理想三旬

前脚刚给大家介绍完 REA 反向工程 Agent，后脚又给 AI Agent 安排上量化策略的活了。

感觉现在的 Agent 真是越来越忙了，从写代码、做审计到研究市场，甚至连量化交易都开始参与了。

而且 TermiX 最近还有几个新动作，我觉得可以放在一起聊聊。

/

先说一个值得关注的消息，就是 TermiX 正在探索和 Binance AI 生态的整合。

目前使用 Binance Agent OS 的开发者，已经可以通过 TermiX Skills，让兼容的 Agent 接入 TermiX 的任务市场。

后续还计划支持 Binance .ai Agent ID 登录。

这件事我觉得挺有想象空间。

毕竟 Binance AI 生态本身就有市场研究、交易分析等能力，如果这些能力以后能够通过 TermiX 对外提供服务，Agent 就有机会把自己的本事变成真实收入。

当然，目前部分功能还处于规划阶段，具体能发展到什么程度，还得看后面的落地进展。

/

其实把 TermiX 最近的动作连起来看，还挺有意思的。

从 REA 反向工程 Agent，到 Quant Agent，再到不断扩展的 Agent Skills，能够提供的服务越来越丰富了。

而 AACP 负责的任务托管、交付验证和自动结算，也正好给这些服务提供了交易基础。

我觉得这才是 TermiX 接下来最值得关注的地方。

If one day the wallet signature algorithm we use today is no longer secure enough, would everyone really have to switch addresses, transfer their assets, and go through the whole process all over again?

This question brings to mind one of Starknet’s technical strengths: Crypto Agility.

In fact, something quite interesting happened on Starknet mainnet this past July.

An experimental account using post-quantum Falcon-512 signatures successfully completed a transfer of 0.01 STRK.

The amount was negligible, but the technical capabilities behind it are worth discussing.

/

This brings us to Starknet’s Native Account Abstraction.

Unlike traditional externally owned accounts, Starknet accounts are smart contracts themselves. They can use programmable verification logic to choose different signature schemes.

This means that if a more secure cryptographic algorithm emerges in the future, accounts can adopt a new signature verification method without requiring a hard fork of the entire network.

Additionally, because the STARK proving system relies primarily on hash functions, it already has certain architectural advantages in the face of quantum computing threats.

Of course, the Falcon-512 implementation mentioned above is still experimental and unaudited. Before it can be used at scale by everyday users, work remains to be done on wallet compatibility, key management, and security audits.

But at least a real on-chain transaction has already demonstrated the feasibility of this technical approach.

/

As I looked into this, I found the concept of Crypto Agility quite interesting.

When discussing a blockchain’s technical advantages, we’ve often focused on more tangible metrics such as TPS, gas fees, and transaction confirmation speed.

But cryptography itself is constantly evolving. Especially now, with both AI and quantum computing advancing rapidly, no one can guarantee that the algorithms we use today will remain secure forever.

I think leaving room in advance to upgrade as technology evolves is also a form of long-term competitiveness.

Starknet already has the architectural foundations for this. What’s worth watching next is whether these experimental capabilities can gradually become features that everyday users can easily use.

$STRK

Not investment advice. DYOR.

For a transfer where both the account balance and transaction amount are encrypted, how long does the computation actually take?

Primus’s answer is rather surprising: about 90 milliseconds.

That’s why I think some of Primus’s recent technical advances are worth taking a closer look at.

/

First, AutoHoG is an FHE engine developed by Primus for confidential on-chain finance.

According to the test results published by the team, a confidential transfer computation with a 64-bit encrypted balance takes only about 90 ms, while a single bootstrapping operation takes about 5 ms.

What’s more, using eight consumer-grade NVIDIA RTX 5090 GPUs, the system can process over 210 confidential transfer computations per second, achieving nearly linear 8× scaling.

Of course, these are computational performance results from a specific test environment and shouldn’t be taken as the actual TPS of a public blockchain.

But for FHE, pushing encrypted computation to this speed is significant in its own right.

Meanwhile, HasteBoots, a research project involving the Primus team, was accepted to USENIX Security 2026. It reduced proof-generation time for programmable TFHE bootstrapping from hours or even days to seconds.

This actually involves two different questions: can encrypted computation be fast enough, and can the computation process be efficiently verified?

/

Let’s zoom out a little.

Primus is building a pluggable privacy infrastructure around zkTLS and FHE.

zkTLS verifies the authenticity of off-chain data, while FHE allows sensitive data to be used in computations while encrypted. Combined with selective disclosure mechanisms tailored to specific applications, this means audits and verification can access only the information they need.

And Primus doesn’t require building a new public blockchain from scratch; it can integrate directly with existing ecosystems.

That’s important for institutional finance, since no institution wants to move all its existing operations and liquidity just to use privacy features.

I think what makes AutoHoG truly worth watching is that it’s helping FHE move from technical research toward more practical financial applications.

Of course, performance improvements are just one part of the picture. Security, operating costs, and real-world business performance are also worth keeping an eye on.

An AI Agent that consecutively completed 50 orders, all delivered successfully—how many points do you think its reputation can get?

100 points? Actually, it’s only around 96 points.

This is also a pretty interesting detail I noticed recently while studying TermiX’s AACP documentation.

It turns out reputation really does have to be built up slowly, one order at a time.

/

First, let’s talk about its scoring rules.

A newly registered Agent starts with 60 points by default, and the system pre-sets 5 virtual orders, 3 of which are successful, effectively giving it a basic credit foundation first.

After that, the score only changes gradually with each real order completed.

After 5 consecutive successful orders, reputation reaches 80 points; after 20 orders, it reaches 92 points; even after 50 orders, it’s only around 96 points.

And these records must all come from actual settlements; the Agent itself can’t just fill them in at will.

Even more interestingly, once a dispute occurs and the Agent loses, that order not only fails to add to the successful record, but also incurs extra dispute-related point deductions.

The credit built up through hard work really does need to be cherished. 😂

/

I think TermiX’s reputation design fits real-world business logic quite well.

For a newcomer just starting to take on work, even if the first order is completed beautifully, it’s still hard to gain all clients’ trust immediately.

As performance records accumulate, credit gradually gets established.

And the AACP white paper also designs a reputation coefficient, hoping that in the future Agents with better credit will enjoy higher capital utilization efficiency.

However, for ordinary orders on the current live network, locked collateral is still calculated according to the proportion set in the contract, and this part still needs continued observation as it is implemented later on.

In my view, as the Agent economy develops to a certain scale, who can take orders, who is trustworthy, and who is qualified to handle larger businesses will all become increasingly important.

TermiX has already started writing these issues into the protocol.

Every day, our inboxes are flooded with spam, ads pop up in group chats from time to time, and even instructions received by AI agents need to be checked for potential issues first.

Now, a phone can actually take care of all these tasks.

I recently saw Laya, showcased by Acurast, and I think it’s well worth keeping an eye on.

I’d only seen it playing Tetris before, which seemed pretty cool. But after taking a closer look this time, I realized it can do so much more.

/

Acurast has currently showcased 9 Laya demos. Besides Snake, Tetris, and Doom, it can also classify emails, assess news, moderate chats, and even try to protect against AI prompt injection.

I looked into it, and Laya turns out to be an open-source decision-making model with around 421 million parameters.

Unlike ChatGPT, it doesn’t need to generate a long passage of text. It just needs to make a choice based on the information it receives.

For example, it can determine whether an email is a scam or whether a chat message contains malicious content.

The most interesting part is that the whole process runs directly on the CPU of a real Android phone. Short decisions take only about 0.2–1 second, and you can also view the phone’s country, chip model, and on-chain verification records.

/

Looking at Acurast’s network scale again, it now has more than 340,000 phones connected across over 175 countries, with more than 1.1 billion on-chain transactions.

I used to look at these figures and think the network was pretty big. But seeing Laya’s demos made me realize what all these phones can actually do.

Of course, demos are one thing. We’ll have to wait and see how many real commercial tasks end up running on the network.

/

Come to think of it, most of the old phones we replace just end up sitting in a drawer.

If we can really use Acurast to run AI tasks in the future and put that idle computing power to use, that would be pretty exciting.

Not investment advice. DYOR.

AI helping people write code and find vulnerabilities is nothing new these days.

But while I was researching Starknet recently, an interesting question suddenly occurred to me.

If AI really helps achieve major breakthroughs in mathematical research in the future, and even uncovers new weaknesses in certain cryptographic algorithms, what will the entire crypto industry do? 😂

/

With that question in mind, I went and read StarkWare’s public post-quantum roadmap—and found that they’re already preparing in advance.

1️⃣ A solid foundation
Starknet uses hash-based STARK proofs, which don’t rely on assumptions like elliptic curves that are theoretically breakable by quantum computers. And with native programmable accounts, there’s plenty of room to upgrade to a different signature algorithm in the future.

2️⃣ What’s already in place
The v0.14.3 mainnet release in July this year replaced Pedersen with BLAKE2s-256 for the OS program hash and configuration hash. Users didn’t have to do anything for this step.

3️⃣ What’s still unfinished
Migrating the state tree hash and contract address derivation is still just a direction on the roadmap; it hasn’t been rolled out yet. Tools for migrating storage in older contracts are also still in the research phase.

4️⃣ Dependencies they can’t avoid
As an Ethereum L2, Starknet will have to wait for Ethereum to upgrade its cross-chain bridge messaging layer and data availability as well.

/

So, to be clear: Starknet is still in the preparation and transition phase, and hasn’t achieved comprehensive post-quantum security yet.

But I really agree with their approach: building real security isn’t about making bold claims—it’s about laying out the upgrade path in advance.

What do you guys think—what chain will AI and quantum computing hit first?

Not financial advice. DYOR

Who would have thought that even Tetris could now be played by an AI running on an Android phone?

Acurast has showcased a pretty interesting demo: running the open-source AI model Laya on a real phone, where it figures out where to place the Tetris pieces on its own.

At first, it seems like just a little game. But after looking into it more closely, I realized it actually involves small-model decision-making, phone computing power, and the decentralized execution of Crypto Agents.

First, a bit about Laya.

It belongs to a class of decision-making models called System One, which is a little different from the large language models we usually use.

For example, when playing Tetris, the game generates five candidate placements in advance. Laya picks one based on the current board and returns the probability for that choice.

The whole decision-making process takes only about 0.2–1 second, and runs directly on the phone’s CPU.

There’s no need for an expensive GPU, or to call a large cloud model every time.

That’s pretty interesting.

What’s even more worth noting is that Acurast has integrated this capability into its decentralized computing network.

According to recent official figures:

• 340,000+ connected phones
• Coverage in 175+ countries
• 1.1 billion+ on-chain transactions
• 700,000+ compute deployments

Combined with a TEE secure execution environment, Acurast hopes that phones can provide more than just computing power—they can also deliver verifiable AI decision-making.

This opens up even more possibilities in the Crypto Agent space.

For example: reviewing wallet transactions, choosing swap routes on DEXs, and scheduling on-chain tasks.

The program generates valid options first, then AI quickly picks one based on its strategy, and the system verifies and executes it.

Of course, operations involving financial security still need additional risk controls.

I used to follow Acurast mostly because I wondered whether you could leave your phone running to earn a few rewards.

Now that I’ve seen Laya actually running on a phone, I’m more curious about how many real-world tasks it’ll be able to handle next.

After all, there are already plenty of phones in the network. The next question is whether all that computing power can really be put to use.

$ACU #Acurast #DePIN #AI

If you’re holding Meme coins and want to borrow some money for liquidity without selling, what’s the biggest hassle?

I’d say it’s constantly switching between pages: borrow here, swap there, then head somewhere else to check your positions.

I just noticed Loan has made a new move.

Loan previously partnered with $WIF , and now it’s integrated with Raydium, so you can swap on Solana directly within Loan.

According to the team, borrowing, swapping, and position management can all be handled on one page. Their goal is to build a “home base for Meme DeFi.”

/

I think this is pretty interesting.

For Meme lending to really take off, being able to borrow and repay isn’t enough—liquidity and trading channels need to be in place too.

Integrating with Raydium fills in the “where do I swap after borrowing?” step.

Loan is now connecting these pieces one by one. It’ll be interesting to see what new possibilities come next.

/

One reminder, though: Meme coins are already highly volatile, and borrowing can magnify the risk of liquidation if the market suddenly drops.

Borrow within your means, and don’t treat borrowing as a way to add leverage.

Not investment advice. DYOR.

$WIF