Sabitlenmiş Tweet
arkheτ.hl
3.6K posts
arkheτ.hl
@arkhet
trillions of tokens monthly, powered permissionlessly by @chutes_ai — the most secure, fully open-source, decentralized inference provider 🪂 https://t.co/daLuZBCxwD
bittensor subnet 64 Katılım Ekim 2024
3.2K Takip Edilen3.1K Takipçiler
RFQ is now live on Hyperliquid.
Well... on the Hyperliquid ecosystem.
@hypurrdash, top 25 builder codes according to @hl_eco, has just launched RFQ on its frontend.
You may have noticed a lot of people on your timeline getting bullish on @variational_io because of its RFQ system, which makes it much easier to support a large selection of RWAs without requiring a deeply liquid order book.
In short, RFQ (Request for Quote) lets traders request firm quotes from one or more market makers before executing a trade.
Unlike an order book, where orders are publicly displayed and matched against visible liquidity, RFQ routes orders directly to market makers, avoiding exposure to the broader market.
Hyperliquid can't build everything itself, and that's exactly why innovations from other teams are finding their way into its ecosystem, even when they're not developed by Hyperliquid directly.
Note: This isn't a sponsored post, just sharing something I found interesting on a platform I genuinely like, especially for its cohort feature (go check it out). Hyperdash's RFQ feature is currently available by private access only.
Hyperdash@hypurrdash
Introducing RFQ on Hyperdash. Professional traders require institutional liquidity. Trade larger size with improved execution and better pricing.
English
Long as fuck $BTC . Long as fuck $TAO . Long as fuck $AERO . Long as fuck $LIT .
$ETH validator still humming after 5 years. $LINK isn't my biggest bag anymore and I'm unfortunately out of the native pool but I'm not going to risk seeing the vision realized and not play.
$NOCK $TIG $REPPO $SERV ice cold storage/staked and ready for zero or hero.
Picked up the only stock other than $TSLA i've bought in last 5 years, $ABCL (thanks @cyberprince_rwo )
$SPCX accumulation plan established.
Couple real life biz ventures finally hitting breakthroughs after months of grinding, my squad is going to EAT.
Pushing in every possible direction all at once. I can't fucking quit moving.
English
@arkhet @MiniMax_AI @Hailuo_AI Yeah, seens to be an amazing video model, but most likely closed source afaik
TSUBAKI@AI__TSUBAKI
MiniMax H3 Early Access Made with #MiniMaxH3 on @Hailuo_AI
English
day 50 of posting an update every day till convergence
arkheτ.hl@arkhet
day 49 of posting an update every day till convergence
English
Napkin math on bits per byte vocab adjusted etc, this is also beating olmoe 7b 1ba at matched tokens (which is 30% more active params, 40% more total params). Maybe competitive with MobileMoE if we did a complete several trillion token severely overtrained run.
Quantile balancing
LatentMoE
Hybrid GDN-2/SWAX/MSA
Role differentiated mixed precision
I think we have a winner. More napkin math, I think this 5b model would be able to do around 1200tps inference on 5090s.
Kinda want to do the full 9t token run on some 5090s of this tiny one, for funsies. Parallax-5b-instant (or we could bump total params up to 10b for free if we don't change active params, or maybe higher)
Jon Durbin@jon_durbin
167k tokens per second training throughput on a single 8x 5090 box with parallax (on a toy 5b 1b active test moe)... That's quite high. Every GPU you add reduces compute for the others so that number goes up > linearly. Exciting times. And look at the beautiful loss curve, beating DDP baseline at matched steps (identical warmup/lr/dataset/etc.) With 5.99 GFLOPS per token for this arch and BF16 denominator that's ~60% MFU equivalent. 3.068 nats so far at ~19b tokens (llama-3 tokenizer), beating the DDP baseline at matched token budget and getting there ~15% faster and 64% cheaper, and remember it's ternary routed experts... Best guess as to nats gap is separate expert muon optimizers with the surrogate feedback+fold normalization etc. Almost ready for the big runs...
English
English
Currently, LLM training, inference, and compute subnets (Chutes, Engy, Targon, Lium) are surging on #Bittensor. Most of the major weight is concentrating right here.
It feels like Bittensor is quietly transforming from an intelligence blockchain into a data center.
TAO is called an “intelligence token,” yet in practice it’s being treated more like a profitability engine for compute providers—exactly how traditional data-center companies operate. Subnets are taking models that @nvidia gives away for free and using them to farm TAO injections.
Whether this validates the value of $TAO as an intelligence token or simply reflects @const_reborn’s current philosophy for Bittensor is still unclear.
But if we run Bittensor like a data center, we will devalue TAO.
The TAO paid to miners already exceeds what customers are willing to pay for the service. Treating TAO as free prize money can make the numbers look workable in the short term. But just like Bitcoin, TAO only sustains real value when the reward roughly matches the real cost of the electricity and hardware required to earn it.
We have to protect the original intent: AI innovation.
Recognize that TAO has genuine value. Prioritize what is actually needed. Commit to the long-term vision instead of short-term extraction.
Bittensor is a powerful platform for serious builders precisely because TAO has value. Keeping that value requires discipline, responsibility, and a longer horizon than the next emission cycle.
English
Watch the full episode: youtube.com/watch?v=CLydsO…
The paper behind Parallax: chutes.ai/parallax.pdf
YouTube
English
arkheτ.hl retweetledi
Privacy isn't really about privacy.
Jon Durbin's argument: it's not that a frontier lab wants to read your data. It's that every prompt you hand them trains the next model, the one they decide who gets to use. You are building their moat, brick by brick.
And if your company's edge goes into that training data, the model that comes out the other side doesn't need you anymore.
Parallax aims to be the destroyer of these moats.
From DropZone Episode 1. Full episode in the reply.
English
@dschamis @guru__hl @HyperliquidX @HypeStrat @HyperliquidR @hl_eco liquidity begets liquidity
revenue begets revenue
hyperliquid
English
This is an incredibly helpful tool to have when researching @HyperliquidX and @HypeStrat.
I was sitting with my favorite AI this week, trying to create something like this that could be shared—but @HyperliquidR and @hl_eco beat me to it (which I am perfectly happy about).
As we play with this tool, we will have comments and questions for them and the community.
To my #TradFi friends, I need to point something out here. Because everything on Hyperliquid is open and public, when you look at the financials, they are literally up to date as of the day you are looking at them.
For instance, normally, trailing twelve months (TTM) today would mean the twelve months ended June 30. Here, it literally means the twelve months ended YESTERDAY!
Hyperliquid Research Collective (HRC)@HyperliquidR
Today, HL Eco and HRC are incredibly excited to launch the new "Hyperliquid Financials" dashboard. hl.eco/financials The core reason we built HRC is that Hyperliquid intentionally leaves an information gap. They don't spend time on this, they're just busy building. That information asymmetry clearly affects how the market understands and values Hyperliquid. We started this journey by aggregating the best Hyperliquid research in one place and publishing financial reporting (hyperliquid:native and Trade[XYZ] quarterly/annual reports). Today we're taking that mission a step further, moving to live reporting. It's no secret TradFi is paying close attention to @HyperliquidX and @tradexyz. But that attention needs clear, accessible data, the kind that lets investors, analysts, and every other stakeholder do more than just read the raw numbers. With this, you get the most meaningful information on Hyperliquid in one place, CSV downloads for the underlying data, and analytical and projection tools to help you interpret it. No more raw data and nothing else. This will be improved even further with your feedback. We want this to be the Hyperliquid financials dashboard you've always wanted, so if you have suggestions or want to collaborate, we're all ears. Community is what Hyperliquid has been about from the start. Collaborative work by people who genuinely want to see it succeed in housing all of finance. We're incredibly happy to now be close partners with @hl_eco; this dashboard is what happens when two teams that actually want to provide value build together. Hyperliquid. HRC. HL Eco.
English
@Lighter_xyz @vnovakovski lighter is 0 fee for retail lmao
30% of 0 is 0
English
Lighter's new referral program is now live!
Refer traders, earn up to 30% of the fees they pay, and receive weekly USDC payouts.
Learn more in the link below:
docs.lighter.xyz/referral-progr…
English
We are pleased to present our latest research paper:
From Point Forecasts to Distributions: Adapting @GoogleResearch TimesFM 2.5 to Synth
This paper documents the seven steps needed to adapt this open source Time Series Foundation model to compete in Synth's price volatility competition
Full paper available below
English