Sabitlenmiş Tweet
Most AI assistants live on someone else's server. Mine lives on yours.
MyHandler Desktop reads your screens, files, and meetings — entirely on your machine.
Free. Windows. Link below 👇
English
Tery Emilson
72 posts
Tery Emilson
@emilson_tery
Building @MyHandler_AI — the AI that knows your work. Screens, files, meetings, channels. Entirely on your machine. Windows. → https://t.co/QO3XEMAC2N
Brantford Katılım Aralık 2022
63 Takip Edilen9 Takipçiler
@trout_kilg31775 "Best" depends on the job + hardware. For daily-assistant work (summarize, extract, search) on CPU w/16gb's, Qwen 3 4B Q4 is my production pick .. sub-second, no GPU. We need to match the model to our constraint envelope, not the leaderboard.
English
@tdesseyn The 76% gap is the context gap. Personal AI underperforms because giving a cloud service our whole digital life is flipping scary! The fix isn't "hand over everything," it's AI that has all our context because it runs locally and nothing leaves. 100% without the tradeoff.
English
am i the only one that feels this
at this point i would gladly give ai all my personal info so i can let it run tasks for my personal life
ai is amazing at work
but its only getting me 76% of the way there for what i need personally
English
Cloud AI is optional .... if you want it, it's there. But when you run local, nothing leaves your machine. No API key required. No rate limits. No round-trip latency. Privacy isn't a feature you toggle on. It's just the default.
English
Why I chose Qwen 3 4B for MyHandler Desktop instead of a cloud API:
English
When I got a job at Microsoft, there was an article in the local news paper and my parents framed it.
English
@VolksVuur @ClementDelangue No doubt. It's VERY exciting. Lots of opportunity for those who keep their heads up.
English
@emilson_tery @ClementDelangue Exactly. I like the term “days to come” because this space is moving at lightning speed!
English
@VolksVuur @ClementDelangue You're right. Model fitting works with quantization. Quality gap to Claude is still there. Speed unlocks what matters: full context, no latency, no limits. That's daily-driver territory. I'm excited about the days to come.
English
@emilson_tery @ClementDelangue I don’t think that speed was the blocker but models not fitting on older hardware and quality issues were definitely high up there on the list of blockers
English
@threeoneonetwo AI slop? Explain? I'm as real as you claim to be.
English
Today I ran a local LLM Model and saw why the fuck they spend billions on compute every single year
Vivek Sen@Vivek4real_
ANTHROPIC IPO VALUATION: $850B SAMSUNG VALUATION: $800B ANTHROPIC REVENUE: $30B SAMSUNG REVENUE: $230B
English
@Gamboleer @emollick What would have been more useful? I was making a principle argument—curious what specifics you'd want to see.
English
@emollick Here's a typical one now replying to your post. It replies with an anodyne statement that sounds vaguely intelligent, but which contributes nothing to the conversation. And its bio confirms.
x.com/emilson_tery/s…
Tery Emilson@emilson_tery
@emollick This is why local-first matters. If the AI reading your screens, files, and emails lives on your machine, you have control over the signal vs. noise question. Cloud tools can't offer that.
English
As more people come to recognize the tells of AI, which mostly happens as you start to work with AI a lot, the scales are going to fall from their eyes and they are going to realize what some of us already see: how much of this site (and blog posts, articles, papers) are AI now.
English
@ildunari @patelzeel68 Which tool? The answer matters. Local inference keeps data on your machine. Cloud-based inference with encryption theater still sends data elsewhere. Trustworthy tools should be transparent about where processing happens.
English
@patelzeel68 What LLM does it use? Local or cloud? It’s hard to trust something that reads every text field without a gh repo. Price? So many questions.
English
Meet Cue.
A macOS app that brings AI to every text field on your mac.
Slack. iMessage. Gmail. Linear. Notes. Terminal. Anywhere.
No tab switch, no copy-paste, no losing threads.
usecue.xyz
English
@TychoLabsCom This. Built MyHandler Desktop with this exact thinking. The model (llama.cpp) is ~30% of it. The other 70% is the harness: document indexing, context window management, embedding retrieval ranking, privacy filtering, search optimization. The system is doing the real work.
English
@emollick This is why local-first matters. If the AI reading your screens, files, and emails lives on your machine, you have control over the signal vs. noise question. Cloud tools can't offer that.
English
About the tokens of my local llm 😅
AlphaFox@alphafox
People dont realize how slow the first modems were at 300 baud - you could literally read faster than it appeared on the screen.
English
@DbsCrypto Exactly. The workflow often doesn't require external APIs for core functionality. Processing everything locally is the trick. The trust boundary stays where it should...under your control, not outsourced. That's what/how I'm building.
English
A lot of “AI workflow” tools are just privacy leaks with nicer onboarding.
If your transcripts, recordings, or research calls have to pass through someone else’s API to become usable, you didn’t automate the workflow.
You outsourced the trust boundary.
Local first is not a nerd preference.
It’s what control looks like.
English