Decentralized compute networks (DCNs) offer cheap GPUs. But | 0xPrismatic στο Binance Square

Αρχική σελίδα

Ειδοποίηση

Δημοφιλέστερα άρθρα

Με σελιδοδείκτη και "Μου αρέσει"

Ιστορικό

Κέντρο δημιουργού

Ρυθμίσεις

--

Decentralized compute networks (DCNs) offer cheap GPUs.
But enterprises aren't touching them (mostly).
WHY? WHY? A 🧵

Αποποίηση ευθυνών: Περιλαμβάνει γνώμες τρίτων. Δεν είναι οικονομική συμβουλή. Ενδέχεται να περιλαμβάνει χορηγούμενο περιεχόμενο. Δείτε τους Όρους και προϋποθέσεις.

5

0

Εξερευνήστε τα τελευταία νέα για τα κρύπτο

⚡️ Συμμετέχετε στις πιο πρόσφατες συζητήσεις για τα κρύπτο

💬 Αλληλεπιδράστε με τους αγαπημένους σας δημιουργούς

👍 Απολαύστε περιεχόμενο που σας ενδιαφέρει

Διεύθυνση email/αριθμός τηλεφώνου

Σχετικός δημιουργός

@0xPrismatic

Ανακαλύψτε περισσότερα από τον Δημιουργό

planning a token launch? Pivot to an IPO instead. you can still rewrite that investor deck tonight before sending.

--

We believe Bittensor is the most compelling place to witness Crypto x AI unfold. Here are some lessons we learned from talking to teams and watching subnets in the wild: 🧵

--

Bittensor now has over 110 subnets—and counting Here are 12 Bittensor subnets that have stood out to us, and we're keeping on our watchlist👇

--

Just released a detailed deep dive on decentralized training. We cover a lot in there, but a quick brain dump while my thoughts are fresh: So much has happened in the past 3 months and it's hard not to get excited - @NousResearch pre-trained a 15B model in a distributed fashion and is now training a 40B model. - @PrimeIntellect fine-tuned a 32B Qwen base model over a distributed mesh, outperforming its Qwen baseline on math and code. - @tplr_ai trained a 1.2B model from scratch using token rewards. Early loss curves outperformed centralized runs. - @PluralisHQ showed that low-bandwidth, model-parallel training is actually quite feasible... something most thought impossible - @MacrocosmosAI releases a new framework with data+pipeline parallism + incentive design and starts training a 15B model Most teams today are scaling up to ~40B params, a level that seems to mark the practical limit of data parallelism across open networks. Beyond that, hardware requirements become so steep that participation is limited to only a few well-equipped actors. Scaling toward 100B or 1T+ parameter models, will likely depend on model parallelism, which comes with an order of magnitude harder challenges (dealing with activations, not just gradients) True decentralized training is not just training AI across distributed clusters. It’s training across non-trusting parties. That’s where things get complicated. Even if you crack coordination, verification, and performance, none of it works without participation. Compute isn’t free. People won’t contribute without strong incentives. Designing those incentives is a hard problem: lots of thorny issues around tokenomics which I will get into later. For decentralized training to matter, it has to prove it can train models cheaper, faster, and more adaptable. Decentralized training may stay niche for a while. But when the cost dynamics shift, what once seemed experimental can become the new default, quickly. I'm watching closely for this.

--

Waiting on @PluralisHQ's paper to drop. So we can update our upcoming deep dive report on decentralized training and hit publish. Hopefully within the next 24 hours!

--

Τελευταία νέα

Hut 8 Warns Against Fraudulent Activities in Brazil, South Korea, and Hong Kong

--

Ethereum's Role in Stablecoins and DeFi Gains Attention Amid Market Developments

--

Mempool Warns Against Phishing Scams Using Its Logo

--

ALEX Protocol Faces Significant Loss Due to Security Flaw

--

ETF Issuers Urge SEC to Reinstate 'First Come, First Served' Principle

--

Προβολή περισσότερων

Χάρτης τοποθεσίας

Προτιμήσεις cookie

Όροι και Προϋπ. της πλατφόρμας

Χάρτης τοποθεσίας

Όροι και Προϋπ. της πλατφόρμας

Προτιμήσεις cookie