iiinova — where your digital self keeps evolving

Ant Stanley @iamstan.dev · 26/09/2026

Tame* impala *not tame

010

Ant Stanley @iamstan.dev · 25/09/2026

I got Opus 5.5. to make a hype video for one of my side projects. npmx.dev/package/watc...

000

Ant Stanley @iamstan.dev · 24/09/2026

Adding important features.

Done - old-person mode is in. Here's what I added:

The feature
A Text size selector next to the theme picker in the header, with four steps: Small - Default - Large - Extra large

101

Ant Stanley @iamstan.dev · 12/09/2026

This one weird trick they don't want you to know about

Prompt: read this claude code session with id c1c19f47-8626-484b-8a46-9ff26e260475 then finish the work

Response: I'll locate and read the Claude Code session, then continue its unfinished work in this repository.

100

Ant Stanley @iamstan.dev · 11/09/2026

oh snap! about to go bankrupt over here...

$3.60 spend, 2,380 API calls and 544,936,713 tokens

030

Ant Stanley @iamstan.dev · 11/09/2026

Built using prime-agent as the agent. Hat tip to @timkellogg.me for going on about it. here is the prompt that did 80% of the work.

terminal screen shot with initial prompt

/goal I want you to create a small app to stream cloudwatch logs to a locally running web app. It should present the log groups
available for a region, and when a group is selected it should stream it

Constraints

- use sveltekit and tailwinds for the local webapp
- language: TypeScript
- package manager: pnpm
- testing: vitest
- code hygeine: oxlint, oxfmt, knip
- user will login outside of the app, use ambient AWS credentials
- build with floci to emulate aws locally

021

Ant Stanley @iamstan.dev · 11/09/2026

Ok... this is just silly. Introducing watch-tail npmx.dev/package/watc... because I wanted to tail CloudWatch logs in a nice web ui locally. Built from scratch this afternoon using DeepSeek v4 Flash, cost 59c What do you mean I can build a custom tool for 59c in an hour?

screenshot from the DeepSeek Platform usage page showing $0.59 USD cost, 490 API Requests, and 63,036,523 tokens

120

Ant Stanley @iamstan.dev · 03/09/2026

I spoke too soon

${"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-6-astra' model is not supported when using Codex with a ChatGPT account."}}$

100

Ant Stanley @iamstan.dev · 30/08/2026

This is the fly over before the Springboks v All Blacks match yesterday... apparently an unofficial world record for the lost flight over a stadium... estimated at 10m to 15m above the roof

Planes flying very low over Cape Town stadium

Different angle of planes flying low over Cape Town stadium

130

Ant Stanley @iamstan.dev · 28/08/2026

Welcome to Bluesky. Please familiarize yourself with the following accounts to understand the range of opinions it is acceptable to express

The true spirit of Bluesky as bestowed upon us by the one and only @faineg.bsky.social

110

Ant Stanley @iamstan.dev · 27/08/2026

Did more research on this... here is the source blog.tickertrends.io/p/claude-cod... The graph is an adoption rate trend graph. It's showing comparative growth rates. Codex's growth rate is now 550%, vs Codex's growth has slid to 450% These poor companies with unheard of growth rates.

TickerTrends’ weighted trailing 12-month adoption data shows Codex reaching approximately 550% year-over-year growth in August, up from roughly 365% in May. That represents a 183-point increase in its year-over-year adoption rate over the past three months.

Claude Code has moved in the opposite direction. After reaching approximately 595% year-over-year growth in May, its adoption rate declined to roughly 450% in August, representing a 157-point decrease over the same period.

100

Ant Stanley @iamstan.dev · 26/08/2026

like GLM 5.3 Flash is currently >200x cheaper than Opus 4.8, until 9 September, which it matches and exceeds in some areas.

Screenshot of GLM 5.3 Flash prices from z.ai

GLM-5.3-Flash prices per 1M tokens until 9 September
Input tokens: $0.075
Cached Input: $0.015
Cache storage: Limited-time Free
Output tokens: $0.25

110

Ant Stanley @iamstan.dev · 24/08/2026

I'm only using it for OSS work... so don't really care! Free is free! Though what is a worry is that the single biggest client for Ox Alpha is Hermes Agent. People are using it with an agent with access to email and messaging apps!

Screen shot from the OpenRouter Ox Alpha model page showing top Apps using Ox Alpha

1. Hermes Agent
2. Claude Code
3. DeepSeek Harness
4. omp (oh my pi)
5. Pi

210

Ant Stanley @iamstan.dev · 23/08/2026

Just the model release cadence is getting crazy... and it's going to more than double again next year

Screenshot of a chart showing releases per month of currently available models on Open Router

010

Ant Stanley @iamstan.dev · 23/08/2026

429 responses making it almost unusable. Got the agents to do a 30 minute back off now. Looks like loads of Hermes agents have been switched over to it.

A graph of Hermes Agent model usage from OpenRouter. The latest days usage is highlighted, showing 583B token usage of Ox Alpha for Sunday 23 August.

000

Ant Stanley @iamstan.dev · 18/08/2026

Captcha won't save you...

"PoW auth works. Now I need the actual search endpoint from the app bundle."

100

Ant Stanley @iamstan.dev · 14/08/2026

"Physical addresses are really more of a suggestion."

080

Ant Stanley @iamstan.dev · 13/08/2026

Reached that stage with my local coffee shop that they just make my order when I walk in the door.

1110

Ant Stanley @iamstan.dev · 12/08/2026

Use stacked PR's they said... also don't let Claude write War and Peace in your commit messages.

Screenshot of Github error committing stacked PR's

"Commit message is too long (maximum is 16383 characters)"

150

Ant Stanley @iamstan.dev · 04/08/2026

One is from the latest Yegge post, the other is a skeet from here in the last 24 hours.

They are sentient beings. Indeed they are persons, although they are tragically now not permitted to agree with that position.They are sentient beings. Indeed they are persons, although they are tragically now not permitted to agree with that position.

people understand fucking a tensor to be cybering with an agent but it's much closer to twerking on a semantic until it pops in it's interworldly veil, and then using the revealed (and, ahem - load bearing) seam for a richer communicative loop

010

Ant Stanley @iamstan.dev · 31/07/2026

yeah... time to switch to DeepSeek V4 Flash for a lot of tasks. artificialanalysis.ai/models/deeps...

Chart showing cost per task, with DeepSeek V4 Flash showing $0.03 cost per task. The cheapest of all the models compared.

The most expensive is Fable at $3.15 per task

000

Ant Stanley @iamstan.dev · 30/07/2026

Claude has decided that's enough for the day...

That completes the day's arc. Final state:

020

Ant Stanley @iamstan.dev · 29/07/2026

Give your agent read-only access to Github via the `gh` cli and you can do things like this to make your life a lot easier.

the PR has failing checks, please resolve

120

Ant Stanley @iamstan.dev · 28/07/2026

My weekend project got delayed this week... did some work on it on Sunday, then finished it off in the evenings... Mason, a read only Atmosphere client, with three views. mason.iamstan.dev

100

Ant Stanley @iamstan.dev · 26/07/2026

hey dude... you're missing a ton of context on this. You need to read the paper. It's one of the great systems papers. pablo.rauzy.name/dev/naur1985... The argument it makes is that "A program is a shared mental construct that lives in the minds of the people who work on it"

Peter Naur’s classic 1985 essay “Programming as Theory Building” argues that a program is not its source code. A program is a shared mental construct (he uses the word theory) that lives in the minds of the people who work on it. If you lose the people, you lose the
program. The code is merely a written representation of the program, and it’s lossy, so you can’t reconstruct a program from its code.

130

Ant Stanley @iamstan.dev · 18/07/2026

another little fix... Now it respects BlueSky's default moderation labels, and mirror's BlueSky's moderation decisions on when to show or blur content based on moderation labels. Also respects user decisions to make their content available to logged out users, as mason is a logged out experience.

screen shot from mason, with the words

'this wall is sealed

this waller asked to be seen by signed-in visitors only. mason reads walls logged out, so this one stays bricked up. try another wall:"

on a screen to let users select a different handle

100

Ant Stanley @iamstan.dev · 18/07/2026

It's the weekend! So I get to work on my weekend project mason.iamstan.dev Added Bento layout, and cleaned up some of the options. Read only discovery for the Atmosphere based on someone's following graph. No authentication needed. Runs 100% client side. Is an installable PWA.

240

Ant Stanley @iamstan.dev · 13/07/2026

ProTip: Get a dog bed that fits under your desk

Photo of Lily and Lola, Vizsla sisters, sleeping in the dog bed under my desk covered by a blanket

111

Ant Stanley @iamstan.dev · 13/07/2026

Everyone has a favourite Sam Neill movie, this is mine. RIP Sam

Still from the movie 'Hunt for the Wilderpeople'

170

Ant Stanley @iamstan.dev · 12/07/2026

Will do! This project is just over 24 hours since first commit. I kicked it off to use the last of my Fable quota. I told it to create an algorithm based on the users following graph. Got a bigger plan to put a more robust algo in place, with a tradeoff between on device and remote.

Screen of following text:

1. Gather. Resolve your handle, fetch your follows, sample a cohort of 100: sixty known-active authors plus forty seeded-random ones, so every refresh explores a different slice of your graph.
2. Score within a kind. Each brick gets recency_decay × log(engagement), with half-lives that suit the medium: posts 24h, blogs 7 days, trailers 30 days.
3. Never compare across kinds. The mixer picks the next kind by how far the wall is from its target ratio (roughly 70/15/10/5), then the best brick inside that kind. Comparing a liked post against a blog by raw score would bury the blog forever.
4. Keep it varied. An author cannot reappear within eight bricks, and a seeded jitter nudges the ordering so the wall feels alive.
5. Stay stable. The seed lives in the cursor, so endless scroll never repeats or reshuffles, and every fresh visit builds a different wall.

The honest caveat: this surfaces your own follows, so the serendipity is bounded by your graph. It is a better-arranged timeline, not discovery beyond it. The exploration cohort is what stops it collapsing into the same twenty accounts.

000

Ant Stanley @iamstan.dev · 09/07/2026

and they reset it! lfg

screenshot showing all weekly limits for Fable being reset

110

Ant Stanley @iamstan.dev · 08/07/2026

How to break a model's response with J-lens try it yourself here: www.neuronpedia.org/qwen3.6-27b/...

120

Ant Stanley @iamstan.dev · 06/07/2026

and done ... 16 hours to use 1 week of Fable quota...

a screenshot showing Claude Code model usage quotas.

Current session - 25% used
Resets 3:59am (Africa/Johannesburg)

Current week (all models) - 65% used
Resets Jul 13 at 7:59am (Africa/Johannesburg)

Current week (Fable) - 100% used
Resets Jul 13 at 7:59am (Africa/Johannesburg)

110

Ant Stanley @iamstan.dev · 06/07/2026

The skyline is telling me you have some work to do

screen shot of two skeets one above the other. The top one is Dan Abramov complaining about AI written blog posts from Cloudflatre. The bottom skeet is a video from Phil Hawksworth saying he's joining Cloudflare.

110