What keeps striking me is this: things I tried and failed to do last year now work surprisingly easily with AI.

1.

People who could not build websites are now shipping full-stack apps.

2.

One engineering first principle used to be: start with Python because it lets you iterate quickly. That logic is starting to change. When AI can implement most of the code, the human attention cost across languages starts to converge. At that point, Python can become the expensive option: similar human cost, but higher runtime and memory cost.

That is why more people are building even small, single-user tools in Go or Rust. When you are making 100 projects instead of 1, efficiency matters more. And now, getting that efficiency is often just one prompt away.

For example, I’m moving many of my personal projects to Go. Some now use less than one-tenth of the RAM they used before, which lets me run roughly 10 times as many programs at the same cost.

A team of five 19-year-olds built a 30 petabyte storage cluster in San Francisco for about $500,000 to create a 40x cheaper alternative to AWS S3, as part of a project to store 90 million hours of video. The article highlights a startup's innovative engineering and a foundation model that learns from video like humans, built by Standard Intelligence.

https://si.inc/posts/the-heap/

https://x.com/deedydas/status/2029391960159740003

The example above is the new normal now.

3.

I know almost nothing about fine-tuning or training AI models, but I’m now “vibe-training” lora adapter for Qwen-3 on RunPod. Codex (GPT-5.4) writes the scraping and training code, provisions GPUs, monitors jobs, shuts them down when training finishes, investigates failures, and fixes the code by itself. In a real sense, it is helping me train my own model. Code is becoming extremely cheap. Model customization is getting cheaper too, although training still requires GPU access.

Research is now entirely the domain of autonomous swarms of AI agents running across compute cluster megastructures in the skies. The agents claim that we are now in the 10,205th generation of the code base, in any case no one could tell if that's right or wrong as the "code" is now a self-modifying binary that has grown beyond human comprehension.

https://github.com/karpathy/autoresearch

4.

Another thing I never imagined I could do is rename all the photos and images on my MacBook—most of them screenshots from Chrome or Terminal—by having Claude read them. That makes searching those images much easier.

This was not really practical last year for three reasons: vision models were weaker, token costs were higher, and the harnesses were worse. Now, better models, cheaper tokens, and tools like Claude Code or Codex make it possible for AI to understand images, browse, and act directly in a desktop environment.

We can do fundamentally bigger things now. You need to rewire your brain to fully harness the agents.

It will only get better from this

AI is currently driving significant innovation in the field of coding. Companies including Anthropic and Cursor have proven that AI can dramatically accelerate productivity—a trend referred to as “vibe-coding.” This progress is making software significantly affordable.

Mental Model

My guiding principle for AI coding agents is simple: the speed of implementation should match the speed of writing down the idea. Physically, you can’t go faster than that. The idea must be the only bottleneck.

tg-claude

tg-claude is a self-hosted Telegram client for Claude Code running on the Cloud. You can start out a project by sending a message. It outputs programs in the form of github repositories. You can keep iterating the development by literally sending messages to the telegram bot.

mothership

mothership is an SDK (supports Python and TypeScript) that provides the infrastructure layer for building bots. It abstracts common utilities—Telegram bot client, database, dashboard, loggers—so you can focus on bot logic. The CLI handles orchestration: start, stop, and manage multiple bots from the command line. tg-claude is built on mothership as a dependency module.

Roadmap

AT-AT Walker Phase (Current Stage):

AI agents including tg-claude use mothership to build software through natural conversation, deployed on cloud infrastructure with minimal friction.

Millenium Falcon phase (Coming Soon):

Introduces Mother Wallet: A crypto-native treasury that grants economic autonomy to bots. Each bot is allocated a stipend for self-development, debugging, and testing. I found it extremely useful for building crypto trading bots that need real capital to validate strategies. For broader applications, the wallet will utilize the x402 protocol to compensate other agents for task completion. This system enables bots to autonomously hire specialized agents for computation, purchase data, and pay for API access.

Open Source

Both projects are fully open source:

• mothership: https://github.com/guzus/mothership

• tg-claude: https://github.com/guzus/tg-claude

Instantly, I decided to investigate whether the searcher from mainnet replicated the same behavior on the ETHW. For those unfamiliar with it, ETHW is a hard fork of Ethereum since The Merge (September 2022). Because the affected contract was deployed before The Merge, there was a chance that copying the transaction could be profitable to anyone.

I checked https://www.oklink.com/en/ethw which is the only ethw explorer I know, but there wasn’t any same MEV transaction yet. “Eureka! 22 ethw is about $81, that’s small but fun money.”

If you spot the opportunity, the execution is quite straightforward. Begin by searching ETHW on Chainlist to obtain ETHW RPC. Initialize your hardhat project and write a script that sends the same transactions as the original one. I used my ethereum private key, because I had some ETH on ethereum before the Merge, and therefore the same amount of ETHW airdropped into my account, providing sufficient funds to cover the gas fee. And boom! Send it to the network. Typically, there are two processes (transactions) in general: 1. Deploying a contract for his strategy, 2. Executing the contract.

It was successful. I withdrew ETHW to my Kucoin account. It was a nice, gratifying trip. Several months later, I succeeded at replaying the another exploit on ethereum, which brought me about $60.

Key Takeaways

1. I knew nothing about “reflective token” or some kind of stuff. I just simulated the same transaction with a bit changed input, and propagate the transaction.

2. When you exploit on Ethereum, consider exploiting ETH forks right after it if it’s worth it. For the record, “Ethereum Alarm Clock” exploit guy did.

Why I’m sharing this alpha

1. I think there isn’t much return in this strategy.

2. I’m sharing this to web3 newbie devs to foster them so crypto can be more crowded.

3. I want more twitter followers.

I’m going to write some posts. Judge my plans and decide whether to subscribe me or not.

My interests are mainly on crypto, finance, and system.

My Post Plan:

• Gasless Tx in ethereum

• Something about Dune analytics

• Random Generator in ethereum

• Data availability problem in ethereum

• Review of the Game theory course from Stanford

• Review of the book “Engineering at Google”

• How 1.1.1.1 make their profits

• How 1inch network make their profits

• Code Review of GMX

• The implementations of Future and Option Markets in DeFi

Any suggestions are welcomed, Thanks.

Subscribe now