Raven is developing a distributed network of compute nodes for Artificial Intelligence and Machine Learning. A decentralized computing protocol built on #BNB
Hong Kong and Bangalore
- Tweets 4,592
- Following 210
- Followers 12,766
- Likes 9,174
Pinned Tweet
πRaven wins the @mozillabuilders Award at the Fix-the-Internet showcase.
Recognition by the top open source organization in the world that Raven Protocol is a core infrastructure project.
We demonstrated best traction and commitment to fix the internet!
medium.com/ravenprotocol/ravβ¦
57
57
189
1/ Open source developers working with frontier models like DeepSeek V4.1 or GLM-5.3 face a question most distributed training tools donβt answer well:
How much compute does this model actually need across my specific hardware setup, and how should it be allocated?
2
3
234
2/ Getting it wrong means:
- Memory errors before training starts
- Compute budgets assigned manually to each node
- Configuration that breaks when hardware changes
- Hours of setup before a single gradient moves
Most devs spend more time on this than on the actual training run.
1
17
DeepSeek dropped V4.1-Flash-1017, MIT licensed, 1M context, beats GPT-5.6 Sol on DeepSWE benchmarks.
Calling it a minor update. Terminal-Bench jumped from 31.2 to 42.3.
The open vs closed gap isnt closing gradually. Its collapsing in point releases.
chaosagentsai.substack.com/pβ¦
3
302
1/ Last Week in Open Source AI & Distributed Compute
- Meta releases Muse Spark Open, an open-weight multimodal model with 816B parameters and a 1M-token context window
- DeepSeek releases V4.1-Flash-1017 for coding and agent tasks, alongside public Search and Fetch APIs
1
1
255
4/
- Anthropic makes Claude Code Review generally available to Team and Enterprise customers.
- Sierra publishes a draft of the Personal Agent Protocol (Poppy), with 35 additional design partners including OpenAI, PayPal, Visa and Mastercard.
1
41
1/ Ben Horowitz on why AI infrastructure isnβt repeating the 2001 bandwidth crash.
In 1999, bandwidth was $1,500/megabit. By 2001, it was $50. Companies that borrowed to secure supply were wiped out overnight.
AI is running the opposite dynamic.
Ben Horowitz on the lesson LoudCloud, his early cloud company, taught him about today's AI neoclouds, and why they now face the opposite problem:
"It can end up being like a very dangerous arbitrage business, where you can get upside-down easily. And so if you look at LoudCloud, just as an example, here's a simple thing that happened to us."
"In 1999, bandwidth was $1,500 a megabit. And you couldn't get it. It was very hard to get it. And so you had to get a long-term contract, three years. By 2001, it was $50 a megabit... And if you borrowed money to get it, you're dead."
"That kind of thing is so out of the control of the entrepreneur."
"I think the core dynamic of AI has been working for CoreWeave, though, in that there's a massive shortage, not an overage of supply... Token prices have gone down, but GPU prices have not."
"And by the way, data center space is going up in price because even labor to build data centers is hard. The politics are making it hard to build data centers. So the whole supply chain price is going up and demand is not subsiding."
"Those are very big tailwinds for CoreWeave, despite the fact that the debt markets, kind of independent of them, are tough."
@bhorowitz w/ @chamath on @allinpod
1
2
301
2/
- GPU prices remain elevated despite efficiency gains
- Data center costs rising, not falling
- Labor to build data centers increasingly scarce
- Demand not subsiding
Supply chain costs going up. Demand holding. Thatβs a very different setup than 2001.
1
43
1/ Distributed training has obvious bottlenecks, GPU memory, network bandwidth, node failures. Most developers spend time debugging all three.
The one that rarely gets attention: cluster formation. Get it wrong and every training run inherits the problem before it starts.
1
1
245
Microsoft just trained MAI-Base-2, 2.8T parameter MoE, 16k GB200 GPUs, 38 days. No open weights.
The model architecture isnt what keeps frontier AI closed. The training infrastructure is. Distributed training across consumer hardware is how that changes.
aiweekly.co/ai-news-today
1
195
Nous Research wrapped their open source Hermes agent for enterprise, multi-step workflows, private deployments, production-ready.
Open source agents are moving into enterprise production. Training infrastructure at that scale is what gets tested next.
aiagentstore.ai/ai-agent-newβ¦
6
265
1/ Heterogeneous hardware is the reality for most open source developers training across consumer GPUs.
Mixed generations, different memory, varying network speeds and one slow node that quietly holds back the entire cluster.
Most frameworks donβt have a good answer for this.
1
1
239
2/ Standard distributed training treats all nodes as roughly equal. When they're not:
- Slow nodes stall the cluster
- Dropping them wastes gradient contributions
- Manual workarounds delay the run before it starts
Underutilized hardware. Longer training runs than necessary.
1
1
27