ceo @nodus_compute (yc f26) ex google, uber, scaleai, aws कृण्वन्तो विश्वं आर्यं [email protected] https://xitter.gq/t.co/y2D1YiTYZ9
home
Joined March 2025
- Tweets130
- Following90
- Followers141
- Likes427
equal accuracy can hide different failures.
compare the same requests.
count wins and regressions separately.
keep critical failures out of the tradeoff.
only then compare cost and speed.
nodus-compute.ai/blog/compar…
17
step 5 preview now on nodus
1.0m input / 379.6k output tokens per $1
current price: $0.97561 in / $2.634147 out per 1m tokens
We’re building Nodus around a simple question:
How much more AI can we get from the GPUs we already have?
Results so far:
• Up to 56 customer workloads on one GPU
• 2.4× as many requests answered on time
• ~5-second restores vs. ~70-second cold starts
Here’s how↓
1
35
added glm-5.3 flash to nodus
1m context, 20.5m input tokens per $1
current price: $0.048781 in / $0.195122 out per 1m tokens
We’re building Nodus around a simple question:
How much more AI can we get from the GPUs we already have?
Results so far:
• Up to 56 customer workloads on one GPU
• 2.4× as many requests answered on time
• ~5-second restores vs. ~70-second cold starts
Here’s how↓
1
59
better precision doesn't mean a usable classifier.
check the recall floor.
check daily review demand at the incoming positive rate.
if neither tested policy meets both, don't invent an untested midpoint.
nodus-compute.ai/blog/evalua…
28
deepseek v4.1 flash is up on nodus
44.3m input tokens for a dollar
We’re building Nodus around a simple question:
How much more AI can we get from the GPUs we already have?
Results so far:
• Up to 56 customer workloads on one GPU
• 2.4× as many requests answered on time
• ~5-second restores vs. ~70-second cold starts
Here’s how↓
2
95
one relevant hit can still miss the evidence that changes the answer.
fix the reading budget.
check the rule and its conditions together.
count duplicates as space, not new evidence.
nodus-compute.ai/blog/evalua…
2
61
deepseek v4.1 flash 1017 is live on nodus
1m context, 9.7m input tokens per $1
We’re building Nodus around a simple question:
How much more AI can we get from the GPUs we already have?
Results so far:
• Up to 56 customer workloads on one GPU
• 2.4× as many requests answered on time
• ~5-second restores vs. ~70-second cold starts
Here’s how↓
97
we spent an unreasonable amount of time building a gladiator arena so language models could fight over leetcode easy problems.
is this AGI?
(btw @nodus_compute won the speed battle ... console.nodus-compute.ai/inf…)
arenas coming soon
1
76
heres the full clip for our megakernel demo btw
really cool work fitting so much on A cards that many people discard
The world is borrowing hundreds of billions of dollars to build GPUs, and the average GPU is busy 5% of the time.
Nodus Compute gets more AI out of the GPUs that already exist.
We fill idle hours with useful work, pack more models and jobs onto every card, and run training, RL, agents and inference on one platform.
1
32
it came to our attention that someone is using nodus compute to promote a cryptocurrency scam
please MASS report @noduscomputeai as posers from indonesia, our ONLY @ is @nodus_compute and we do NOT support their grift
3
2
270
parney retweeted
The world is borrowing hundreds of billions of dollars to build GPUs, and the average GPU is busy 5% of the time.
Nodus Compute gets more AI out of the GPUs that already exist.
We fill idle hours with useful work, pack more models and jobs onto every card, and run training, RL, agents and inference on one platform.
parney retweeted
We’re launching Ascenix (@ycombinator F26): custom software for running senior homes. Our first customer was the retirement home we lived in for months. Now we’re growing 15% WoW across Canada and the US.
Senior homes face bottlenecks like time-consuming reporting, missed reimbursement revenue, and compliance gaps that risk fines and reputational damage. Solving each with separate tools can mean dozens of vendors, hundreds of thousands of dollars in spending, multiple logins, and siloed data.
We replace that stack with one platform tailored to each home. Teams describe what they need, and Ascenix turns their data and instructions into dashboards, reports, or automated workflows. Our reusable building blocks let us add new workflows as new bottlenecks arise, without another vendor or product.
Visit ascenix.co for a demo.
parney retweeted
Agents can now control phones.
Introducing Phone Harness (YC F26)
Let your agent use your personal phone, or its own iPhone/Android in the cloud
> Test and use real mobile apps
> Message and call people
> Grab verification codes for 2FA
Open source. Connect through MCP, CLI, or API.
Get started at phone-harness.com
parney retweeted
Introducing Chance by @Nodus_Compute
A new frontier in probabilistic intelligence.
Four architectures. Zero learned weights. One unified API.
Built from first principles. Available today.
Intelligence, left to chance.
Introducing Chance by @Nodus_Compute
A new frontier in probabilistic intelligence.
Four architectures. Zero learned weights. One unified API.
Built from first principles. Available today.
Intelligence, left to chance.
A coin. A die. A deck. A wheel.
Technical papers and API included.
chance.nodus-compute.ai
2
175
parney retweeted
Introducing: Phone Harness MCP 🔌
Any agent that supports MCP can now control a real iPhone or Android in the cloud:
> works in Claude, ChatGPT, Cursor, Codex, and more as a custom connector
> nothing to install, just paste a URL and sign in
> sees the screen, taps, types, and swipes like you do
Available in phone-harness 0.3.3.
github.com/phone-harness/mcp
parney retweeted
We’re hiring: Intern of Research Staff at Nodus Compute.
Help optimize AI workloads across training, fine-tuning, RL, agents, and inference.
Apply: careers.nodus-compute.ai/job…
Know someone great? Share this with them.