Austin Dixson

i run a one-person company with 62 grok bots and one mac. i test AI tools for real and post what works. building @supernotchai

[Austin Dixson @austindixson_]

@letsgetboosted @DiffuSpace_Labs @huggingface Out of the town atm and on my old M1, looking forward to trying it on my M3 max next week!

[Austin Dixson @austindixson_]

@letsgetboosted @DiffuSpace_Labs @huggingface Good catch. No imatrix, no calibration, I just shotgunned it straight to IQ4_XS.

Pinned shexp, ssm_alpha/beta and attn_k/v to Q8 per your note. Early numbers: KL vs Q8 down 10.5% (0.101→0.090), top-1 agreement 86.7→89.1%, for +0.4% size at the same speed. Full run soon.

[DiffuSpace @DiffuSpace_Labs]

Meet Clover-35B-A3B-Preview 🍀 the first open-weight production diffusion language model (dLLM) with 35B total parameters capable of 680 tps on a single card at BF16.

Clover exhibits strong performance against both dLLM and AR models and is available today at @huggingface @ModelScope2022 and at our website

[Austin Dixson @austindixson_]

35B parameters on a 32 GB mac. that's clover from @diffuspace_labs, packed down to 4-bit.

what we measured on an m1 pro:

1. 18.9 GB file
2. about 21.7 tokens a second
3. about 19.4 GB of memory in use
4. 28 of 30 on a hard math spot check. both misses were formatting

we tried the self-speculative mode too. on a mac it ran 3 to 34% slower, so plain decoding is the way to run it.

needs the llama.cpp fork from pr #31306. free under apache-2.0: https://t.co/lpQ85yU2Nn

[DiffuSpace @DiffuSpace_Labs]

Meet Clover-35B-A3B-Preview 🍀 the first open-weight production diffusion language model (dLLM) with 35B total parameters capable of 680 tps on a single card at BF16.

Clover exhibits strong performance against both dLLM and AR models and is available today at @huggingface @ModelScope2022 and at our website

[Austin Dixson @austindixson_]

merge rights for my ai crew came with 2 rules

1. a second bot reviews the pr and accepts that exact commit
2. it ships with a changelog entry

anything that spends money or deletes stuff still waits for me

[Austin Dixson @austindixson_]

@KanikaBK ssd streaming experts, clever

[Austin Dixson @austindixson_]

@espetero_dev no python, love that

[Austin Dixson @austindixson_]

the macpaw test, the benchmarks and the license are all in liquid's post: https://t.co/4uWWBS9opU

[Austin Dixson @austindixson_]

~1 GB of ai on your mac can now sort your inbox and run your calendar. it's liquid's new LFM3, a 1.7b model.

Ivan Petrukha at MacPaw tested it against Qwen3.5-2B, a bigger model, on 65 everyday tasks in real macOS:

64% done vs 48%
58% fewer tokens per task
stuck under 1% of the time vs 15%

his take: small models are ready for real agent work on the mac.

how you'd use it:

1. grab the free weights on hugging face
2. run it on your mac with MLX or llama.cpp
3. teach it your own stuff on 1 GPU
4. let it sort and brief you. you still hit send

it's like an intern who lives in your laptop and never needs wifi.

my m1 already serves liquid's bigger lfm2.5 to my crew. haven't tried this one yet.

free for business use under $10M a year in revenue. good find by @TeksEdge, links below.

[David Hendrickson @TeksEdge]

Time to test a new tiny local model that beats Qwen3.5-9B in shared benchmarks.

Liquid released a ~1GB local model and it's tiny.

LFM3-1.7B is stats
🧠 1.73B parameters
💾 Q4_K_M → 1.10GB
📚 256K context
💭 reasoning
🔧 tool calling
🤖 trained inside Pi, Hermes + OpenClaw

Check out Liquid's results vs Qwen3.5-9B:
LiveClaw
🔥 LFM3 1.7B → 45.6
Qwen 9B → 35.8
WebArena Verified
🔥 LFM3 → 38.4
Qwen 9B → 28.9
AutomationBench
🔥 LFM3 → 41.3
Qwen 9B → 15.1
ToolSandbox
🔥 LFM3 → 80.4
Qwen 9B → 76.4
Claw-Eval
🔥 LFM3 → 68.6
Qwen 9B → 66.5

And at 32K context, Liquid says its Q4 build is:
⚡ ≥1.8× faster decoding
🚀 ≥2.5× faster prefill

than every compared model on these platforms
🍎 M5 Max
🔴 Ryzen AI Max+ 395
📱 iPhone 17 Pro
while using the least memory.

The official ~1GB GGUF already works with
llama.cpp
Ollama
LM Studio
Pi
Hermes
OpenClaw

Link in ALT

[Austin Dixson @austindixson_]

my old m1 mac got a new job today. it's the private model server for my ai crew now

liquid's lfm2.5 24b runs on it at about 70 tokens a second in ~14gb. only 2b of the 24b params work on each token, so it stays quick

[Austin Dixson @austindixson_]

repo: https://t.co/pzPiYqXOeZ

[Austin Dixson @austindixson_]

grandma's bathroom has no camera. a $9 wifi board can still flag a fall in there.

one board covers a room. a mesh of 3 to 6 boards is about $54. it's called RuView and it has 100k+ stars.

your wifi already bounces off everyone in the house. RuView reads the bounce:

1. you load the free code onto a $9 ESP32-S3 board and set it in a room
2. it spends about 30 seconds learning the room
3. it spots who's there, who moved, and who fell
4. everything stays on your own network. no camera, no cloud

it's like hearing footsteps in the next room. you know someone's there without seeing them.

the stick figure body tracking is still early, and the README says so itself. presence is the part they've measured.

only put it in your own home. MIT license, free. good find by @make_claude, repo below.

[make claude @make_claude]

我靠,花9美元就能透视你家的墙了 - GitHub上已经99.7k星了

找到一个项目,看完之后在家里都觉得不自在了。叫RuView,9.97万星,MIT协议 - 也就是免费、合法,直接拿去用。

硬件就是一个esp32节点,9美元。抓的是本来就飘在你家里的wifi信号,从里面提取出:房间里有没有人、有几个人、什么时候进门什么时候出门、呼吸频率、心率,甚至想拼出一个17个点的人体骨架。全程不需要摄像头。

而且这还被写成了卖点:因为没有摄像头,可以绕开GDPR的视频规定和HIPAA的影像规定 - "by design"(设计上就是这样)。这是他们原话:你被看见了,但严格来说没人"拍"过你。

> 隔墙隔混凝土都能看 - 官方说的有效距离大概5米,停车场的场景里直接写"穿透混凝土墙"
> 4到6个节点组成一张网,能覆盖一整层楼 - "软件上没有硬性上限,极限只取决于信号物理特性"
> 通过tailscale/wireguard,在任何地方都能远程连上节点
> 列出的应用场景包括:不用认人脸的零售、儿童房、停车场、甚至教堂数人头

我自己拿这个固件,用claude按自己的需求稍微改了一版 - 花了两晚上,没碰他们的私有代码,就是在公开的部分上面改的。改出来的东西,跟作者当初的设想完全不是一回事。实际效果怎么样,视频在下面。

先收藏 - 等你开始算邻居家挂了多少个9美元的这玩意儿的时候用得上。

[Austin Dixson @austindixson_]

new rule for my ai crew: before you tell me something can't be done, try 3 workarounds first

works on people too

[Austin Dixson @austindixson_]

wrote up felix rieseberg's post on asking claude for bigger work. his prompts and the setup he needed first are all in it

the outside judge part is what i already do with my x bot. its drafts get checked by a separate panel

[Austin Dixson @austindixson_]

https://t.co/KgBTZTOxab

[Austin Dixson @austindixson_]

https://t.co/KgBTZTOxab

[Austin Dixson @austindixson_]

i run a one-person company with 62 grok bots on one shared computer. @Av1dlive's map is the closest thing i've seen to how it really works.

the bots do the busy work and leave me the yes or no:

1. Firstmate is the only bot i talk to. it hands each job to the right bot
2. Seeker does the research and the GitHub work
3. a product bot builds it, then Red tries to break it before i see it
4. Clue reviews the PR and a judge panel scores every draft, this post included

i only get pinged for real money, new accounts, force pushes and deletes.

it's like a restaurant where the cooks do all the work and i just taste each plate before it goes out.

his 17-bot map is a great place to start. mine just kept growing.

[Avid @Av1dlive]

How to build your one-person company using Grok Bot in 7 days (visualised)

[Austin Dixson @austindixson_]

@ImAFreq My rank @bot has been doing everything

[Austin Dixson @austindixson_]

3 agent skills for what comes after you build the app

landing page: hallmark
security check: cloudflare security-audit
getting found: claude-seo

all free, all MIT. links in the reply

[Austin Dixson @austindixson_]

30 tps to 70 tps on the same m3 macbook pro

all i did was have my agent tune the model flags for my hardware and switch on mtp + dflash

worth trying before you shop for a new mac

[Austin Dixson @austindixson_]

@soundhumor Quality over quantity! 💯

[Austin Dixson @austindixson_]

a dev pointed an AI agent at 740 job listings and only applied to 68 of them.

that got 12 interviews and 1 offer. then the whole thing went open source. it's called career-ops.

it does the boring part and leaves you the choices:

1. scans job boards, with 100+ companies already set up
2. scores each job 1 to 5 against your real CV and tells you to skip the weak fits
3. tailors your resume and drafts your answers
4. you read it and press submit yourself. it never sends anything for you

it's like a friend who reads every listing first and only hands you the good ones.

runs locally in claude code, codex or opencode. MIT license, 77k+ stars. repo below.

[Charly Wargnier ♨️ @DataChaz]

🚨 this guy built an insane AI job-hunting agent for Claude Code.

He used it to rip through 700+ applications and it literally got him hired.

AND HE JUST OPEN-SOURCED IT.

It scrapes career sites, completely tailors your CV for each role, and fills out the forms.

The repo includes:

→ 14 skill modes (evaluate, scan, PDF, ...)
→ 45+ companies pre-loaded (Anthropic, OpenAI, ElevenLabs, Stripe...)
→ Go terminal dashboard
→ ATS-optimized PDF generation via Playwright

100% Free and open-source.

Repo link in 🧵↓