A Decentralized AI Inference Journey: Darkbloom vs Common Compute
“By pitting Darkbloom against Common Compute on identical, high-end hardware, I aim to provide real-world data on which decentralized AI network offers the most compelling opportunity.”
https://t.co/chCsBeKwyd
For last 3 days, our objective has been simple: reliability and performance.
I'm excited to share with a lot of the improvements we've done to the load balancing algorithm and growth of the nodes to 340 machines (51kW):
We were able to achieve 99% uptime (from 89-90%). Absolutely exciting paradigm.
Few years back: you would have never related a distributed high-workflow network achieve this.
Welcome to the new era of computing!
Next objective: continue to improve performance and network utilization while maintaining 98%+ uptime
Rumble in the Jungle: Testing Darkbloom vs Common Compute
“For those evaluating entry into decentralized AI inference, Darkbloom’s setup experience lowers the barrier to entry while maintaining enterprise-grade control.”
https://t.co/dAxhcuFhq9
@darkbloomai I just spun up my MacBook Pro M5 Max and the dashboard’s showing “Everything’s Earning”. However, the “Trust & attestation” tab’s showing an “X” next to “ACME device-attest-01”. Is this cause for concern? Could my Firewalla Gold Plus be interfering with attestation?
Decentralized AI, simply explained:
- Anyone with spare hardware runs local AI and sells inference compute to others.
- To run massive models, multiple distant hardware units work together simultaneously.
- Data privacy is built-in during the process so absolutely nothing is left behind.
- Buyers get uncensored, private models for cheap.
This is an Uber-like moment.
Just as Uber changed the taxi ecosystem by directly connecting consumers, this will completely change the AI ecosystem.
darkbloom update:
- 1milly requests in total!! :)
- we hit ~2B tokens served this week;
- near 300 machines live on the network
- we're still early and there have been numbers floating around on how much you can make on darkbloom providing inference -- due to our early stage: we're conservatively sending requests and observing.
meantime: we have introduced an alpha program -- this will be base level incentive on top of your work. we will slowly take it off as we maximise the usage of the unused compute.
this week our focus is reliability -- scaling the requests, while keeping stability. a lot of bug bash.
thank you for tuning into our journey. there's also a surprise later this week!!
Wow. We closed in June 20th a little over 500M tokens served.
I guess, the next mark is 1B tokens in a day. We have found more efficiencies in the network where our providers can do 2x more capacity + concurrency.
The world's sleeping compute is waking up!!
When markets are bad.. pay attention to the people who still show up every single day building, educating, cooking and providing value even when engagement is low and timelines are quiet. The builders in the trenches during the hard times are usually the same people leading when the market turns around again.
$WMTX is the ticker 💪
Coverage. Privacy. Price.
That is the foundation of the new World Mobile consumer campaigns now rolling out across the U.S.
Better signal where people actually live.
Privacy built into the network itself.
Straightforward pricing without the usual telecom games.
A new generation of acquisition campaigns is now live.
Built for scale.
Just adding yet another UFO video to your timeline 😜
This one:
- stays mid-air for 6+ days
- flies at ~60,000 feet
- carries 500+ kg payload
- generates 100 kW 24/7
- has a 56m wingspan
- serves 500k+ concurrent users to traditional handsets
@WorldMobileTeam $WMTx