Your Life at Every Level of Running AI Locally (Laptop to 4×3090)

Comments · 285

• TH

  @TheDevPsyche · 2 weeks ago · pinned

  So: what level are you actually on, and what&apos;s in the machine?<br><br>Post the specs. I want the range — someone is running a 3B on a ThinkPad from 2014 and someone else has a 20-amp circuit they had to explain to a professional, and both of you are in this comment section.<br><br>And be honest about one thing. You told somebody it was for fine-tuning. How many things have you actually fine-tuned.

  15

• PI

  @piexlpalace · 2 weeks ago

  This video definitely just warmed a room by a few degrees

  110

• WE

  @weykn · 2 weeks ago

  no, the limit is the money

  123

• WE

  @wesss9353 · 2 weeks ago

  There is copper in Flock spying cameras.

  54

• KA

  @kaczynski2333 · 12 days ago (edited)

  That&apos;s why I&apos;ll have commercial 3-phase in my next house. &nbsp;I&apos;ll gut it down to the dirt, do a very expensive renovation, and then install 3 racks. <br><br>It really is about the money.

  11

• CH

  @ChilenonetoYoutube · 2 weeks ago

  With my budget there is no way this affects me any time soon.

  33

• 00

  @00000a0009 · 13 days ago

  <a href="https://www.youtube.com/watch?v=0YdsvCJXeic&amp;t=226">3:46</a> I always say &quot;I&apos;m just 3TB short&quot;!

  8

• YA

  @yankoaleksandrov9050 · 10 days ago

  Great ladder. The axis I&apos;d add next to the &quot;most people are hybrid&quot; point is uptime.<br><br>Almost everyone benchmarks tokens/sec, but once you actually run agents on the box — scheduled jobs, overnight tasks, anything long-lived — the binding constraint stops being peak throughput and becomes a machine that never suspends. A level-zero laptop that sleeps or reboots silently kills a job; the 4×3090 frame is brilliant for a burst but burns idle watts just staying available, and that 220W power-cap trick matters twice as much once the thing is up 24/7.<br><br>What has worked for me is splitting the two design points: a cheap low-power always-on host (~7–15W, Jetson Orin Nano-class or a small mini-PC) that owns the scheduler, session state, logs and memory, while the big GPU box only wakes for heavy inference. Same hybrid move you point at, just on the machine axis instead of the model axis. (Disclosure: we build ClawBox, which is exactly that always-on box — but the pattern works with hardware you already have.)<br><br>Did level six / &quot;the way back down&quot; end up being mostly power and noise, or the realization the workload fit somewhere smaller?

  2

• AR

  @arnverbelen4648 · 2 weeks ago

  I hate to be the one that stops when people say it&apos;s just a worse ChatGPT😢

  15

• PY

  @pycontiki · 2 days ago

  These videos are comic genius!<br><br>Bravo my dude 🙌🏼🏆💯🫡

  1

• MF

  @MFEeee · 2 weeks ago (edited)

  Started with an Intel arc a770 with 16GB. Upgraded to an Intel arc pro B60 with. Then purchased an Intel arc Pro B70 with 32GB. Then bought a second Intel Arc Pro B70. Each GPU runs its own instance to avoid slowdowns. Running qwen3.6 35b a3b as MoE. And qwen3.8 27b as dense model. Both at q6 with 65k context on one B70. Running comfyUI video generation on the second B70. Running comfyUI image generation on the B60. Have 128GB of system VRAM and Intel core ultra 7 cpu. Everyone is happy.

  16

• DJ

  @djfremen · 2 weeks ago

  Loved this. Excellent SF and animations

  1