Tenzin
673 posts

Tenzin
@10_zin_
MLE @togethercompute EX @amazon MSCS @GeorgiaTech understanding the inference stack.
Katılım Ekim 2019
916 Takip Edilen342 Takipçiler
Sabitlenmiş Tweet

Super excited to be building this ! Provisioned Throughput is a win-win for customers and providers 🫸🫷
Together AI@togethercompute
We're introducing Provisioned Throughput: reserved inference capacity for frontier open models, with token-based pricing and a 99% uptime SLA. Serverless simplicity, guaranteed capacity, up to 90% lower cost vs. Opus 4.8. Get started with MiniMax M3 + GLM-5.2, read more 🧵
English


@cursor_ai This is such an impactful research field, that engineers can lead.
Almost akin to babysitting while models train.
A lot of space for defining best practices and design tradeoffs.
English

Claude/Codex are the generators.
Humans are the discriminators.
Even if coding agents get 1000x better.. if the engineer using it does not develop their understanding in tandem. The output result will remain slop for difficult problems.. since the agent will easily fool the humans.
We must maintain Nash Equilibrium.
English

Reference:
Highly recommended read !!
Physics and Economics based principled introduction to the Space DC industry.
vixra.org/pdf/2510.0065v…

English

@dylan522p @OnodaCapital Hey @dylan522p would be great to open a subscription to DC Models + Core Research. We want the hard core stuff.
English

@OnodaCapital Accelerator, TCO, datacenter models + core research had this earlier + fewer words.
But those don't explain tech
What intellectual beat off
What section isnt needed and doesnt contain value except for the sales blurb
We try to keep it short as possible so please provide feedback
Jacksonville Beach, FL 🇺🇸 English

Dunking on the customers because they don’t feel the need to intellectually beat off to 10,000 words
Always a shame to see great success lead to great hubris
(Paying sub, would pay more!)
Dylan Patel@dylan522p
Since peoole had trouble understanding the report, we've decided to make a major change. We will no longer drop 10k word nuanced pieces that describe what's happening in AI from models to datacenters to silicon. Instead we will do infographics even the buy side can understand.
English

If you use "rate" or "irate" in prometheus for measuring qps.
Please STOP !
You are missing all the spikes in your measurements.
valyala.medium.com/why-irate-from…
English

@AravSrinivas Alright good enough reason to re-install.. been slacking on mandarin.
English







