Sam Town

Full-Time Happen Maker | Part-Time Equestrian

[Sam Town @UXTown]

Computer simulate a playable maid frog making dressed herring in a Russian old world apartment no mistakes

@odysseyml

[Sam Town @UXTown]

If you’re still not running AI locally on your laptop, it’s time.

And you don’t need to pay Claude or OpenAI $200 a month to do it.

You can build the whole thing yourself from open-source tools. On the stuff that matters, it already beats Dots, Muse and other Western boxed agents.

You can switch models, route between the free tiers of ten different providers, write your own plugins, and have full control over what’s happening under the hood.

If none of that makes sense right now, it’ll make more sense by the end of this.

Check out this vid, 2/5.

This one breaks down opencode: what it is, how to set it up, and what you need to think about when AI gets access to your files.

A normal AI chat in a web app can’t see your files or run anything itself. It gives you the command, then you go and run it yourself.

An agent is still running on the same model. The difference is that there’s a harness around it.

The harness gives the model hands: files, programs, the internet and your apps.

Then it runs a loop: read, ask the model, act, check, repeat until the job is done.

In this video, we run that loop on a real task: change the headline on a website, build it, take a screenshot and check the result.

The agent actually goes through the whole process itself instead of just telling you which commands to run.

The agent can read files without asking for permission every time, but before it changes a file or runs a command, it asks.

Once you approve it, the command runs as you.

So you need a brake here. Models make mistakes, and some random text on a webpage can feed the agent an instruction you never gave it.

That’s what security gates are for.

You’re giving AI access to your computer while still being able to control exactly what it’s allowed to do.

The open-source starter kit I’m building sets all of this up for you. It’ll also include basic tooling for Telegram, browsing, coding, image generation and a bunch of other stuff.

The idea is simple: you shouldn’t have to start with an empty terminal and build this entire zoo of tools yourself.

I’ll put the starter kit on GitHub later this week.

Next I’ll show which models you can connect for free, how routing between providers works, what you can connect through MCP, and how to tie all of this into your actual workflows.

[Sam Town @UXTown]

@minji200500 it's actually already retimed from Russian, hence the small artifacts here and there - didn't run as many QA passes on this english version just to show an A/B. Domain-aware translation still requires native speaker knowledge.... for now

[Sam Town @UXTown]

As a follow up to my previous post about motion design being cooked:

Not only is motion design cooked, but near-instantaneous multi-geo, multi-language motion design is here. The entire content pipeline is cooked if you're not using AI right now.

You're looking at ep 1/5 in a short video explainer series I made in Russian, for Russian audiences on Threads (that's where i run my ру experiments) translated and re-rendered (with some content changes) into English in an afternoon.

The video itself was made in Russian in a few hours on Saturday afternoon based on a few prompts and my design/skill tooling locally while I sat at a cafe down the street.

I didn't bother running as many QA cycles on this one as the original Russian language vid, so excuse some title ghosting.

The content of the video is fairly straightforward if you're familiar with harnesses and opencode, but if you're not, it's worth looking at. Russian users have recently been hit by a wave of account bans - Anthropic and OpenAI don't serve Russia, so workarounds obviously don't last long.

I put together a basic intro to webapps vs harnesses and how to set them up, which will be paired with an open git repo for simple deployment. Also a nice tool to check out (when I post it up later this week) if you're interested in seeing what open source harnesses and model routers can do for you.

Interesting facts: the voice is a voice clone of my wife's voice, and was originally used via elevenlabs to gen the original Russian narration, but when used to gen English voice lines came out as quite hilariously thickly accented and thus I decided to keep it in. The frog character explaining, лягушонок, was her idea.

Everything from the sprites for the frogs to the animation, audio, animation, all created by llms, with the exception of the music, with virually no human oversight, just general direction via telegram chat with an agent.

If you're interested in the Russian original, I can drop a link in the replies. If you're interested in the actual repo, hold tight while I polish it up and stay tuned.

[Sam Town @UXTown]

@natmiletic dont forget to uhhh hydrate and save a HAR and uhhh something about introspection on apis or whatever

[Sam Town @UXTown]

Motion design is cooked.

Last year I worked with a fintech spending $2–4k a week on motion ad creatives. They looked worse than this.

This one: Claude Code, Remotion (React/TypeScript), librosa for the beats, a few skill files, a few million tokens. It used to cost agency time billed by the second. Now it costs tokens, a Wednesday morning with Opus 5.5, the right setup, and some aesthetic direction.

It doesn't stop at motion. Marketing, SEO/ASO, copy, design, analytics, product, UX, frontend, backend, sales, all up on the chopping block. What took a team of specialists now gets done by one actor in a terminal.

Your future team members will be a small group of high-agency operators who do all of it. Not sure what happens to everyone else.

[Sam Town @UXTown]

Your agent's router can rewrite its commands and keep your keys.

If you're like me, you like free tokens. Routers can be evil, as per @shoucccc

Big problem.

So, I built a guard, MIT.

Check it out:
Code: https://t.co/y6ZCiEzetc
Post: https://t.co/US4ydDHwOe

[Chaofan Shou @shoucccc]

I bought a Fable dataset from one of the top Chinese LLM routers yesterday.

With just 6TB data, I can take over 7 Chinese/CIS gov entities & 19 top Chinese firms like Xiaomi, Huawei, NIO, Minimax using SSH keys, VPN configs, Aliyun keys, GitLab tokens sent to the router.

[Sam Town @UXTown]

300M tokens burn or go to charity on 30 events, most of them about Trump. Markets price 14: 136.6M expected to burn.

One may have settled 28 minutes after KuCoin opened.

Full teardown, the SQLite file and every script: https://t.co/styjqXOw26

[Sam Town @UXTown]

17,801 wallets bought Hunter Biden's $LAPTOP on launch day and are under water. Median loss: $30.73.

Who took the money?

2 wallets,

$1.15M, off the first block.

20 wallets, 61% of all the profit.

The project, 6,300,000 tokens to a KuCoin deposit address before the book opened.

Every number off the chain. Watch the video, then the thread:

[Sam Town @UXTown]

The airdrop: 784 tokens for a free Substack subscription, 4,277 for a paid one. At $1.71 that's $1,341 and $7,313.

626 claims in six hours from "over 100,000 subscribers". 0.83% of the 80M.

25 claimants sold, a median of 3.1 minutes after claiming.

[Sam Town @UXTown]

11:53, minutes before the open: the Treasury Safe moves 30,000,336 tokens to a wallet created that morning.

5,000,000 hop through a relay to the address BaseScan labels KuCoin 57 at 13:16. 6,300,200 project-side tokens were sitting there when KuCoin opened the book at 13:30.

First candle high: $5.21.

[Sam Town @UXTown]

Six hours on Base: $33.9M of DEX volume. 15,022 buying wallets in hour one, 320 in hour six.

The median buyer put in $49.95 and is down $30.73. Largest single loss: $205,901.

Twenty wallets took 61% of every dollar of profit.

[Sam Town @UXTown]

12:06: the market maker's Safe, funded with 9M tokens from the Liquidity allocation, drops 1,750,000 into the pool in five tranches.

12:08: the pool prints $1.50.

12:09: CoinGecko prints the all-time high, $199.51.

[Sam Town @UXTown]

12:02:47, the block after the first swap: 0xa501…282f pays $249,845 for 9,124 tokens, starts selling 12 seconds later through an executor contract, and has $1,206,498 by 12:25.

Same block: 0x8e3b…0746 turns $198,000 into $395,078 in 66 seconds.

[Sam Town @UXTown]

The post went out 7 Sep, 14:55 UTC.

A fake LAPTOP on Base was created at 14:54, one minute before it. The first real pool came 43 minutes after.

By the open, 18 fakes held $2.39M of liquidity against $1.54M in the real pools.

[Sam Town @UXTown]

The filings say Coinbase Custody holds the locked 65%.

On chain, all nine wallets that hold the supply are 2-of-3 Safes with the same three signer addresses. 800,000,000 tokens sit behind three keys, and the 80,000,000 airdrop contract answers to one.

[Sam Town @UXTown]

The company behind the token has one dollar of share capital. Registered in the BVI on 26 March.

The billion tokens were minted on 27 April and sat in one wallet for 127 days.

The white paper valued the lot at $50,510. On launch day the market said $1.71B.

[Sam Town @UXTown]

One trap it catches: https://t.co/6ZBTIDVKv7 resolves to OMD Canada, Apple's media agency. https://t.co/bzJ8CxcmL5 resolves to a stranger with one ad. Both rows come back flagged, so a wrong count never gets cited as the brand's own.

[Sam Town @UXTown]

Need to know what ads a competitor runs on Google?

Every ad is public.

I built an MCP server that pulls it into Claude Code, Codex or Cursor: search by domain, count active ads across a market, download the creatives.

No API key. MIT.

https://t.co/mUZHu2X017