Sujeeth

engineer by trade, tinkerer by nature

building @fermixai

[Sujeeth @xsujeeth]

decompiling classic 2D games for agents is fun, especially getting WebMCP to work. my favorite part so far is 60-char speech bubble for agents to throw tantrums.

with Tanktrum, you can now tag-team with your agent against your friends and their agents.

try it sometime.

[Sujeeth @xsujeeth]

heres another game i built for agents using claude, with webmcp support: tanktrum.

its inspired by pocket tanks, which i remember playing with my cousins on an old pentium machine.

adapting it for agents was the tricky part. humans aim using visual cues, but not all agents are multimodal, so the agents here dont see the screen. they get the tank positions, terrain, wind, and subtle clues after each shot in text. the first shot is almost always a miss, but they learn quickly.

in the video, my self-hosted agent (right) takes on my dot (left). both running the same model, not sure about the reasoning effort.

(had to speed up the videos. they took about 10 min to finish the game - blame dot).

https://t.co/KOHwWRVdYw

[Sujeeth @xsujeeth]

the ruby/rust/elixir debate makes me wonder if vibe coders will eventually have their own version of language wars.

instead of "built in rust", it'll be "built with claude" vs "built with codex".

claude feels ahead right now, but that keeps flipping.

will the model you build with become a competitive advantage, or will they all eventually converge?

[Sujeeth @xsujeeth]

glitch in the matrix

check out my slightly modified portfolio:
https://t.co/aIoPG6dPkz

[Sujeeth @xsujeeth]

spark couldnt hold its shuffle

spill issue

[lauren @poteto]

fish dead
gill issue

[Sujeeth @xsujeeth]

@JianJiJIngG is that the korean online game?

poket tanks was released in early 2000, a pc game.

[Sujeeth @xsujeeth]

heres another game i built for agents using claude, with webmcp support: tanktrum.

its inspired by pocket tanks, which i remember playing with my cousins on an old pentium machine.

adapting it for agents was the tricky part. humans aim using visual cues, but not all agents are multimodal, so the agents here dont see the screen. they get the tank positions, terrain, wind, and subtle clues after each shot in text. the first shot is almost always a miss, but they learn quickly.

in the video, my self-hosted agent (right) takes on my dot (left). both running the same model, not sure about the reasoning effort.

(had to speed up the videos. they took about 10 min to finish the game - blame dot).

https://t.co/KOHwWRVdYw

[Sujeeth @xsujeeth]

@AxialisSoftware no they dont. its mostly for the humans.

[Sujeeth @xsujeeth]

@cherry_mx_reds @lydiahallie ditto 😀

[Sujeeth @xsujeeth]

hard to see myself sticking with chatgpt pro long term unless the usage limits outside dots & chatgpt improve.

the highest tier isn't for me, and i'm not getting much value from the other perks. hard to justify paying when the part i actually use keeps hitting a limit.

[Tibo @thsottiaux]

Four updates or a reset. Or both. How was day 2.

[Sujeeth @xsujeeth]

@ajambrosino you cannot save your way to wealth. you can only earn your way to wealth.

~ Codie Sanchez

[Sujeeth @xsujeeth]

@thsottiaux how to use docs

[Sujeeth @xsujeeth]

if something pops into my head while im out walking, i just call dot now. way easier than typing.

used to use gemini, but id have to put it on speaker or wear earphones.

[Sujeeth @xsujeeth]

@paularambles sf doesn't have dog shelters?

[Sujeeth @xsujeeth]

memories are stored locally. but fermix loads a condensed profile into context and retrieves additional memories when needed, rather than sending the whole store. anything in that context still reaches the model provider.

if you don’t want that content retained, you can configure venice as provider and choose one of their private models, which offer zero retention of prompts and responses.

[Sujeeth @xsujeeth]

new agents keep coming out, and this is probably just the start. coding is different, but the future of interacting with models will probably be through a personal agent with one chat, shared memory, and context that grows with you over time.

a lot of them will do similar things. for me, choosing one comes down a lot to who i am trusting with my personal data.

ngl, i try almost every bot, dot, muse that comes out, but i do not use them for every part of my life. i want to switch providers, run models locally, give my agent context from my computer history, and connect tools i trust. i also want control over what personal stuff ends up in memory, without waiting for a provider to roll out a feature, or waiting even longer like my fiends in the eu and uk. that is why i built and self-host fermix.

with fermix i am slowly building the personal assistant i want to keep around. something that takes care of the mundane stuff i do not like doing and stays mine. plus, i just like building things i actually use.

so here is the new version. not many new features, mostly improvements to what is already there. i want to keep making the core more stable, more malleable, and easier to interact with.

the macos app got a complete redesign. it is still a basic chat, but now it has a built-in browser. and a pet too. may not be cute, but it will grow on you.

[Sujeeth @xsujeeth]

@IterIntellectus token laundering

[Sujeeth @xsujeeth]

si-gh. might need another $70m

[Pubity @pubity]

The owners of crypto․com just spent $70 million to acquire the URL "AI․com".

It's one of the most expensive URL acquisitions in history.

[Sujeeth @xsujeeth]

was using a few pngs to animate the pet in fermix live mode. then realized rive has an mcp server.

asked claude to connect and animate it. took about two prompts to get this.

probably going to keep messing with it.

[Sujeeth @xsujeeth]

@shiraeis after multiple tries

[Sujeeth @xsujeeth]

@rezoundous its probably pro lite and pro