iiinova — where your digital self keeps evolving

Reflection AI's Beam 1.0 is pretty great. I just wish I could consistently use it. Something is off about their setup Not sure how one place says I'm nowhere near my limits, but another has me over... it's not at all clear what the limits actually are.

000 0

New day. New MoE spread-weight model to test. Today I ran the new NVIDIA Bolt through my custom benchmark (built around common homelabbing tasks). Also included last nights run with Beam 1.0 for funsies Qwen 3.6 35b a3b continues to reign supreme.

060 0

I normally don't go this hard on new releases, but it's been an especially great month for books. Finished Exit Party (4/5) & Hollywood Ending (4.5/5) prior to starting up Victorious yesterday. It's been such a joy getting re-immersed into V.E. Schwab's Villains series. #BookSky

060 0

Joshua White @jrw14.whnc.me · 14/10/2026

Okay Beam! I see you!!! Wasn't the smoothest start, but now that we're rollin... Reflection AI might have a banger on their hands.

060 0

Ope... I stand corrected. One test through LiteLLM did finally go through. Only took 553 seconds...

000 0

Joshua White @jrw14.whnc.me · 14/10/2026

Ran a quick experiment with the new LFM 2.5 24b a2b this morning. It's a speed demon on my 9070XT/9800 X3D/32 GB RAM machine... by far the best tps I've seen from a local model. But. (why is there always a "but"?) What's the point if all the tps is unnecessary? So much thinking. Not enough doing.

390 0

Joshua White @jrw14.whnc.me · 10/10/2026

The amount of joy this just gave me cannot be quantified... 😂 Betting coding? Nah... just let me nerd out. Thanks @sethkarten.ai @primeintellect.bsky.social

040 0

Joshua White @jrw14.whnc.me · 07/10/2026

I'm going to wait a few more days to truly celebrate... but Qwen 3.6 35b a3b is impressing me a lot right now. Full 256k context, 70+ tps supporting up to 4 concurrent streams off my gaming PC... and I just ran a couple quick coding tasks that it handled great. Also, not used to models so... terse?

5120 0

Joshua White @jrw14.whnc.me · 07/10/2026

Yesterday's experiments with Strata and Qwen 3.8 Next made me realize what is possible with local inference. I ran into some issues that were a bit disappointing with Strata last night, so I decided to try a similar approach, with a smaller MoE model to address concurrency. DIDN'T EXPECT THIS!

1100 0

Joshua White @jrw14.whnc.me · 06/10/2026

Bonsai 27b 2 vs. Qwen 3.8 Next Flash IQ1 (Strata) An LLM entirely in GPU vs. one spread over GPU, RAM, SSD and even the processor doing some worok. This is my custom benchmark, so of course YMMV. Honestly, not too revealing. I don't think I'll be able to say how good Strata is until I just use it.

000 0

Joshua White @jrw14.whnc.me · 05/10/2026

Beam is here. Reflection AI's new 501b MOE model. A promising addition to the open weights space from an American lab. The numbers alone aren't enough to make me want to switch from GLM 5.3 Flash, so we'll have to see how it actually performs. reflection.ai/blog/introdu...

2170 0

Joshua White @jrw14.whnc.me · 05/10/2026

Well, this is a new record...

000 0

Joshua White @jrw14.whnc.me · 04/10/2026

A few things of note from this, that shows what kind of day its been: - Finished a months-long TUI project earlier today. yay. - Mercury crossed 1000 tps for me for the first time. Yay! - LFM2.5 is my local background/memory model. If it's getting ANY cache at all, that's the sign of a bad day.

000 0

Joshua White @jrw14.whnc.me · 03/10/2026

Wiring up a custom e-ink dashboard for the homelab. 🤓

120 0

Joshua White @jrw14.whnc.me · 02/10/2026

It's been a busy year.

000 0

Joshua White @jrw14.whnc.me · 30/09/2026

I recently hit a new personal best at the scale, which made me go digging through old photos. I'd say I barely recognize who I was four years ago, but body-dysmorphia is a real pain in the ass, and the first pic remains all I see. Still, losing 256 pounds has been absolutely life changing.

4570 0

Joshua White @jrw14.whnc.me · 29/09/2026

Finally got around to getting Excalidraw set up on the server. There's something quietly endearing about prompting an agent in Hermes with very little ("a drawing inspired by the movie Interstellar") and just seeing what happens.

Joshua White @jrw14.whnc.me · 28/09/2026

Over the weekend, I split up some inference over my potato M1 Mac Mini 16 GB and 9070 XT powered PC to make Sub/Wave entirely self hosted. Gemma 4 E4B is surprisingly great at writing the DJ scripts, and Breeze TTS 2 is mind-blowing-ly good at TTS. Sure it could be faster via API...

3141 0

Joshua White @jrw14.whnc.me · 28/09/2026

I've been using Jev now for a week or two (however long its been since it came out). I've spent all of...

Joshua White @jrw14.whnc.me · 25/09/2026

Voice upgrade. Breeze TTS 2 is very good. Set up a bridge between Sub/Wave (server) and Breeze (running on a pc previously set up for gaming) with hard-coded voice settings. It takes 3x as long as Kokoro running in the same container, but since it's got 3-5 minutes (song) it isn't noticeable.

051 0

Joshua White @jrw14.whnc.me · 25/09/2026

I've been working on setting up a new TTS engine (Breeze 2 is incredible, but still optimizing it). In the meantime, this is what the DJ transitions sound like. Honestly, pretty freaking good for Kokoro's 82m STT engine (af_heart). Completely automated.

061 0

Joshua White @jrw14.whnc.me · 25/09/2026

Spent a little more time on BubbleRadio tonight. Added support for themes, album art, dj avatars, dynamic re-sizing, and cleaned up several bugs. Including reducing resource usage pretty significantly. TUI's are magic. (Sub/wave, Go, Charm's Bubbletea)

030 0

Joshua White @jrw14.whnc.me · 24/09/2026

Had some fun with @charm.land's Go goodies tonight and made a custom TUI audio player for SUB/WAVE complete with real-time spectrum analyzer. Linux distro: Endeavour OS Shell: Noctalia WM: Niri Charm goodies: bubbletea, lipgloss, bubbles, harmonica

262 0

Joshua White @jrw14.whnc.me · 22/09/2026

It's a small thing many might not see the value in, but Sub/wave, the FOSS, self-hosted, AI-powered internet radio station is still impressing me days later and is amongst my very favorite things I've put in my homelab.

000 0

Joshua White @jrw14.whnc.me · 21/09/2026

The sub-500ms ttft I was seeing at first didn't hold under continual agentic workflows, and tps, while respectable at 31, just don't make up for how much thinking this model does. That said, pretty happy with it's caching, ha. Though I suppose it matters a lot less when I'm running it locally.

020 0

Joshua White @jrw14.whnc.me · 21/09/2026

Eating some crow tonight. Remember that coding harness showdown I ran about a week ago? I tried it again tonight. Just one harness (Pi), but switched out the model from GLM 5.3 Flash to Bonsai 2 27b. And, it wasn't pretty.

170 0

Joshua White @jrw14.whnc.me · 19/09/2026

My biggest question going into setting up Sub/Wave today was what token consumption would look like. I wanted a "best" baseline, so turned everything on; agentic song picking, aggressive dj segments, medium length inserts, etc... And after a LOT of hours left on, I'm kinda shocked < 1 million tks.

000 0

Joshua White @jrw14.whnc.me · 18/09/2026

I do a lot of cool things with my home lab that mostly revolve around self-hosting and data sovereignty. But every once in a while, you just need a new toy to play with. Sub/wave might be the coolest thing I've put on my server yet. github.com/perminder-kl...

041 0

Joshua White @jrw14.whnc.me · 14/09/2026

Lets make this happen, people!

0121 0

Joshua White @jrw14.whnc.me · 13/09/2026

May 2025: I moved 500 miles from Ohio to North Carolina for work. My wife and daughter stayed in Ohio for another month to finish out the school year. Needing something to do, I bought a $30 Raspberry Pi to experiment with and set up Pi-Hole on. I'd never used Linux before. 16 months later...

0150 0

The overloaded Hermes agent got back to its real job and helped me look at/understand the results. Thought you might like its paragraph on Polytoken :)

120 0

Okay so what'd we learn... @polytoken.dev made the best looking one-shot tetris clone, and it wasn't particularly close. By a landslide. But it was also the slowest, yet 2nd most efficient. Hermes isn't built for this. Which, to be fair; unsurprising. omp is a monster that I do not understand.

3294 0

<<<HARNESS THROWDOWN RD.3>>> Polytoken vs. Hermes Same-ish as the last two rounds, mostly. Hermes has my ENORMOUS system prompt (31k!!) and skill tree to deal with, as I typically don't use it for coding work. Polytoken Tokens/Time: 142,379!! / 11:39!! Hermes Tokens/Time: 1,214,642!!! / 2:54!!!

230 0

<<<HARNESS THROWDOWN RD.2>>> omp vs. Crush Model: Same GLM 5.3 Flash Global AGENTS prompt: Same Skills: Same Prompt: Same (Specific rules, expected deliverables, but basically; build me a self-contained, playable Tetris). omp Tokens/Time: 143,734!!!! / 4:43 Crush Tokens/Time: 600,866 / 3:19

020 0

Joshua White @jrw14.whnc.me · 09/09/2026

<<<HARNESS THROWDOWN>>> Pi vs. Prime Agent. Model: Same GLM 5.3 Flash Global AGENTS prompt: Same Skills: Same Prompt: Same (Specific rules, expected deliverables, but basically; build me a self-contained, playable Tetris). Pi Tokens/Time: 715,975 / 3:28 Prime Agent Tokens/Time: 433,004 / 3:56

3161 0

Peaceful mid-vacation morning out on Harris Lake. Abraxos sails again!

000 0

Joshua White @jrw14.whnc.me · 08/09/2026

Today I learned that there is a GNOME-adjacent launcher for Android called Lynx (natch). Of course it's the most "at home" I've felt on this os since migrating from iOS a few days ago.

041 0

Joshua White @jrw14.whnc.me · 07/09/2026

One year ago I barely could restart (Debian) Linux. I was FULLY in Apple's eco-system for 20+ years. Now? Custom setup of EndeavourOS w/ Hyprland on a Thinkpad X1 Carbon. Fairphone 6+ running e/OS. Writing my own custom android app to connect to Hermes on my homelab. What a time to be alive.

Joshua White @jrw14.whnc.me · 06/09/2026

I've had an Android phone for three days, and I just finished screwing around and making my own custom app for e/OS that connects over headscale to my homelab, allowing me to chat with my Hermes agents.

Joshua White @jrw14.whnc.me · 04/09/2026

First non-Apple phone since the original iPhone. It’s a Fairphone Gen. 6+ running e/OS. Hooray for repairability and data sovereignty.

Joshua White @jrw14.whnc.me · 02/09/2026

Poor little guy is _not_ wild about my ~12k image/video import and transcoding happening right now.

000 0

Joshua White @jrw14.whnc.me · 29/08/2026

Learned some things today... - /loop is magic - I was wrong to leave oh my pi last month. It's so much better for sub-agents - Don't change LiteLLM model names in the middle of busy part of the day and expect things to not break - four tiled omp's at once... my first time. That was, a lot. 😂

070 0

Joshua White @jrw14.whnc.me · 28/08/2026

Cue existential crisis....

120 0

Joshua White @jrw14.whnc.me · 27/08/2026

One of the best page turns in all of literary history.