Entrepreneur founding AREO | Building AI-native Company | VP of Marketing &CEO of oversea markets @OKX(2016-2022) | Crypto一级投资Loser,不投只孵化
Joined August 2017
- Tweets761
- Following1.3K
- Followers347
- Likes144
CryptoJames retweeted
🧠 Bittensor $TAO upgrade: Root Reborn is live
Bittensor’s root validator layer has changed.
Instead of root dividends remaining in the previous claim system, each validator now directs them into a validator-curated beta basket — a compounding fund containing subnet alpha and, optionally, root TAO.
Here’s how it works:
1️⃣ Validators set the allocation
Each root validator chooses how its basket is distributed across the subnet economy through its Root weight vector.
2️⃣ Dividends are reinvested
Root dividends are converted into TAO and redeployed according to the validator’s allocation. The resulting positions are held in escrow, staked, and able to compound over time.
3️⃣ Stakers accrue basket shares
Root stakers earn shares in their validator’s basket based on their root stake. New shares are minted at the basket’s current NAV, preventing later stakers from diluting previously accumulated value.
4️⃣ Claims redeem the full basket
When a staker claims, their proportional share of every basket holding is redeemed into TAO and returned to Root.
5️⃣ Validator strategy becomes transparent
New on-chain views expose basket NAV, holdings, validator allocation weights, and the TAO owed to individual stakers.
The main change is that root validators are no longer differentiated only by fees, stake, and headline yield.
Their subnet allocation strategy — and the performance of the basket built around it — now matters too.
Root Reborn connects root staking more directly to the subnet economy and turns validators into active on-chain capital allocators.
Explore Bittensor validators → tao.bot/validators
CryptoJames retweeted
为什么美国科技公司要为开源模型（中国开源模型占了绝大多数）争取发展空间？昨晚最大的新闻无疑就是老黄@JensenHuang 首开X账号，第一条推文就是一份由a16z发起、英伟达、微软联合PLTR、PerplexityAI、Ollama等多家科技公司共同签署的公开信，明确主张"前沿开源模型与前沿闭源模型应并存发展"。看到都在聊公开信本身，但是对背后的动机和意图聊的不多。个人角度，核心原因就是之前下面这张图和引用的之前长文里聊到的内容：
1、美国企业使用的中国开源模型每周token占比正逼近60%，开源模型在美国企业token消耗里占比这么大，一旦被禁或者限制、那么企业的成本剧增。
2、7月中的这篇长文里（引用的长文就是）聊到：企业AI采用进入工程化时代，AI在企业级市场的采用越来越变成分场景应用模型阶段：
1）高价值场景使用继续使用闭源前沿大模型的高价值Token；
2）高频、低复杂度、高体积的任务场景下，企业越来越偏向高性比价比Token（基本上来自开源模型）。
3、这个趋势不可避免：
对于芯片厂商来说、用谁的模型都要训练和推理；
对于云厂商来说，给企业客户模型选择越多，毛利越高；
对于部署层（如pltr）或者应用型公司来说，更是天然有选择高性价比token的动力。
这份公开信的联名企业基本上就是这三大类，所以公开信背后其实有着极强的商业诉求和巨大的商业价值。
4、Authropic在国会呼吁限制或者禁止中国开源模型的发展，实质上是明显感受到了这股势头的威胁和影响。kimi k3只是一个催化剂，在k3之前中国开源在Token消耗量的份额占比已经在快速飙升了，即使没有k3，Authropic也有动力这么做。
本条由@bitget_zh赞助，「Bitget 买美股：秒级入场，丝滑交易 」
CryptoJames retweeted
I think Conviction is going to be a massive net positive for the ecosystem
it won't prevent subnet owners from selling to cover expenses, it'll just require them to be more intentional and transparent about it.
Teams can run programmatic unlock schedules that will get priced in by the market since they'll be predictable.
This is basically a push toward professional treasury management for teams
One of the largest Bittensor upgrades, called Conviction, is slated to arrive as early as next week.
The overarching goal of the whole Conviction Upgrade is to add a layer of governance at the subnet layer and provide investor protections that currently don't exist natively within the protocol.
The Conviction Upgrade in a nutshell:
• Subnet owners can choose to lock their token holdings to signal their "Conviction" to the subnet. All of their ongoing/future subnet owner emissions are, by default, auto-locked.
• Subnet owners can unlock tokens, which is an onchain action, at any point. But the unlock has a delay function applied to it. After ~30 days, 63% of the unlocked tokens will be spendable, 95% after ~90 days. The delay means subnet owners can't unload their entire holdings without giving an onchain alert to current holders that the owner is exiting.
• Any token holder can lock their tokens toward any address. This could be the current subnet owner or a potential new subnet owner.
As it currently stands, the initial Conviction Upgrade only implements the locking mechanism. There won't be the ability to "elect" new subnet owners yet.
This is a big economic change and does present risks to existing/future subnet owners, so I expect there to be enthusiastic debate from both sides.
CryptoJames retweeted
如果大家对 Obsidian 的基础使用上有疑问
欢迎看这个博主：Blink
制作精良画面清晰，讲解透彻
值得关注：
youtube.com/@blinkq3370?si=I…
CryptoJames retweeted
资产链上化时代，也是加密黑洞开启的时代，它会虹吸并吃掉传统金融的一切。
资产上链是这个黑洞的事件视界。一旦现实世界资产（如国债/黄金/股票/私募信贷/房地产/基金份额……）被搬到链上，就变成了可编程、可组合、24/7 全球流动的数字资产。
流动性、透明度和效率会逐渐赶上并超过传统金融：
• 稳定币是第一个被吸进去的“庞然大物”（链上稳定币市值已接近/突破 3000–3200 亿美元）；
• 链上黄金、Tokenized Treasuries 和链上股票/权益类资产都在快速增长中；
• 贝莱德、富兰克林等传统巨头早已排队上链，BUIDL 等产品领先市场；
• 香港、美国、新加坡、欧盟的监管正在加速这个过程。有了牌照+清晰规则，机构才敢大规模进来。
传统金融曾经的壁垒（牌照、时区、中间商、纸质合同），如今正在变成劣势。
就像当年互联网把传统媒体、零售、银行业务一口吞掉一样，这次轮到整个金融基础设施。
加密黑洞已经打开，吸力只会越来越强。
在这个过程中，谁会是获益者？
CryptoJames retweeted
AIsa is proud to be one of the first data providers integrating Circle nanopayments — and a tech partner for the Agentic Economy on the @arc hackathon.
This is the shift:
AI agents don’t want API keys, plans, or dashboards. They want to act — and pay as they go.
So we built for that.
An agent calls AIsa → pays $0.002 in USDC → gets real-time data → moves on.
No subscriptions. No gas. No friction.
We’ve already processed 1M+ calls on traditional API infra (keys, quotas, rate limits).
It works — but it’s not agent-native.
Nanopayments by @circle change the primitive:
From “pay to access” → “pay per action”
Same 100+ endpoints. Same infra.
New layer: programmable, per-call USDC payments.
This is what the agentic economy runs on: transactions, not SaaS.
Hackathon starts April 20.
lablab.ai/ai-hackathons/nano…
If you’re building agents, build systems that can pay, decide, and act end-to-end. 🚀
ALT Arc hackathon
CryptoJames retweeted
我看很多炒合约亏钱的散户都在骂项目方或者做市商，因为有些币拉砸的K线实在是匪夷所思，放传统金融市场可能要进去一堆人了。
但我想他们骂错了人，或者说没意识到谁是最该骂的。 想一下这种暴涨暴跌的K线是什么时候成为行业常态的？ 又为什么会这样？
币安开始通过alpha大量上币之后，不允许项目方按照协议TGE给VC发币，导致几乎所有币都有高控盘的阶段，使得K线操纵的成本非常低。
也就是从那个时候开始，CEX彻底赌场化了，涨跌跟基本面不再相关，价值投资成了笑话。
更重要的，行业从业者的荣誉感降到最低。KOL觉得自己是叠码仔，VC纷纷转行，对投资web3羞于启齿，散户觉得自己是赌客.
CryptoJames retweeted
Quasar mission is now twice as hard, and that’s what makes us special
We’re not just solving the hardest problems in AI, like memory, but also tackling MoE decentralized training.
And the funny part is we have a path for both we’ll utilize every bit of power from Bittensor
CryptoJames retweeted
Yes emissions are used to bootstrap innovation, same as Uber, Amazon and countless of other big companies
You can chose between these 2:
-give those emissions to VC’s
-give those emissions to builders who devote their whole time to build out the network
Vc’s hate it because they can’t apply the VC playbook/had discounted access compared to the masses
CryptoJames retweeted
Building a multi-agent swarm this weekend shouldn't involve fighting five different provider SDKs.
Every time you want to test a new model or give your agents a new skill, you end up integrating a new API key and refreshing another billing dashboard.
Instead, why not let AIsa handle the entire routing layer for you. You integrate our unified endpoint once, and we natively route your agents to the best tools available:
✅ Top-tier AI models (Claude, GPT, Kimi, Qwen, Minimax)
✅ AIsa skills (live market data, code reviewer)
✅ Native execution APIs (financial, search, video, x/twitter)
your custom logic stays perfectly intact and you're also able to track every millisecond of compute across all providers in one clean dashboard.
So stop doing infrastructure maintenance on a saturday and start shipping faster and cheaper with AIsa!
docs.aisa.one/docs/welcome-t…
CryptoJames retweeted
What if you could tell an AI to improve itself overnight and wake up to a better model in the morning?
@Hevalon built exactly that on Basilica. No babysitting. No manual tuning. Just one command.
Full writeup: templarresearch.substack.com…
CryptoJames retweeted
Advancing confidential computing for a more secure AI future.
Together with @manifoldlabs, we’re exploring how Intel TDX and Intel Trust Authority help enable confidential workloads across decentralized infrastructure, including @TargonCompute's Targon Cloud platform—protecting data at rest, in transit, and in use.
CryptoJames retweeted
Quasar kernels are way faster much faster!
This is Quasar Attention, the mechanism behind the upcoming Quasar models, designed to support context lengths of up to 5 million tokens.
Attention has long been a bottleneck for processing extended context. Standard attention mechanisms struggle to scale beyond ~200k tokens in training, creating a ceiling on how much information models can reliably use.
One approach to solving this has been linear attention methods, such as gated delta attention (used in Qwen 3.5) or Kimi delta attention. These improve efficiency and allow longer sequences, but introduce trade-offs: instability at extreme lengths, quality degradation, and in practice, they are not strictly linear.
Quasar Attention takes a different approach. It uses a continuous-time formulation, implemented as a fully matrix-based system rather than relying on vector-state approximations. In practice, this improves stability, reduces cost, and maintains performance as sequence length increases.
In internal stress tests at 50 million tokens, KDA-based approaches begin to lose stability, while Quasar Attention remains stable. This allows performance to hold as sequence length increases, rather than degrading beyond a fixed threshold.
On BABILong, a Quasar-based model pretrained on 20B tokens and fine-tuned on 16k sequences was evaluated on contexts ranging from 1 million to 10 million tokens, maintaining consistent performance across that range. By contrast, models using gated delta attention show significant degradation at longer lengths, in some cases dropping to ~10% performance at 10 million tokens. (Note: results are indicative; setups are not directly comparable)
On RULER benchmarks, a Quasar-10B model (built on Qwen 3.5 with frozen base weights and Quasar Attention added), pretrained on 200B tokens, achieved 87% at 1 million tokens, outperforming significantly larger baselines, including Qwen3 80B, under the same evaluation conditions.
Taken together, this points to a shift in where long-context performance is won or lost: not in model size alone, but in the attention mechanism itself.
Quasar Attention represents a step change in long-context modelling, setting a new standard for stability and performance at scale.
We thank @TargonCompute for the compute and for being our compute provider and long-term partner in training the upcoming Quasar models
Here is the link to our paper 👇
