed private LLM endpoints, now available for everyone in @modal. Deploy in a few clicks with the UI or a few keystrokes with our CLI. The coolest thing is that these are not black boxes – customers have full access to the code underneath.
Modal
@modal
Jun 23
It is not too late to _actually_ own your inference. Introducing: Modal Auto Endpoints.
0:52
21
20
170
27,452
At the same time, open source LLM inference is moving super fast – with amazing projects like vLLM, SGLang, FA4, etc. We're actively contributing to those because we think everyone benefits from moving the frontier.
1
2
25
3,259
This is why we're taking a different approach with LLM inference. We want customers to have full control over the serving code. When you're running Modal Auto Endpoints, you run the best open source models with the best open source engines, and you have full access to the code.
2
3
18
1,404
Modal is about having the best building blocks and it's easy to forget in the era of the cloud that they are actual physical objects. Luckily I was able to tour the blocks factory the other day.
Monishee Matin @ICML
@monishee
Jun 23
“Let me give you a tour. Isn’t this place so cool!?!” dear god pls don’t ask me what anything costs
2
2
84
12,757
If you want to use LLMs in production for serious stuff, you should probably use open source models
Adam Azzam
@AAAzzam
Jun 19
x.com/i/article/206808131498…
10
8
252
44,468
Erik Bernhardsson reposted
Charles 🎉 Frye @ ICML
@charles_irl
Jun 20
Speculation Is All You Need. In this blog post, we announce the co-release (w/ Z Lab) of six more state-of-the-art DFlash speculators for @Alibaba_Qwen 3.x. Over 1k output tps for 3.5 122B-A10B on a B200. Read the blog for why we're all-in on spec dec. modal.com/blog/spec-is-all-u…
0:02
35
101
700
188,178
Running 100k sandboxes is a hard scaling problem that's not for everyone – @modal is very excited to be one of a small set of providers that can handle this. In fact, @modal was the only one to nail every single iteration of this test leading up to the final results.
ComputeSDK
@computesdk
Jun 19
🚨 Benchmark Friday 🚨 After weeks of work from us and the participants, today we are excited to release the results of the 2026 100k Scale Invitational. Over the next hour we'll be releasing each set of results. Huge shoutout to @namespacelabs for powering our benchmarks!
9
7
109
24,431
What's our secret? That this is business as usual for us, and has been for a long time. We have many customers running 100k or even 1M sandboxes with us on a regular basis. Large RL runs for instance. Looking forward to the 1M sandbox challenge – we are ready any time!
37
6,047
Want to set your agents loose on Microsoft Word (or Minesweeper, Notepad, Diablo 4, or whatever?) Your prayers have been answered! github.com/modal-projects/wi…

GitHub - modal-projects/windows-sandboxes

Contribute to modal-projects/windows-sandboxes development by creating an account on GitHub.

github.com
Peyton Walters
@peywalt
May 22
on the day of modal's series c announcement i am ... getting microsoft word to run in a modal sandbox
3
3
65
15,139
Our sandbox team has been on a crusade against every millisecond of latency and it's paying off. More cool results coming very soon!
10
7
207
17,934
67% now x.com/jolandgraf/status/2065…
Erik Bernhardsson
@bernhardsson
May 19
There’s like 6-7 NYC infra founders I occasionally get drinks with and ~50% of them just got acquired: - @graphite - @astral_sh - @StainlessAPI Congrats!! 🍻
8
3
135
56,105
I have had a weird theory for years that Uber fits their ETA model using mean absolute error regression instead of ordinary least squares. Since wait-times have a long right tail, this will cause a bias that underestimates mean wait times (median < mean).
John Loeber 🎢
@johnloeber
Jun 6
Standard experience booking an Uber “5 minutes away” I order it “Finding your driver…” “Pickup in 7 minutes” I open my timer 9 minutes and 21 seconds later, it arrives This is so tedious and I am tired of being deceived at the margin
7
1
73
27,441
Load more