@saeid

Machine Learning | Data Science & Al | Python scikit-learn SQL FastAPI | Top 10 Finalist - Google Cloud Gen Al Hackathon

[@saeid @SaeidSahil9064]

@AdamZinebi1 Makes sense 🙂

[@saeid @SaeidSahil9064]

@droke_daniel I find it funny how fast pure RL finds a shortcut — one word it saw twice became the whole output by step 700. What was the reward actually scoring?

[Simon Høiberg @SimonHoiberg]

You can now get Opus 5.5 quality out of Qwen 3.8 27B.

You just have to teach it the job first.

We let a frontier model work through thousands of realistic support cases, keep only the ones that meet our standard, and train a small LoRA on Qwen with them.

Now it runs on our own hardware. No API bill, no rate limits, and nobody can change the deal on us 👇

[@saeid @SaeidSahil9064]

@0xSero A 1.7B model making HTML files in the browser is the part that stays with me. How long can it keep going in Pi before it loses track on a long task?

[@saeid @SaeidSahil9064]

@SimonHoiberg The filter step is what caught my eye. I train small models too, and the examples I keep matter more than the ones I make. How many of the thousands did you end up keeping?

[@saeid @SaeidSahil9064]

@alexandr_wang Adding Muse to a thread by email feels useful, because that is where the real back and forth already lives. Does it pick up the full thread when you add it halfway through, or only what comes after?

[@saeid @SaeidSahil9064]

@amypretzel @earthtojake I like the one feature at a time rebuild, so you can watch the part grow instead of trusting one big run. Do the fit and clearance checks run on their own after each rebuild?

[@saeid @SaeidSahil9064]

@SarvasvKulpati The live edit part is what got me, the doc fills in while you talk and the links get added on the way. How does it handle the half thoughts you change your mind on mid sentence?

[@saeid @SaeidSahil9064]

@romankhrupa @liquidai Zero tokens and it runs on your laptop is the part I liked most. A small model doing a real job like this in one weekend is really cool.

[@saeid @SaeidSahil9064]

@kirtandopamine Letting agents use a real phone from the terminal is the part I liked most. Can it run the same test script on more than one phone at once?

[@saeid @SaeidSahil9064]

@MengTo Walking through your own room in the browser is such a cool idea for tutorials. The lighting change part stood out to me, it makes the scene feel real.

[@saeid @SaeidSahil9064]

@Israfilv2 The 2.6 GB size is what caught my eye. It still keeps tool calls and 1m context, which sounds great for a normal laptop.

[@saeid @SaeidSahil9064]

@b1rkh0ff The score only going down part is scary. A 0 for a pasted reply means it stays buried even if the next reply is typed by hand.

[@saeid @SaeidSahil9064]

@TheAhmadOsman The 1B beating Opus part got me thinking. If clean data can do that much, small models running on a normal laptop start to feel real, not just a dream.

[@saeid @SaeidSahil9064]

Fine-tuning a 1.7B model into a voice typer that beats Wispr Flow on latency is wild. 65ms vs 520ms is a huge gap. What kind of data did you train it on to get it that natural? #AI #MachineLearning

[Elliot Arledge @elliotarledge]

i reverse engineered wispr flow, then used deepseek to fine-tune LFM3-1.7B. the result is on par with the wispr flow models you pay for, and it's free, open source and runs entirely on your mac.

key-up latency is 65 ms on my m4 max vs ~520 ms for wispr flow on the same clips. a 6 minute yap session pacing around my office still lands in about a quarter second.

> rust backend with a native metal engine: each decode step runs as one fused megakernel
> a dflash2 speculative drafter is built into the engine, trained on my local 4 x rtx pro 6000 server
> parakeet transcribes, LFM3-1.7B refines. both stream while you talk, so key-up only waits on the last chunk
> never leaves your macbook, so slow internet or bad ping is no longer an issue
> import your own dictionary of terms for your workflow
> i picked every default myself. it feels amazing out of the box
> adapts to the app you're in: discord is lowercase casual, slack and email are formal, claude code gets clear prompts
> training data: 77k transcripts of public audio and synthetic dictations, each rewritten by deepseek v4.1 flash 1017 following a written spec of what "what you meant" means, in 4 writing modes. none of my own dictations went into it
> blind judge (claude opus) on an hour of lectures: phonon's text preferred 90 times, wispr's 23
> built with claude opus 5.5 [xhigh]

brew install --cask infatoshi/tap/phonon

[@saeid @SaeidSahil9064]

@dhvanil The part where it stopped at 4am to mix its own shades is the best bit. 303 brush marks in one night is a lot of work for a robot.

[@saeid @SaeidSahil9064]

@sanketsahu Ah, that makes sense 🙂

[@saeid @SaeidSahil9064]

@Arnesh_24 100 tokens a second on one m1 core is wild. And just 4 kb of machine code for all of it is the part that got me.

[@saeid @SaeidSahil9064]

@Arnkum_Avar @AarushYadav47 85 to 90% word accuracy is a huge step, and just one minute of setup for each new person makes it feel close to real use.