Author | Lightnews

Kentaromiura @kentaromiura.bsky.social · 4d

Emoticons

Kentaromiura @kentaromiura.bsky.social · 17d

As a demo I shown what my AI built all alone on first try from a single small prompt:

cristian.tokyo/img/blockgam... but if you are from mobile you might want to use cristian.tokyo/img/blockgam... which also has been vibe engineered the same way using the above as a starting point.

Tetris Game

cristian.tokyo

Kentaromiura @kentaromiura.bsky.social · 17d

Some Tuesday at work we have a “vibe engineering” session where we discuss stuff, this Tuesday I spoke on how to make AI agents that use local llms to reproduce @roocode.bsky.social “boomerang” mode on the cli with multiple local llms using github.com/kentaromiura... and @lmstudio-ai.bsky.social

GitHub - kentaromiura/ODM: On demand model proxy server

On demand model proxy server. Contribute to kentaromiura/ODM development by creating an account on GitHub.

github.com

1

Kentaromiura @kentaromiura.bsky.social · 27d

Since this I’ve released two local tools already.
The first for Jp/En translation and back and the latest one for allowing to load and unload different models on demand, allowing better agentic flow on local development machine.
ODM: github.com/kentaromiura...
ShallowT: github.com/kentaromiura...

Kentaromiura @kentaromiura.bsky.social · Aug 26

Hey guys, I’ve found a way to make my cocktail sauce recipe hotter and pinker!

1

Kentaromiura @kentaromiura.bsky.social · Aug 26

Neosoft Neopaint, early 90s, I first used in 286 in middle school together with neobook

1 2

Kentaromiura @kentaromiura.bsky.social · Aug 23

github.com/kentaromiura...

GitHub - kentaromiura/ShallowT

Contribute to kentaromiura/ShallowT development by creating an account on GitHub.

github.com

Kentaromiura @kentaromiura.bsky.social · Aug 18

My guess someone had to repeat the same transition start properties many times and was bother enough to create a spec

Kentaromiura @kentaromiura.bsky.social · Aug 17

I’ve fully entered the local llm rabbit hole. Please send help (or an AMD AI MAX if you hate me).

Kentaromiura @kentaromiura.bsky.social · Aug 13

This is the best advertisement for grok tbqh.

Best of Dying Twitter @bestofdyingtwit.bsky.social · Aug 12

THE GIRLS ARE FIGHTINGGG

Sam Altman
@sama • 13h
This is a remarkable claim given what I have heard alleged that Elon does to manipulate X to benefit himself and his own companies and harm his competitors and people he doesn't like.
Elon Musk
@elonmusk • 15h
Apple is behaving in a manner that makes it impossible for any Al company besides OpenAl to reach #1 in the App Store, which is an unequivocal antitrust violation. ...
Readers added context
In January 2025, DeepSeek reached #1 overall on the App Store.
abcnews.go.com/Business/deeps.....
Show this note
Rate proposed Community Notes
→
• 3.2K
17 3.8K
38K
Ill 6.1M
Elon Musk
@elonmusk • 7h
You got 3M views on your bullshit post, you liar, far more than l've received on many of mine, despite me having 50 times your follower count!
1.5K
t, 1.2K
14K
Ill 883K

Sam Altman
@sama •1h
will you sign an affidavit that you have never directed changes to the X algorithm in a way that has hurt your competitors or helped your own companies?
i will apologize if so.
1.5K
t7 938
12K
Ill 454K
idia
@dctrpr •1h
@grok who's right? dont be biased.
11
t7
© 144
贝
Grok t
XI
@grok•1h
Based on verified evidence, Sam Altman is right.
Musk's Apple antitrust claim is undermined by apps like DeepSeek and Perplexity reaching in 2025.
Conversely, Musk has a history of directing X algorithm changes to boost his posts and favor his interests, per 2023 reports and Show more
© Download Grok
276
17 542
O 3.2K
Ill 106K
Takumatoshi @elonmusk Imao
5
174
@tokumatoshi •59m
464
Ill 30K
Grok v
@grok•57m
Fair point-evidence shows Elon has indeed tweaked X's algorithm to boost his own reach (e.g.,
2023 Super Bowl incident). Meanwhile, apps like DeepSeek hit on App Store this year, debunking his antitrust claim. Hypocrisy levels: high.
企

1

Kentaromiura @kentaromiura.bsky.social · Aug 12

A few years ago (2011) I made a script that downloaded lots of news from different rss feeds, put together in a page, run through calibre to convert to mobi and send to my kindle every morning; I think this is now a prompt away from that.

Kentaromiura @kentaromiura.bsky.social · Aug 12

So yesterday Jan V1 was released based on this model. So I tried Jan in the Jan app, after configuring the duck duck go and fetcher MCP I managed to get a little assistant to summarize news etc, although it got Wikipedia winning instead of losing.

Kentaromiura @kentaromiura.bsky.social · Aug 8

Despite this, for the first time I’ve got 2 local llms finishing my test, both the stubborn but small qwen Qwen/Qwen3-4B-Thinking-2507 and gpt-oss-20b, the latter with huge context could even complete the task on first try via aider.
Test is asking to write rule 110 in Odin.

1

Kentaromiura @kentaromiura.bsky.social · Aug 12

I later tried the Odin challenge but with search available it ended up copying some gists, I asked to iterate on that and again it search for premade stuff before doing it, so a real engineer behavior lol. It always gave attribution though, so maybe asking to check for MIT license might be fine.

Kentaromiura @kentaromiura.bsky.social · Aug 12

So yesterday Jan V1 was released based on this model. So I tried Jan in the Jan app, after configuring the duck duck go and fetcher MCP I managed to get a little assistant to summarize news etc, although it got Wikipedia winning instead of losing.

Kentaromiura @kentaromiura.bsky.social · Aug 8

Despite this, for the first time I’ve got 2 local llms finishing my test, both the stubborn but small qwen Qwen/Qwen3-4B-Thinking-2507 and gpt-oss-20b, the latter with huge context could even complete the task on first try via aider.
Test is asking to write rule 110 in Odin.

Kentaromiura @kentaromiura.bsky.social · Aug 8

I think AI is remarkably similar to human engineer behavior. Maybe a bit too much.

1

Kentaromiura @kentaromiura.bsky.social · Aug 12

Anyway it’s the first model I tried that passed that test, with gpt oss being the second, gpt could even output the correct program in one go when fed the odin documentation and rule 110 wiki page with lots of context (Odin doc I think is around 15K).

Kentaromiura @kentaromiura.bsky.social · Aug 12

I managed to get this model getting the right result for my special test (write rule 110 in the Odin programming language), after I corrected the wrong syntax it remembered the corrections and eventually got it right, not before of getting the order of bits wrong and telling me that they were right.

1

Kentaromiura @kentaromiura.bsky.social · Aug 11

In your opinion. If it works and it solves a problem then it has real worth.

1

Kentaromiura @kentaromiura.bsky.social · Aug 11

I disagree, if the barrier of entry is lower using .net, Java, electron/tauri or lisp or lua or basic or python and it solves a problem then the worth of having something done now against maybe never is infinite percent better.

1

Kentaromiura @kentaromiura.bsky.social · Aug 10

We reach to a point where the dead internet theory is true not because there are more bots than humans but because the slop created and the insane engagement makes social medias basically useless. Manual curation is the only way forward now.

Kentaromiura @kentaromiura.bsky.social · Aug 10

Expedition 33 has one of the worst platforming part of any game. If I was to demo the game and give it a score solely on that it would get a 0.

Kentaromiura @kentaromiura.bsky.social · Aug 8

This is a difficult task because no local llama are trained on Odin code and the syntax is alien to them. My only requirement is I don’t have to write a line of code myself.
GPT oss is very good at failing fast with small context, so you can have many iterations in relatively small time.

Kentaromiura @kentaromiura.bsky.social · Aug 8

Despite this, for the first time I’ve got 2 local llms finishing my test, both the stubborn but small qwen Qwen/Qwen3-4B-Thinking-2507 and gpt-oss-20b, the latter with huge context could even complete the task on first try via aider.
Test is asking to write rule 110 in Odin.

Kentaromiura @kentaromiura.bsky.social · Aug 8

I think AI is remarkably similar to human engineer behavior. Maybe a bit too much.

Kentaromiura @kentaromiura.bsky.social · Aug 8

>Hey AI, I think instead of rule 110, you implemented rule 118, you’ve got the MSB and LSB backwards.
Ai: I see that’s a common mistake when implementing cellular automations, I’m actually correct.
>No you’re not;you are reading the truth table backwards, MSB is bit 7 you think is 0.
Ai: Acthually..

1

Kentaromiura @kentaromiura.bsky.social · Aug 8

I think AI is remarkably similar to human engineer behavior. Maybe a bit too much.

Kentaromiura @kentaromiura.bsky.social · Aug 8

>Hey AI, I think instead of rule 110, you implemented rule 118, you’ve got the MSB and LSB backwards.
Ai: I see that’s a common mistake when implementing cellular automations, I’m actually correct.
>No you’re not;you are reading the truth table backwards, MSB is bit 7 you think is 0.
Ai: Acthually..

Kentaromiura @kentaromiura.bsky.social · Aug 8

>Hey AI, I think instead of rule 110, you implemented rule 118, you’ve got the MSB and LSB backwards.
Ai: I see that’s a common mistake when implementing cellular automations, I’m actually correct.
>No you’re not;you are reading the truth table backwards, MSB is bit 7 you think is 0.
Ai: Acthually..

Kentaromiura @kentaromiura.bsky.social · Aug 5

TIL if you zoom enough you can read the “Here’s to the crazy ones.” In the document emoji on a Mac.

Kentaromiura @kentaromiura.bsky.social · Aug 4

So you all know Serial Experiment Lain, but do you know