@everythingmeta

All opinions somebody else's Computers and stuff MTS at Parallel Web Systems. Previously Head of Forward Deployed Engineering at Baseten.

 Location, Location
vlad.cc
Joined January 2011
  • Tweets599
  • Following221
  • Followers332
54Media
iykyk
3
108
$1B used to mean something
2
13
669
20% of YC startups whenever a tweet like this drops
kevin
@YngkvnOct 20
Inference is a beautiful business. deepseek reports 155T tokens/day from one inference cluster at $38/gpu-hr. I estimate their cluster at ~5 MW, implying ~$160M/MW/year in API revenue and 84–92% margins after $3–6/GPU-hour rent working assumptions: 96% input-cache hits, ~107 input token/output token ratio -> ~$14.35 revenue per billion tokens at off-peak prices. working backwards from 155T tok/day, 60,000 uncached input tok/sec/gpu prefill + 14,000 output tok/sec/gpu decode, each two-GPU pair handles approximately 1.5m input + 14k output tokens/sec -> 2400 gpus. on blackwell hardware I assume 2 kW/gpu including external networking, cache storage, etc. No wonder everyone wants be part of the token flow / bid up GPUs :)
6
466
Vlad retweeted
yang you
@yangyouOct 21
we used Parallel to build Parallel’s first brand campaign. here's a look at the research, connections, and creative decisions behind the artwork.

See the whole picture: A brand campaign built with Parallel

Article · @yangyou
21
18
21
300
36,397
Horrified by what happened to our dear friend Heidi Levine while she was on assignment today. “They just kept kicking me and kicking me,” she said. washingtonpost.com/world/202…

Israeli settlers attack Palestinians and Washington Post photographer in West Bank

The settlers attacked Palestinians harvesting olives and beat Heidi Levine, a veteran photojournalist. Witnesses said Israeli soldiers did nothing to halt the violence.

washingtonpost.com
39
206
17
487
39,802
Typesafe's CEO @CompleteSkeptic does not trust public benchmarks. Neither do I, Diogo. So @p0 and I wrote a blog evaluating 8 decision models on real company training datasets we have personally audited and iterated on. For what it's worth though, Jev is still on top 👑 parallel.ai/blog/are-all-sys…

Are all System One models created equal?

At Parallel, we train a lot of our own classifiers. But for every classifier we train, there are use cases we can’t prioritize, because each classifier…

parallel.ai
11
15
6
177
49,447
can't next-token predict any of this. pure chaos.
10
381
parallel.ai/blog/are-all-sys… We spent two weeks extensively testing decision models on real production use cases. It's been super fun, we learned a lot, and we are sharing our findings! 👇
8
10
3
58
6,488
more replies
3. We got big lifts applying DSPy-like iteration to the prompts. One of the many benefits of having a universal classifier is that you can n-shot prompt it rather than fine-tuning a classifier
2
2
303
4. At real scale, we still get better performance from our own tuned and trained models (much smaller models) but there is no compelling reason to start this way until you know what you are optimizing for
1
234
POV: it's 2026 and you want to make a reservation at your local restaurant that has 3.2 stars on Google Maps
2
6
228

Using specialized data to improve research agent accuracy

Article · @ethank_6
2
9
5
32
5,725
Vlad retweeted
vicki@vboykisOct 8
trying to imagine explaining to myself 10 years ago that we made a Markov chain so long that the pope is mad about it
17
147
6
2,570
58,575
The problem with a lot of AI assistant use cases is that specifying the work to be done in sufficient detail actually IS the work.
64
124
23
2,964
64,342
my entire timeline rn is versions of: “Feynman Cubes are non-transitive in polyamory space?!? This is a nuclear bomb for math PhDs”
2
250
OpenDots now has @p0 Web Search. Every agent in OpenDots can search the web, with sources linked and no API key needed to get started. Thanks to the Parallel team for integrating this into OpenDots!
Atai Barkai
@ataiiamOct 6
OpenDots 🤝 @p0 Web Search OpenDots can now search the entire web! No need to point your agents at a specific page. Available out of the box with every template.
2
6
33
1,900
Not too closely please
6
283
Load more