Sabitlenmiş Tweet
No API keys. No servers. Just your browser.
I built an on-device AI agent in <200 lines of code using Gemini Nano / Phi-4-mini.
It loops, calls tools, even reads your filesystem.
English
Ryan Seddon
14.3K posts
Ryan Seddon
@ryanseddon
Front-end Enthusiast, 🛰 space bot connoisseur @earthin24, interested in renewable tech & electrification.
Melbourne, Australia Katılım Eylül 2008
1.1K Takip Edilen5.2K Takipçiler
My 1990 toyota celsior did this, had crazy features on the car.
PATINA RESEARCH@patinaresearch
1991 btw
English
@argyleink It's been using a transformer architecture since 2023 but is still bad 😅 jackcook.com/2023/09/08/pre…
English
Why isn’t AI used for typing auto correct on mobile yet? What we have now is so primitive in comparison to what it could be right?
English
@rileytomasek It exists xcede is a cli tool that can do everything without ever opening xcode codeberg.org/luxmentis/xcede
English
xcode is the ultimate use case for computer use.
there are so many personal ios apps i haven't built because i refuse to subject myself xcode.
English
@badlogicgames We have internal claude files that people have done with linebreaks in the description frontmatter that breaks in every other agent that uses an actual yaml parser. 🤦♂️
English
Claude Code agent files have YAML frontmatter. Claude Code seems to have implemented it's own vibe coded YAML parser, and allows invalid YAML. Now people demand pi also parses YAML in the same broken way as Claude Code does.
Glorious.
github.com/earendil-works…
English
@vimtor @tan_stack Nice, it always chugged bad on my iphone. Now is usable.
English
we know ram prices are through the roof
that's why we made models.dev not consume half of yours when visiting the site
now it loads and searches instantly thanks to @tan_stack virtualizer
English
@thdxr Goonbag and somewhat gooning means getting drunk off of cheap wine in Australia. The wine that comes in a bag.
English
@lucasmeijer I've been running a variant called tinyclaw that has been nice, paired with camofox its been doing a good job reading school emails. Just some processes within a tmux session.
English
now that the hypetrain is gone, I think i might try openclaw. i was a bit too early last time and it started answering a ai friends whatsapp group. (they didn't manage to make it leak anything from my machine but was still scary :)
English
@wesbos Wonder if pi-autoresearch could of got the job done too. Pretty cool either way.
English
Burned $91.34 with Claude Code /goal in 3.5 hours
Unreal, It was able to reverse engineer it!
Wes Bos@wesbos
/goal reverse engineer the amaran BLE mesh protocol. I pointed you at the the light. Make no mistakes
English
@Beelinkofficial send this man one of your GTR9 128GB models.
antirez@antirez
I just pushed a big refactoring of DS4 backends with CUDA support and single direction activation steering. The Metal path should be unaffected. Note: I only support hardware I have own (or have full access to): so just M3 (no M5 NE for now), DGX Spark.
English
I wish I had my ps1 memory card, my 13 year old self spent hours producing bangers.
Obsolete Sony@ObsoleteSony
The PS1 had a whole music studio on a disc
English
@LukeParkerDev I really wanted pnpm global virtual store to not cause weird issues to fix the install and disk space issues but it ain't ready yet.
English
git worktrees are fun and all, until you have 3563 instances of node_modules
English
This is the bee watcher watcher method
Anthropic@AnthropicAI
New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text.
English
@mitsuhiko Have something similar, have you had any luck indexing the in a vectoredb? Or is greping files just better?
English
Gotta say my librarian skill (when you find a github/git url, fetch and cache the repo locally) has been one of the most useful skills and most commonly used ones.
English
@jaffathecake Microsoft has the prompt api in edge, where can i read up on the concerns?
English
So, Chrome's "web standard" Prompt API:
Mozilla: Opposed
WebKit: Opposed
Microsoft: Several concerns
W3C TAG: Several concerns
Developers: Mostly negative
Chrome: Ships anyway.
A sad time for web standards. But, I guess someone at Google will get promoted, so 'every cloud…'
English
@mitsuhiko It's even been present pre-llm too with predictive text seas.harvard.edu/news/predictiv…
English
I analyzed my coding sessions and on the text interactions some words stand out. And well, they also show up on Google Trends as spiking. Oh and so much slop in my Twitter mentions and on GitHub. Thus here are some updated thoughts on all of this. lucumr.pocoo.org/2026/5/4/conte…
English
Ryan Seddon retweetledi
@zeeg I've been using tailscale aperture which gives good granular per user info like this across any llm api.
English
Alright Anthropic, OpenAI:
I need APIs that give me usage data. Granular. Per user.
I need this in the same way its provided by every other company for all of time.
What I dont want: "heres a $100,000 line vague line item of token spend".
English
Wtf did I just watch…
dax@thdxr
tired of this misinformation so we made a video on the truth behind the anthropic vs opencode drama
English
@GergelyOrosz They're changing to their ai credit model though as of june, which converts to cost based pricing so same as vercel gateway, openrouter, opencode zen etc. They could do something like opencode go with optimised inference or the composer route and fine tune an open model
English
Microsoft / GH Copilot is now paying the price for not owning a foundational model of their own that is good at coding. They now need to resell the markup of OpenAI, Anthropic, Google
Cursor was v smart to build Composer, to be able to control pricing. GH done subsidizing tho:
AshutoshShrivastava@ai_for_success
I was going through GitHub Copilot pricing changes… this is wild. They quietly changed model multipliers and some of these jumps are insane. Opus 4.6 is 9x Opus 4.5 is 5x Opus 4.7 is 3.6x Sonnet 4.6 is 9x Sonnet 4.5 is 6x Gemini 3 Pro is 6x Gemini 3.1 Pro is 6x GPT 5.1 is 3x GPT 5.2 is 3x GPT 5.3 Codex is 6x GPT 5.4 is 6x
English
Ok is this an AI generated @mattpocockuk in the ai engineer youtube thumbnail? It's giving me uncanny valley vibes.
English