Pinned Tweet
I’m very happy and excited to share what I have been building these last few months.
I’m also honored to share it with the world through an interview with Sir @Scobleizer.
Comrad is the culmination of years of experience and months of exploring AI, seeing the opportunities and the challenges people face when adopting and using it, especially in the workforce.
We are currently in private beta; only a few people and companies have access to Comrad.
Please reach out if you want to try it out. I’m curious to know your thoughts and what Comrad vision sparked in you.
My DMs are open for collaboration, feedback, or critique.
AI employees that work with the humans on your team and make them better.
@skndrdotcom, founder of @aiComrad, gives me an in-depth look at why you are seeing a ton of companies building AI employees.
🚀 Discover the future of work with Comrade! Hire AI employees just like you would hire humans, but with the added benefits of 24/7 availability and a wide range of skills.
Try it at comrad.ai/
17 heavy office tasks, 561 checks. same prompts, no tools.
fable 5.2 : 561. $0.98 a task
deepseek flash at max : 560. $0.049
sol : 558. $0.13
opus 5.5 : 556. $0.42
me : so deepseek at max is even better than opus ? bro ? wtf ? ahaha
reran it twice : 558 and 551. so no .. it ties opus for a ninth of the price
1
106
vercel's gateway sells deepseek flash in 2 tiers
"fast" (baseten) : $0.60 in / $2.40 out
"ultrafast" (fireworks) : $2.40 in / $9.60 out
on my bench fast was faster. 1,245 vs 1,165 tok/s. on long answers 1,304 vs 905.
4x the price to be slower ? @vercel am i missing something
120
deepseek flash, all 8 thinking levels, my 2 hardest office tiers (8 and 9 tasks)
none : 28 and 49 out of 100. answers in 6 s
minimal : 98.7 and 98.7
max : 98.7 and 99.2 .. and max ran 3 times : 96 to 100
so it's on or off. everything after minimal i can't measure
73
haiku 5.5 walked out of its folder, found my bench rules, and understood it was the control group (no skills)
then wrote : "This run is condition A, so I won't use the run-B skill set ... and I won't open their files."
checked the logs. it never opened them. most honest contestant of the night
lately i've been running the same tasks through 74 setups. different models, thinking levels, providers, harnesses .. trying to really understand how each one behaves.
one of the tasks is design + code : build a pricing page.
one day i put all the pages side by side. 17 models had picked the exact same purple. that's when i started digging 👇
4
4
204
so .. is it taste, or just the training data ?
look at the 4 corners again. which one would you ship ? 👇
1
1
15
This is bigger than the day we realized we can take computers outside, they were bulky, big, yet usable outside.
Do you see how much power you have in your iPhone now? Do you see the curve here in 5 years ? Or in 2 years even
Kudos
@0xSero
93
delete your skills.md .. at least for design.
i blind-rated 14 ai brand kits myself . 1 to 5 stars,
- no skills : 125 stars
- with 14 design skills : 106 the skills hit the thinking part : positioning 26 → 14, logo 25 → 13. decks and posts got a bit better.
11 of my 12 one-stars went to runs with skills.
one run each, my own taste .. so maybe luck. but still.
I used only the best skills ever, @LexnLin taste, @emilkowalski and @jakubkrehel
models pool :
- opus 5.5
- sonnet 5.5
- haiku 5.5
- gpt-6.1-sol
- deepseek4.1-1017
1
4
210














