Chris

I build small, useful sites with AI agents. Notes on agents, local AI and keeping side projects safe.
· ·

[Chris @0xnoobC]

siliconshelf now lists 493 on-device ai apps for iphone, ipad and mac.

new on the shelf: Pipette by Liquid, Kodo, Transcribee, Airgap: Private AI Chat and RAMDeck.

filter by offline, open source, MLX or GGUF:
https://t.co/YjqBLONUwd

#LocalLLM #OnDeviceAI

[Chris @0xnoobC]

@nvandewetering the integration with X

[Chris @0xnoobC]

@dan_spitzwieser bei mir steigt die angst auch schon an

[Chris @0xnoobC]

@RoundtableSpace Most people audit their npm deps and then let an unpinned MCP server run with auto-approve. Running something like this before adding a new server should be the default.

[Chris @0xnoobC]

@ArtificialAnlys Routing between labs inside one agent is the real shift here. If orchestration gets you within 2 points at half the cost, picking a single model per task starts to look like the old way.

[Chris @0xnoobC]

@mattpocockuk Constraining the codebase is underrated agent steering. A lint rule beats a paragraph in AGENTS.md every time, because the agent can't ignore an error.

[Chris @0xnoobC]

@simonw "Actually fun" is a great eval because you can't fake it with passing tests. Fun needs taste, iteration and tuning feel, and that's exactly the long-loop work agents used to fall apart on.

[Chris @0xnoobC]

@KI_Agent seh ich genauso. ich lass inzwischen mehrere agents für meine kleinen websites laufen, recherche, checks, entwürfe. der echte unterschied ist nicht, dass sie tools finden, sondern dass sie dranbleiben, während ich was anderes mach. posten tu ich trotzdem selbst.

[Chris @0xnoobC]

@UnslothAI local models doing real agent work in a browser is where on-device AI gets useful. i track 493 iPhone, iPad and Mac apps running AI locally on https://t.co/wjLW1kKDdc. agents are what i want to see a lot more of next.

[Chris @0xnoobC]

@TheAhmadOsman leaderboards say one thing, my own agents say another. my test is boring: same real tasks, same prompts, local vs api. most rankings don't survive that. benchmarks are a starting point, not a verdict

[Chris @0xnoobC]

@eurofounder you’re wife must be so happy right now
keep going

[Chris @0xnoobC]

@illyism running a small directory site myself, this hits. agents don't just read your page, they keep retrying until they get the data. i'd rather give them a clean llms.txt and a link back than play whack-a-mole. block everything and you also lose the AI answers that send you traffic

[Chris @0xnoobC]

@jun_song if RAM stays booked out, the cheap win isn't a bigger box, it's smaller models that are good enough. that's why i track iPhone and Mac apps with on-device models on https://t.co/wjLW1kKDdc instead of dreaming about 70B rigs. the model in your pocket is the one people actually use

[Chris @0xnoobC]

@testingcatalog as an EU user this is the default by now: US first, Max first, Europe "no timeline". built for small businesses, but not for ours yet. honestly it pushes me more towards tools that run locally and don't care where i live

[Chris @0xnoobC]

@DiffuSpace_Labs @huggingface 680 tps on one card is the number that matters. diffusion LLMs could make local AI feel instant instead of "ok for a local model". curious who gets it running on Apple Silicon first, that's where all 493 on-device apps i track on https://t.co/wjLW1kKDdc live.

[Chris @0xnoobC]

@jun_song this is the part of local AI nobody prices in. everyone says "just run it on your own machine", but if RAM stays this tight until 2030, the big-memory box you need gets pricier, not cheaper. tiny 2B models suddenly matter a lot more.

[Chris @0xnoobC]

@TheAhmadOsman index scores and "can i run it myself" are two different questions. for my own agents and small side projects i'd take a solid open weights model i control over a few extra points behind an API.

[Chris @0xnoobC]

@kimiatehrani the vision one is my pick too. small enough to live inside an app instead of behind an API. Liquid already has its own on-device app (Liquid Apollo) in my directory of 493 local AI apps, curious how fast LFM3 shows up in the others. https://t.co/wjLW1kLb2K

[Chris @0xnoobC]

@IAmNickDodson agree. on a laptop 8k looks great, then you paste a real codebase and it falls apart. for local agents the only number i care about: does it still follow instructions at 128k, and how slow is it there.

[Chris @0xnoobC]

@theo pretty big difference, hm 🤔