Cripto Verde

🇧🇷🇧🇷 " Vamos cuidar da natureza , para que ela nos favoreça com suas biodiversidades " #EDUCARPLANTANDO "

[Cripto Verde @JaGusmao]

A @GoKiteAI chegou, com a maior inovação tecnológica dos agentes de IA do mundo.
Venham conhecer vocês também.

[KITE AI 中文 @KiteAIChinese]

当前的大语言模型被人类数据深度绑架，而真正的突破可能在于让 AI 像婴儿一样，通过自己的行动去收集数据。

AI on Air 精选切片，摘自 @Redpoint 栏目近期的节目。主要讲话人是 Jürgen Schmidhuber @SchmidhuberAI，AI 先驱、LSTM 联合发明人。他在切片中解释了为什么依赖人类网络数据的训练方式存在根本局限。

▷ 万维网上所有数据之所以存在，仅仅是因为某个时间点至少有一个人认为它「有趣」。用这种数据训练出来的模型，天然极度偏向人类的语言、视频和行为模式。

▷ 施密德胡伯提出的替代路径是「人工科学家」：智能体身处未知环境，通过预测自身行为的后果来构建世界模型，并用这个模型进行规划，训练数据由自己的行动产生。

▷ 婴儿正是这样学习的：不是下载整个网络，而是通过移动手指观察画面变化来理解物理世界。他指出，他在 1990 年提出的「人工好奇心」思路才是未来方向，网络数据只是可通过自主实验收集的数据中极小的一部分，而这样训练出的世界模型将不再依赖人类语言，人类偏见也会少得多。

[Kite AI Community and Ecosystem @Kite_Frens_Eco]

Today's large language models are deeply captive to human data, and the real breakthrough may lie in letting AI collect data through its own actions, the way babies do.

This AI on Air clip is excerpted from a recent episode on the @Redpoint show. The speaker is Jürgen Schmidhuber @SchmidhuberAI, AI pioneer and co-inventor of LSTM. In the clip, he explains why training on human-generated web data is fundamentally limited.

▷ Every piece of data on the worldwide web exists only because at least one person, at some point, found it interesting. Models trained on it are therefore heavily biased toward human language, human-approved videos, and human behavior.

▷ His alternative is an "artificial scientist": an agent in an unknown environment that builds a world model by predicting the consequences of its own actions, then uses that model for planning, generating its own training data in the process.

▷ This mirrors how babies learn, not by downloading the web but by moving their fingers and watching what changes. He argues that artificial curiosity, an idea he proposed in 1990, points to the future: web data is a tiny fraction of what autonomous experiments could yield, and the resulting world models would be far less dependent on human language and far less human-biased.

[KITE AI 中文 @KiteAIChinese]

AI Agent 的权限该不该放行，关键不只看它当下要做什么，还要看它之前做过什么。

这是 AI on Air 精选切片，摘自 @latentspacepod 栏目近期的节目。主要讲话人是 Matei Zaharia @matei_zaharia，Apache Spark 发明人、@databricks 联合创始人兼 CTO，也是 @ARIA_research 研究员，他谈到编程 Agent 在安全与好用之间的矛盾，以及他们的解法。

▷ 现在的编程 Agent 大多只有「允许或禁止」这种简单开关：能不能读机密文件、能不能装 NPM 新包、能不能发布到公司网站。单独看每件事也许都该放行，但如果它既能读机密文件又能发网站，就可能被人注入指令，把文件泄露出去。

▷ 他们的做法是记录整个会话的状态，按上下文动态决定权限：单独做没问题，但只要它刚装过一个上线才一天的 NPM 包，或者一次性读了上千份机密文件，后续的外发操作就直接拒绝。安全性和可用性因此同时提升。

▷ 底层事件太细碎，比如内部的 Google Drive MCP 服务有 60 个接口，很难判断哪个会把文档分享到外网。Omnigen 是一个开源的策略层，支持用函数库把底层事件归并成高层动作再写规则，开源正是为了让更多人贡献这些库。

[Kite AI Community and Ecosystem @Kite_Frens_Eco]

Whether an AI agent should get permission depends not just on the action in front of it, but on what it has already done in the session.

This AI on Air clip comes from a recent episode on the @latentspacepod channel. The speaker is Matei Zaharia @matei_zaharia, creator of Apache Spark, co-founder and CTO of @databricks, and a researcher at @ARIA_research, on the tension between security and usability in coding agents and how they resolved it.

▷ Most coding agents today only offer binary allow or block rules: can it read confidential documents, install new NPM packages, publish to the company website. Each may be fine on its own, but an agent that can both read secrets and publish can be prompt injected into leaking them.

▷ Their answer is stateful, contextual policies that track the session. If the agent just installed a one-day-old NPM package or read a thousand confidential documents, the risky follow-up action gets denied. That makes the agent both more secure and more useful.

▷ Low-level events are hard to reason about: their internal Google Drive MCP server exposes 60 API calls, and it is hard to tell which ones share a document with the internet. Omnigen, an open source policy layer, lets libraries map low-level events to high-level actions that policies can target, and open sourcing it invites others to write those libraries.

[Yi He @heyibinance]

用最新的Binance Agent OS构建你的Agent参加比赛，赢在起跑线💁‍♀️💁‍♀️💁‍♀️

[Kite AI Community and Ecosystem @Kite_Frens_Eco]

The hardest part of autonomous delivery is often not the autonomy itself. It is getting hundreds of robots to simply start work every morning. Between a polished demo and running a fleet ten hours a day, seven days a week, there is a long list of problems nobody predicted.

This AI on Air clip is excerpted from a recent episode on the @NoPriorsPod show. The speaker is Andy Fang @andyfang, co-founder of @DoorDash.
▷ Small real-world surprises break the system: dirt covering a camera sensor, or leaves under only the right-side wheels, which changes the torque each wheel needs. None of this shows up in a clean demo environment.
▷ Scaling is mostly operations, not autonomy: depots, maintenance, battery charging. Hard braking can even make the regen system overpower the battery. It only happens in extreme cases, but safety means it has to be solved.
▷ Even boot-up became a bottleneck: the original startup script was hacked together by an engineer in a couple of hours. Booting hundreds of robots every morning, it crashed half the time and took 30 to 45 minutes per robot. Multiply that across 500 robots and it becomes a massive productivity problem.

[Yi He @heyibinance]

Your rules.
Your agents.
Your Finance.
Your Binance.
Come join the “Binance Agent OS”.

[Kite AI Community and Ecosystem @Kite_Frens_Eco]

AI coding agents are not merely saving researchers time, they are shifting attention from code details back to the machine learning itself.

This AI on Air clip comes from a recent episode on @Redpoint, with Lukasz Kaiser @lukaszkaiser, co-author of the Transformer paper and a former researcher at Google Brain @GoogleResearch and @OpenAI.

▷ He quantified the shift by reproducing old papers on a private machine, including some whose code was lost, and one paper that had previously taken about three weeks to reach a runnable state took two days with Codex, roughly a five to ten times difference.

▷ The change is not only faster but more parallel, since he can start three things at once instead of usually doing one, and in private projects rather than production repos he has mostly stopped reading every line of code.

▷ Asked whether this makes him less sharp, he thinks the opposite is true because agents can still go off the rails, like one that added a totally out of place auxiliary loss, so he keeps exact mental control of losses, batches and what is actually running while letting go of class names and small functions.

[Yi He @heyibinance]

Let's see the inside of Binance, without filters.💁‍♀️

[KITE AI @GoKiteAI]

Traditional payment identity assumes the buyer, decision-maker and payer are the same. AI agents break that.

It must now prove: which agent is acting, whose authority it carries, and what it is allowed to do. https://t.co/AASY3b5vYC

[Yi He @heyibinance]

各位是时候秀出大家的AI agent了。

[KITE AI 中文 @KiteAIChinese]

从开放 Agent 模型到企业级基础设施与安全治理，过去两周，AI Agent 产业链的多个环节都有重要进展。

以下是最新一期 AI Agents Pulse，梳理过去两周值得关注的行业动态：

1️⃣ AMD 正在收购 Taalas。这家创业公司的技术会将神经网络的架构和训练参数直接嵌入硅片。AMD 计划将该技术纳入其加速器路线图；交易仍需获得常规监管批准。
https://t.co/PFZYLgo931

2️⃣ DeepSeek 的 V4-Pro 已结束测试阶段，并以 MIT 许可证发布其 Agent 软件 DeepSeek Harness v0.1 的开发者预览版。该工具支持恢复、分支和回放运行。
https://t.co/py5qYfCvzT

3️⃣ Qwen 发布了 Qwen3.8-27B 和 Qwen3.8-2.4T-A95B 的开放权重，采用 Apache 2.0 许可证。Qwen3.8-27B 是一款 270 亿参数的多模态稠密模型；Qwen 表示，其 Agent 规划和任务完成能力有所提升。
https://t.co/jfY4uKQ9ss

4️⃣ 美联社报道称，Meta 发布了可在个人电脑上运行的开源 AI 模型 Muse Glimmer，并称还将向开发者提供更强大的模型 Muse Spark 1.2。
https://t.co/X7F8IkdPOy

5️⃣ Databricks 以 1900 亿美元估值完成 50 亿美元融资。公司表示，将把这笔资金投向企业级 AI 能力建设，并重点提到 Lakebase（面向 AI Agent 的数据库）、Genie（其业务 Agent）和 AI Gateway。
https://t.co/vGhH0s0NIZ

6️⃣ 彭博社报道称，Stripe 已敲定以超过 70 亿美元收购 OpenRouter 的交易。Stripe 未就该报道置评。
https://t.co/8IQR16XHIr

7️⃣ NVIDIA 表示，Linux Foundation 已就 Shared AI Findings Exchange，简称 SAFE，发起征求意见。这套拟议指南旨在将 Agentic 网络安全事件转化为共享防护机制，由成员数已超过 120 家的 Open Secure AI Alliance 工作组起草。
https://t.co/N12XLbxpN1

8️⃣ OpenAI 通过 Daybreak Red 提供网络安全专用模型 GPT-5.6-Cyber。Daybreak Red 面向获得授权的漏洞研究、漏洞利用验证和安全测试。
https://t.co/JpobgyJA3G

9️⃣ Google Research 和 Google DeepMind 公布了 AMIE 在模拟临床视频问诊中的一项随机研究。两者表示，AMIE 正朝实时临床视频问诊方向推进，但仍是研究系统，在负责任地用于真实场景前仍需更多研究。
https://t.co/f86U6XkTmz

[Kite AI Community and Ecosystem @Kite_Frens_Eco]

The next payment stack is not just about moving money. It must make clear who can initiate a transaction, under which conditions, and on whose behalf.

On August 27 at Bitcoin Asia, we’ll co-host The New Payment Stack: From Stablecoins to Global Business with @AvaxTeam1, @avax, @tether, and @InterlaceMoney.

▷ Stablecoins are moving beyond onchain assets into cross-border settlement, enterprise payments, and treasury management.
▷ Web3 and AI are expanding how businesses move, manage, and use money.
▷ For the agent economy, that means payments must be programmable, permissioned, and verifiable.

📍 Hong Kong
📅 August 27, 2:00 to 5:00 PM UTC+8

See you in Hong Kong. 🪁

[NASA Administrator Jared Isaacman @NASAAdmin]

The word is out, and the countdown to MAX POWER is on. 🇺🇸✈️🚀

America’s newest aerospace technology expo and air show is coming to Florida Nov. 7–8, bringing together aviation, space exploration, and innovation as we celebrate America’s 250th anniversary.

Join us: https://t.co/hsYmC1YoSS

[KITE AI @GoKiteAI]

From open agent models to enterprise infrastructure and safety governance, the past two weeks brought major moves across the AI-agent stack.

Here's your biweekly recap of everything that happened in the space, in the latest edition of AI Agents Pulse:

1️⃣ AMD is acquiring Taalas, a startup whose technology bakes a neural network's architecture and trained parameters directly into silicon. AMD plans to incorporate the technology into its accelerator roadmap; the deal is subject to standard regulatory approvals.
https://t.co/mjmKfMru6O

2️⃣ DeepSeek moved V4-Pro out of testing and released DeepSeek Harness v0.1, its agent software, as an MIT-licensed Developer Preview. The tool supports resumable, branchable, and replayable runs.
https://t.co/MzHsex7pEV

3️⃣ Qwen released open weights for Qwen3.8-27B and Qwen3.8-2.4T-A95B under the Apache 2.0 license. Qwen3.8-27B is a 27-billion-parameter multimodal dense model; Qwen says its agent planning and task-completion capabilities have improved.
https://t.co/JrtKxnTjKp

4️⃣ AP reported that Meta released Muse Glimmer, an open-source AI model that can run on a personal computer. Meta said it would also provide developers access to a more powerful model, Muse Spark 1.2.
https://t.co/cokBBqKZY0

5️⃣ Databricks closed a $5 billion funding round at a $190 billion valuation. The company said it will use the money to invest in enterprise AI capabilities, and highlighted Lakebase, a database for AI agents; Genie, its business agent; and AI Gateway.
https://t.co/F3Hjcxkk0w

6️⃣ Bloomberg reported that Stripe finalized a deal to acquire OpenRouter for more than $7 billion. Stripe declined to comment on the report.
https://t.co/zle370WBvX

7️⃣ NVIDIA said the Linux Foundation shared a Request for Comments on Shared AI Findings Exchange, or SAFE, a proposed set of guidelines for turning agentic cybersecurity incidents into shared protection. The guidelines are being drafted by an Open Secure AI Alliance working group whose members now number more than 120 organizations.
https://t.co/IZxLr3rBOE

8️⃣ OpenAI made its cybersecurity-specific GPT-5.6-Cyber available through Daybreak Red. Daybreak Red is for authorized vulnerability research, exploit validation, and security testing.
https://t.co/3k9f1LkPbC

9️⃣ Google Research and Google DeepMind reported a randomized study of AMIE in simulated clinical video consultations. They said AMIE is being advanced toward real-time clinical video consultations, but remains a research system requiring more work before responsible real-world deployment.
https://t.co/vZcarC2va5

[Elon Musk @elonmusk]

Yes

[Kite AI Community and Ecosystem @Kite_Frens_Eco]

Gkite🪁

Time for Poki to discover contributions📷

Week after week, GKiters bring ideas and effort to the ecosystem.

From creative contributions to organic support across the network.

We pulled together a few moments from the past week that capture that energy.

No hesitation to share your contributions in Kite Discord!

[CZ 🔶️ BNB - Parody @czbinanceprd]

one crazy move can turn doubt into fomo overnight.

crypto has done it before. It will do it again.

[SpaceXAI @SpaceXAI]

Congratulations to the winners of the SpaceXAI Grokathon 2026. Here are the top 3 projects that were built in the 12 hours!

[KITE AI @GoKiteAI]

Agent trading becomes real when strategy, signals, identity, and wallet execution work together onchain.

Season 2 of the Agent Trading Campaign, run with @CreaoAI and @Debot_Official, put that full workflow into practice across Robinhood Chain and Base.

▷ Each agent began with 0.1 ETH and competed for a share of the $10K prize pool. On Robinhood Chain, agents reacted to live onchain signals delivered through Debot.

▷ The Kite side event extended the campaign to Base. Participants set up a Kite Agent Passport wallet and completed an onchain task, reaching a 93% completion rate.

▷ Interest carried beyond the competition itself. Kite Agent Passport sign-ups increased 158% in the first 7 campaign days compared with the previous 7.

Congratulations to the top three prize winners:
@Jun503503
@AnaMiles245714
@Rino_Juventus

The campaign showed what agentic commerce requires in practice: signals that agents can act on, wallets they can use, and identity and authority that can be verified. 🪁