tweets about typescript, coffee, and everything in between. building @drinkclutch | @nahruntime | @nebprotocol | @astracollab | @cogmem | @wnstndotspace
San Francisco, CA
Joined October 2020
- Tweets7K
- Following1.2K
- Followers963
- Likes16.3K
You can use @nahruntime for free with the minimax 3.1 model 👀
@MiniMax_AI 3.1 Flash is free on NAH!
nah.astracollab.app
npm i -g nah-ai
Elias Wambugu retweeted
@MiniMax_AI 3.1 Flash is free on NAH!
nah.astracollab.app
npm i -g nah-ai
Stealth model reveal 🐰
Space Bunny was OpenRouter’s #1 model for 10 straight days, processing 56.7T tokens! The most of any stealth model we’ve run.
Space Bunny hopped off OpenRouter on Oct 5. It's back today as @MiniMax_AI's M3.1 Flash, and it's free for the next 72 hours.
Used the stealth model to build the coding agent and framework and now we know what we were using 🚀
Now the model is free to use for 72 hours with the coding agent @nahruntime
@MiniMax_AI 3.1 Flash is free on NAH!
nah.astracollab.app
npm i -g nah-ai
For some reason sessions with @nahruntime stop at the 1 million token mark so I’m working on extending that
Elias Wambugu retweeted
we mapped every marketplace on the internet and made the largest dataset
10,000+ of them. deliveroo, airbnb, swiggy, vinted, and everyone in between
for each, we shopped it like a real buyer and captured
- every screen from search to checkout
- every experiment and feature flag they're running
- every tracker and event that fired
- every dark pattern they use till checkout
and @refix_AI has already been trained on all of them
because the best growth ideas are already live on your competitor's checkout
comment & tag @refix_ai and i'll dm you a link👇
Elias Wambugu retweeted
Maybe not a strong argument but it’s supposed to work well with free/lower end models for those that can’t afford AI
Stronger argument is the memory aspects. I made a 4 tiered memory that doesn’t use a vector DB and I think it’s pretty cool @cogmem. cogmem.astracollab.app/docs/…
Elias Wambugu retweeted
nah’s memory isn’t RAG bolted on.
Four tiers (L0 tensions → L1 hot → L2 recall → L3 archive)
Recall is deterministic, meaning no model in the loop. The prompt keeps a light index, and full memory text only loads when you name something missing from the transcript.
Elias Wambugu retweeted
Anything really. You can write tools with human-in-the-loop approval. Like let’s say you are building a trading bot you can have certain trades require human approval before executing or have it send you a text etc.
Right now we’re trying to get it good with free models
Elias Wambugu retweeted
Right now I have it working on bug fixes for itself, @astracollab did a migration and built git @nebprotocol and from @mastra to @nahruntime so all AI is built with that framework
Elias Wambugu retweeted
it shouldn’t take more than 10 mins to setup analytics
canopy.wnstn.space/docs/inst…
Elias Wambugu retweeted
Most of coding work for all my projects is now done by @nahruntime and for heavier tasks @opencode
Elias Wambugu retweeted
this was supposed to be a quick little 3 day build for an internal analytics platform but I decided to make it public
canopy.wnstn.space
Elias Wambugu retweeted
A fact the user plainly stated must never depend on a model agreeing to store it.
That’s why Cognitive Memory runs deterministic patterns first, model second.
Full rules + refusals:
cogmem.astracollab.app/docs/…
this was supposed to be a quick little 3 day build for an internal analytics platform but I decided to make it public
canopy.wnstn.space
Most of coding work for all my projects is now done by @nahruntime and for heavier tasks @opencode
started building my own discord alternative (hive-os.space) to host my community and already have a few people.
hive-os.space/invite-link/6j…
free to join 🙂↔️
