VelVet @finality

Stop dreaming & keep Building | @allezlabs | Former avocado farmer youtube.com/@ConsoleLog_Gm Joined July 2016

Tweets

14K
Followers

197
Following

2K
Likes

10K

VelVet @finality

3 days ago

- One shot - No manual edits afterward ~3m32s to generate the map ~52k tokens used DeepSeek Harness: github.com/deepseek-ai/de… Archify: github.com/tt-a1i/archify

0 0 1 71 1

View Details

Just tested DeepSeek V4 Flash in the new DeepSeek Harness I also discovered archify, a skill that turns a codebase into a visual architecture map. Gave it a try and asked it to explain the Harness itself. Pretty satisfied with the result for a first pass:

1 3 11 2K 10

View Details

VelVet @finality

4 days ago

@PrimeIntellect Prime Agent: github.com/PrimeIntellect… @_can1357 OMP: github.com/can1357/oh-my-… @NousResearch Hermes Agent: github.com/NousResearch/h… @jasonkneen Lazar: github.com/jasonkneen/laz…

0 1 0 265 0

View Details

VelVet @finality

4 days ago

Although that may be false soon as models are getting better & cheaper. I’m still exploring this, I’d be curious to hear if anyone sees this differently. Sources:

1 0 0 19 0

View Details

VelVet @finality

4 days ago

@PrimeIntellect ’s announcement of Prime Agent got me looking more closely at self-improving harnesses. I wanted to understand what self-improving really meant, so I compared 4 harnesses: Prime Agent, OMP, Hermes and Lazar. 🧵

1 1 1 210 2

View Details

VelVet @finality

7 days ago

@WinterArc2125 @dexhorthy I agree and made a similar tool last week that looks pretty much like yours I think node workflows are really great to visualize any workflow !

0 0 1 13 0

View Details

Allez Labs @AllezLabs

a week ago

Monthly @kamino Lend Insights - July 2026 Following June's SOL annual low, supply contraction eased in July; boosted by four new markets launched and @onrefinance's Market jumping 26% to new highs. Continues

4 4 17 1K 1

View Details

VelVet @finality

a week ago

@pivi___ Same !

0 0 0 20 0

View Details

VelVet @finality

a week ago

I reran the same benchmark after the last @polygres update: Before → after Availability: 79.4% (131/165) → 100% (165/165) Evidence recall @5: 62.4% → 80.6% Luna-judged accuracy: 47.0% → 98.0% Not bad @daleverett ! Curious to test it on a larger, harder data set next

VelVet @finality

2 weeks ago

Tested Polygres on 165 second-brain questions with GPT-5.6 Luna. Luna only: 24.8% Local search: 42.9% Polygres Hybrid: 47.0% When it returned context, Polygres reached 59.2% vs. 60.7% with the full library, using 94.5% less context. But availability varied from 58% to 79%.