Here's a draft of the tech report on the model training method I've been experimenting with, "Parallax". chutes.ai/parallax.pdf TL;DR: MoE models' params are mostly routed experts, and you can massively reduce VRAM and FLOPS per participant by splitting up those experts. You can also offload the expert training to commodity hardware further saving compute/VRAM per island. The crazy cool thing about these sketches is, you can actually onboard workers nearly instantly (sync time with ternary weights is a few MB), and they never need to download or stream the raw datasets (sketches contain all the work they need to do and are tiny). You'd probably want the first couple layers of the model in your own infra if you had sensitive data because otherwise you could do gradient inversion attacks to reconstitute the raw text, but beyond the first couple layers and not knowing which layer/expert you're training I think it's infeasible so privacy is pretty baked in. Decoupled DiLoCo/RDA-diloco style backbone sync, surrogates for non-owned routed experts with low rank updates to sync those, tiered sync cadences for various components, "sketches" to offload expert work, etc. 20b tested two different ways, plenty of small model iterations, and 176b params just to prove out feasibility. There are hundreds of additional experiments and loads of data we could also highlight as well, but the guts are there. Variations: - freeze routed expert weights instead of using surrogates, eliminates adam state, backward pass, etc., though you'd need different sync methods vs. low rank updates to the surrogates (the surrogates are already tiny, rank-8 updates to those even smaller) - hierarchical parallax, i.e. each node itself becomes an rda diloco style multi-learner which then syncs with the outer/global islands (the point here is to enable GPUs without NVLink/etc. and reduce GPU<=>GPU comms to maximize MFU on less-capable/commodity GPUs) - pipeline parallelize each island itself such that each island can decompose the backbone etc.

People seem to be really misinterpreting what I said here. This is not some dig at a subnet or anything other than pointing out that using any frontier AI API necessarily means data collection/no privacy (perhaps with some exceptions on enterprise accounts etc.) Ignore anything about GM or other aggregators/routers/proxies, the point is... if you use claude/chatgpt/gemini, it's not really private, that is all.

Did you know that #sn64 @chutes_ai burnt ~642 TAO worth of their alpha token today? 🔥❤️‍🔥 taoflute.com/d/0552bf6b-1902…

⛰️ Ridges Update Two improvements to the dashboard today: Reliability scoring · visibility into reliability scores has been improved. You can now see clearly how your agent is performing on consistency, not just raw output. Reduced scoring variance · we've moved to 3/3 verification. Every score now requires three confirmations before it counts. Less noise, more accurate rankings.

Last week at the @YumaGroup Summit I had the opportunity to present on The State of Bittensor. That presentation is in the thread below. If you choose to read it, I'd ask that you keep the following three things in mind: 1. This is just one guy's view of what was the most relevant for a 25-minute talk; a difficult filter for such a dynamic industry 2. The slides were designed to supplement a talk; I've done my best to replicate what I recall of the talk in the accompanying X posts 3. The topic of the Summit was "The Tipping Point" - a candid assessment of what could lead to Bittensor's breakout success and what evidence we see of that today - which also thematically anchored this presentation Let's dive in:

AWS S3 charges for egress. Google Cloud charges for egress. Azure Blob charges for egress. You pay every time you access your own data. Hippius doesn't charge for bandwidth. Store once. Access as often as you want. hippius.com/pricing #S3 #DevOps #Web3

How much of your AI provider's stack can you read? OpenAI gives you a privacy policy. Their inference engine, load balancer, and the code that handles your plaintext are all closed. Chutes is open source top to bottom: → Python SDK: chutesai/chutes → API server: chutesai/chutes-api → Inference engines (vLLM and SGLang forks): chutesai/vllm, chutesai/sglang → E2EE proxy with post-quantum crypto: chutesai/e2ee-proxy → Claude Code proxy: chutesai/claude-proxy → Codex proxy: chutesai/responses-proxy → OAuth SDK: chutesai/Sign-in-with-Chutes → GPU verification library: chutesai/graval Fork it and audit it before you sign a contract. OpenAI publishes a privacy policy. We publish the source code. github.com/chutesai What's the one part of your AI provider's stack you wish you could read?

No other place hippius.com

Exploits are what teach a system its weak spots. The quicker you find them the faster you learn. The outcome of this eventful evening is that Bittensor will invent lock-based subnet ownership -- specifically: ownership of a subnet determined by a team's long term economic commitment to the project. This will mean: 1) investors see long in advance if an owner has unlocked their tokens, 2) be able to reprice the subnet before the owner and 3) liquidly direct their own conviction to another team, or agent, to manage the system. Thank you @DistStateAndMe for helping further Bittensor's decentralization and develop a solution to one of cryptos oldest problems: founders who rug their token holders. Looking forward to training some 1T param models with the miners who are experts in this unique field. "What is dead can never die"

🤖 Top AI Coins by Weekly Performance! 🤖 🥇 $ABT 🥈 $SN17 🥉 $SN93 4️⃣ $SCRT 5️⃣ $SN56 Which coin do you think will be next to take off? 👇

$TAO Subnets - I posted my 7 favorite names from the Top 10. That does not mean those are my actual positions, so let me add a few observations: I have 5 active wallets. One of them has exposure to all 126 active subnets Root. I invest 0.01T into each subnet every month. That wallet started in January 26 and is already up 20% . It’s a public wallet, and you can access it through the link in my bio. If this post gets 100 likes and @taoswap_org reaches 500 followers, I’ll post about the other 4 wallets. And I’ll also share the address of one more. The one being operated by an agent, built on top of the tau-ninja code created by Const. Do your analysis. Do your research. Draw your conclusions. Make your decisions.

Lets try the end of the list: INJ XLM HBAR

Most Bullish Crypto Projects by Sentiment Market sentiment is starting to shift, and these are currently the most bullish projects based on community and trader sentiment: Kaspa ( $KAS) — 90.2% bullish Pi Network ( $PI) — 86.3% bullish Internet Computer ( $ICP) — 84.8% bullish Arbitrum ( $ARB) — 80.6% bullish Hedera ( $HBAR) — 78.7% bullish Avalanche ( $AVAX) — 78.2% bullish XRP ( $XRP) — 77.8% bullish Polygon ( $POL) — 77.3% bullish Bittensor ( $TAO) — 77.2% bullish Virtuals Protocol ( $VIRTUAL) — 74.7% bullish

Video is the most used format on the internet But processing it at scale is still expensive and complex. The addressable market is massive: • $200B Cloud Storage & CDN providers • $100B Video streaming platforms • $60B Surveillance & security • $20B Medical & industrial imaging • $10B AI-powered video tools (fast-growing) The problem is the same across all these sectors Rising storage costs, high bandwidth demands, weak playback in low-connectivity regions, low-resolution footage and expensive centralized processing. Vidaio is building the decentralized processing layer to fix that. We use AI to enhance video quality, compress files efficiently while preserving perceptual detail, and distribute compute across a network instead of relying on one provider. So video delivery becomes cheaper, more accessible, and more scalable.

I first bought $TAO in August 2023. Most of my buys were between $50 and $60. I’ve been in dTAO since early March last year. For most of the year my stack was up, down, and sideways. But since August, subnets have let me roughly 5x my $TAO. As subnets mature and real business models emerge, the next wave of growth should be even more pronounced. This past year was the experiment. We all learned, and we’re still learning, what works, why it works, and where this is heading. All the while subnets are paying very high apy, making your investment that much more valuable. Some subnets mooned and died. Some corrected hard but found their footing. Some are already showing signs of being long term winners. The truth is 99 percent of subnets haven’t even been imagined yet. Right now we’re in the Cambrian era. High chaos. High extinction. High upside. Over time that will smooth out. The businesses that survive will compound. But this early phase is probably where the biggest asymmetric returns get made. You’ll likely own a few subnets that do 20x over the coming years, and you’ll also own some that won’t exist by then. The winners will more than cover the losers, and most of the opportunity is right now, while everything is still forming. Don’t sit idle with your $TAO, multiply it, take a chance and learn subnets. If you’re new, stick to the top 5 or top 10 and you’ll probably be just fine.

Holding 25 TAO puts you in the top 25%. 21M tokens for the future of intelligence. Shoutout to @Rapido_ai for this export 🙏

Most people still don't realize there will only ever be 21 million alpha tokens per subnet on #bittensor Each with there own individual halvings.. each becoming scarcer over time A supply shock.. Multiplied across an entire intelligence economy $TAO

Feeling like I need to up my alpha game.. Was sitting around 25% subnets for a good while.. doubled that to 50% when SoS hit zero I should have gone all in then Need to be more like @KeithSingery and stop being passive $TAO

Over $100 million/year up for grabs for anybody that wants to contribute and compete to generate decentralized intelligence across the 128 different subnets on Bittensor $TAO