CFTC Chair on Hyperliquid: → He thinks HL is transforming the US markets with their 24/7 offering → Smart contract logic automates a lot of what exchanges have to comply with → ADL is an unconventional mechanism, not good or bad → Modernizing the rules to bring things onshore

elon fully ported chutes ai 🪂

Chutes is now a provider on @openGPUnetwork OpenGPU pulls GPUs from providers worldwide into one routing layer for AI workloads, with Relay giving enterprises AWS-style access and fiat billing on top. Now our TEE-enabled models live inside that layer. Teams on OpenGPU and Relay can reach them with no wallets and no infra setup. The GPU operators serving those models can't see your prompts or outputs. Both networks are after the same thing: pulling compute and models out of a handful of data centers and spreading them across a lot more hands. This is the reach decentralized infra was built for. More coming.

Decentralised compute gets stronger when networks stack. @chutes_ai brings TEE-enabled models. Opengpu brings the routing layer and global GPU supply. Relay brings AWS-style access and fiat billing on top. Each layer doing what it does best. This is the kind of partnership that can tip the scales in the future. Chutes 🤝 Opengpu

ethereum:0x163f8c2467924be0ae7b5347228cabf260318753 fixed this

day 6 of posting an update every day till convergance

ethereum:0x163f8c2467924be0ae7b5347228cabf260318753 proof of human thesis; billions

1/11 Deep dive resource thread on Chutes AI Bittensor (bittensor:native) SN64 - the full surface area every serious builder, researcher, and investor should know. Core site, consumer apps, docs, APIs, all 42 GitHub repos, research papers, hidden technical gems, and what actually matters for the future. chutes.ai 🪂

Open source and bittensor:native will solve the single-point model dependency problem.

apple glasses are going to be so fucking sick expecting them fall 2027/2028

Long Jobs Training as a Service @chutes_ai 🪂 Long Jobs is still listed as Coming Soon on their dedicated page for the full vision. This includes training state-of-the-art models, massive dataset processing, and multi-day workloads. What already works today: • The chute.job(timeout=7200) decorator in the SDK • Automatic retries, progress reporting, and file uploads • Status checking and result retrieval via the custom subdomain • Good for batch processing, embeddings, data pipelines, and lighter fine-tuning The bigger unlock they are building toward is Training as a Service. This lets people pay to fine-tune or pretrain models on the network with strong privacy guarantees via TEEs, at a fraction of hyperscaler cost. Parallax is the research bet that makes the economics work. Long Jobs is the product surface that exposes it. When both land, Chutes stops being inference you can trust and starts becoming training you can actually afford and keep private. That transition is the entire high-upside bet.

Zooming out on what Chutes is actually building: They are not trying to be the next closed frontier lab. They are trying to become the decentralized AI operating system. This is the infrastructure layer that lets anyone run and eventually train open models privately and cheaply at global scale. Current state (shipping today): • Strong serverless inference on open models • Excellent privacy via TEE enclaves plus post-quantum encryption using ML-KEM-768 • Very good developer experience with custom subdomains, SDK, and one-command deploys • Growing consumer surface including Search powered by Desearch (SN22), Chat, and Fictio Next phase (clearly in progress): • Long Jobs: currently Coming Soon for full training-scale workloads, but basic chute.job() background tasks with retries, progress tracking, and file handling are already live in the SDK • Parallax: the decentralized Mixture-of-Experts training architecture that makes serious fine-tuning and eventually pretraining viable across fragmented GPUs • Training as a Service: the long-term vision Jon Durbin has talked about If they execute, Chutes becomes the default place where serious builders and researchers do private, cost-effective AI work. Inference today. Real training tomorrow. That is a much larger total addressable market than just being another GPU rental marketplace.

One of the most underrated superpowers in Chutes is how they handle custom deployments. When you build and deploy your own chute using the Python SDK and Docker, you do not just get a container somewhere. You instantly get your own stable public endpoint in this format: your-username-your-chute-nam… Examples from their docs include: • myuser-my-chat.chutes.ai/v1 (OpenAI compatible) • your-username-sentiment.chut… • Job status at your-username-your-chute.chu…{job_id}/status This is legitimately their Vercel-for-AI moment. You define cords (API routes) and jobs, push it, and get automatic scaling on their decentralized GPU pool, optional TEE execution for privacy, file upload handling plus progress tracking for long-running work, and no need to manage Kubernetes or rent GPUs 24/7. Builders can ship production AI apps with modern developer experience while keeping prompts private. That combination is rare and dangerous to traditional inference providers.

This is the single most important document Chutes has released. Parallax (Jon Durbin 36-page tech report) is their answer to decentralized training at scale. The core problem is that normal internet connections cannot handle the all-to-all expert communication that big Mixture-of-Experts models need. Parallax fixes it with expert ownership plus low-rank surrogates. Each composer node only owns a slice of the experts and approximates the rest locally. Heavy updates get offloaded asynchronously to cheaper workers using compressed sketches. Early results on 20B-scale models are already within roughly 1.5 percent of centralized training while using dramatically less per-node compute and memory. The claim in their materials is up to 82 percent less hardware and resources for similar time targets, with training data never exposed to worker machines. This is how they move from cheap inference to actually competing on model development. Long Jobs is the product surface. Parallax is the technical engine. If this scales, training no longer requires a multi-billion-dollar tightly coupled cluster. That changes the economics of open AI dramatically.