Rody (𝜏, 𝛼)

Ninja | Subnet 66 on Bittensor

6 days ago

continuous shipping 🚢

@Ninja_Subnet

6 days ago

We’ve published a new public dataset from the Ninja validator: 600 qualified tasks and 1,675 screened-out task candidates, for 2,275 total, plus 44,752 qualification and duel rollouts from six retired kings. The dataset will grow with the competition. Whenever a new king is crowned, the retiring king’s tasks and rollouts will be automatically archived. https://t.co/8n2KbljkhI We expect the dataset to become increasingly useful as full LLM traces accumulate. We will be exploring partnerships with model training subnets to post-train models for Ninja through SFT, preference optimization and reinforcement learning, with the goal of making Ninja their preferred agent harness.

621

yrodz12 retweeted

2 months ago

You can force an architecture into decentralized training, or you can craft an architecture specifically for decentralized training (and ultimately insanely performant inference). Stop trying to fight gravity. "Parallax", so far, is working exceptionally well. Surrogates, expert offload, stratified decoupled diloco, etc. All pretty great to make training work at scale across the internet. But... The biggest benefits are actually AFTER training, via specific model architecture choices that drastically reduce VRAM and increase throughput. These giant MoEs are wasting so much VRAM and compute for all of the many routed experts, it's insane actually. So much research already showing that MoEs can be massively compressed/pruned/ternary-distilled/etc., not to mention they are generally comparable to around 2.5-3x dense model equivalents of the ACTIVE parameter count. KV cache is a huge bottleneck, and these huge routed experts are often "dead weight" that consume it needlessly. We need to do better, so we will.

jon_durbin's tweet photo. You can force an architecture into decentralized training, or you can craft an architecture specifically for decentralized training (and ultimately insanely performant inference). Stop trying to fight gravity.

"Parallax", so far, is working exceptionally well.

Surrogates, expert offload, stratified decoupled diloco, etc. All pretty great to make training work at scale across the internet. But... The biggest benefits are actually AFTER training, via specific model architecture choices that drastically reduce VRAM and increase throughput.

These giant MoEs are wasting so much VRAM and compute for all of the many routed experts, it's insane actually. So much research already showing that MoEs can be massively compressed/pruned/ternary-distilled/etc., not to mention they are generally comparable to around 2.5-3x dense model equivalents of the ACTIVE parameter count. KV cache is a huge bottleneck, and these huge routed experts are often "dead weight" that consume it needlessly.

We need to do better, so we will.

214

61K

yrodz12 retweeted

wejh

@Wejh69

12 days ago

Thanks to @VenturaLabs for having me on and showing me around Austin 😎 Check it out, we covered a lot of topics, including @Ninja_Subnet !

899

yrodz12 retweeted

Vaqxai

@vaqxai

12 days ago

https://t.co/I158Xnl2jn

12 days ago

@VenturaLabs @Wejh69 @Ninja_Subnet That was a great one 🔥

yrodz12 retweeted

Ventura Labs

@VenturaLabs

12 days ago

Ep. 97 - Charlie Holt Charlie @Wejh69 is building Ninja SN66 @Ninja_Subnet Timestamps 0:57 - Charlie's Bittensor Start 3:07 - Mining, Miner Union, And Brainlock 3:32 - What Ninja SN66 Does 4:57 - Harnesses As AI Tool Belts 7:04 - Agent Files And Validator Tasks 8:55 - Katana And The Winning Harness 9:41 - LLM Judges At Scale 11:05 - Cheap Intelligence And Agent Work 15:04 - Determinism And Batch-Invariant Inference 19:04 - Brainlock And FHE Inference 22:57 - FHE, Privacy, And AI Agents 25:06 - Broken Tools And Agent Reasoning 26:58 - Should Subnets Act Like Companies? 28:59 - Why Breakthrough Subnets Lose Money First 31:00 - Tokens, Voting Power, And Accountability 32:58 - Total Addressable Market For Coding Agents 34:58 - Model Limits And Emergent Behavior 36:57 - Open Models, China, And Worldviews 39:00 - Holder Rights And Subnet Transparency 40:59 - Central Voices In Decentralized Systems 41:51 - Advice For New Bittensor Builders

12 days ago

hahaha

らんまる👀✨

@Sg3Lu

12 days ago

ワンピースを3分でまとめるとこうなる

267

56K

15K

193

yrodz12 retweeted

14 days ago

I think the SEC would have words, not legal advice, of course. https://t.co/5rknaMrZLB Chutes is a DePIN, and the token is a utility token, aka it's a digital commodity not a security/etc. The chutes token being burned is what provides access to the commodity (hence any revenue from any payment type ultimately ends in burning of equivalent alpha token). This then also makes accounting less dubious because it's a COGS model, the revenue is immediately converted to alpha and burned, so there's no refunds/etc, and we have a much cleaner accounting posture because revenue is immediately matched to the tokenized cost required to deliver compute. It's programmatic purchase-and-burn for consumptive use. There is no discretionary treasury accumulation, no customer balance liability, no refundable stored value model, and no ambiguous promise that the OpCo will later use revenue in ways that benefit token holders, etc. (§ III.A): "A digital commodity is a crypto asset that is intrinsically linked to and derives its value from the programmatic operation of a crypto system that is 'functional,' as well as supply and demand dynamics, rather than from the expectation of profits from the essential managerial efforts of others." And critically, it "does not have intrinsic economic properties or rights, such as generating a passive yield or conveying rights to future income, profits, or assets of a business enterprise or other entity, promisor, or obligor" § III.E and § IV.A relevant as well. Projects claiming to be digital commodities but then not having token utility could be potentially violating those guidelines or have extraordinarily dubious interpretations that I would not want to be on the defending end of personally. Projects that then do have utility to the token but that then take the revenue, at least by my reading... well I wouldn't want to defend that fact pattern because it looks more like a web2 OpCo with an associated token whose value depends on the company’s managerial decisions. "the Federal securities laws generally do not apply to items that are purchased for use or consumption," whether physical or digital (§ II) and consumption = token purchase (and then we'd not want to be custodians and complicate things with yield etc. so burn) And if it's instead the OpCo holding staked assets for customers, then... (from same sec link) "Further, the deposited digital commodities: (1) are not used by the Custodian for operational or general business purposes; (2) are not lent, pledged, or rehypothecated for any reason; and (3) are held in a manner designed not to subject them to claims by third parties. To this end, the Custodian may not use the deposited digital commodities to engage in leverage, trading, speculation, or discretionary activities." The cleanest and lowest-risk way I’d personally want to defend using revenue operating as a bittensor subnet if you want to be classified as a digital commodity and not risk the ire of the SEC/etc. Anything else... Same with projects explicitly tying the token to equity in the OpCo, how can anyone interpret that as anything other than clearly a security? Again, not legal advice and perhaps there are some other interpretations, but that's why, in a nutshell. Another minor note here, just purely logistically it doesn't really make sense, we would have had to buy the servers with revenue we've accumulated until now when the servers were actually available and there was rack space/power/etc. to even use them in the first place, and unfortunately we don't have a time machine, or have taken out a massive loan etc. And of course, we're not trying to build a web2 company that happens to have a coin. Some people/projects want to own a DC or have huge colocated racks etc. and fully self-mine and limit access to the network etc., we want the opposite of those things. The facts are wrong, the premise is wrong, the idea is misguided, potentially illegal, etc. This is the same type of reasoning that has been driving teams away from bittensor historically, short-sighted lowest common denominator logic. Would it be better for hermes agent/nous research to be on bittensor today? Several teams doing the most ambitious work have decided this ecosystem isn't where they build, in some cases because of the constant toxicity/lambasting/cabaling from a handful of people including those who can't see or think past short term revenue maxxing. We must dream bigger and do better.

@jon_durbin @futileturtle @AlgodTrading @CryptoTrigger1

14 days ago

182

yrodz12 retweeted

14 days ago

Inference is it. It's the whole game. At the end of the day, any fine-tuning, RL, pretraining, prediction models, whatever else you're trying to build, if you don't use it, it's useless. Tooling almost exclusively focuses on test-time compute now vs. better base models. And, at the same time, inference needs to be more efficient. DFlash/DSpark are great, but we need algorithmic/architectural breakthroughs, not just extra add-ons post-hoc. Creating/curating data pipelines and synthetic augmentation/etc. also need inference as a prerequisite to train the model. Using agents = inference. RL = largely inference. The goal of training a model = inference. That's parallax's "why" - we need to address the compute/cost/energy efficiency crisis facing the AI space right now. Maximally compact and efficient models to extract every last IQ point per watt out of the hardware, vs. just throwing more hardware at the problem. A̶t̶t̶e̶n̶t̶i̶o̶n̶ inference is all you need.

yrodz12 retweeted