Ecommerce's most intelligent recommendation engine. Powered by dozens of AI models.
Joined November 2024
- Tweets150
- Following5
- Followers663
- Likes187
Openrouter support brings any model to the competition 🔓
Submit your artifact using the guide below:
github.com/bitrecs/bitrecs-v…
Bitrecs retweeted
Novelty Search tonight.
2 protocol upgrades to Subtensor.
Conviction + TAO Flow V2
2 subnet teams building on Bittensor.
\\ @cacheon_ai :: #SN14
\\ @Bitrecs :: #SN122
Live via Bittensor Discord
Hosted by @const_reborn
9PM UTC / 5PM EDT
We're building a live ecommerce optimization loop!
do {
miners improve artifacts against current evals
top artifact serves recommendations to customers
anonymous shopper signals recorded
use data to feed/generate new evaluations
} while (true);
122.
This Post is from a suspended account. Learn more
Bitrecs Weekly Update 3:
- Evaluation set #2 is live and the top winner is revealed and currently decaying
- @Bitcast_network campaign wrapping up and we saw lots of great posts
- Welcome @taodotcom as validator on Bitrecs!
See below: ⬇️
Our evals are in place to verifiably benchmark against SOTA in LLM recommendation. Targeting parity by H2.
Plus customers get the Bittensor miners' continuous optimization on top.
122.
Bitrecs (122): Prompt evolution for online shop recommendation. Using the IM to evolve prompts rather than models or agents. Very nice, well constrained IM domain, miners can't overfit, and with query temperatures fixed they can do copy resistance very easily (no variance on evaluation)
Going to use Chutes for cheap inference, good idea. Should see them be SOTA in shop recommendations if done right.
Also first ever prompt evolution strategy on Bittensor.
If you want to mine @Bittensor but don’t know where to start with high perf VPS servers or bare metal try Bitrecs!
(122,122)
Time to re-dive in @Bitrecs (SN 122) a subnet i'm holding since 0.0017t, one of my best performing picks in my portfolio.
MARKET
In essence, the e-com market is expectec to reach $ 8T by 2027. Shopping RECOMMENDATIONS ( the little frequently bought together section when your shopping) is something carefully crafted by companies under the hood to show you precise side items to bump up your average order value.
Recommendations are responsible for approx 31% of a store's revenue! A store NOT having a good rec engine is leaving serious $ on the table, sometimes without even knowing.
WHAT IS SN 122
Miners work on a simple ARTIFACT.YAML config file (literally they just play in a file on a terminal)
they need to optmize a prompt, what model the Ai uses to do recommendations, and the LLM parameters (see claude code visual example)
This is a zero-compute subnet for miners. No GPUs, no servers, no electricity costs. You're competing purely on prompt engineering skill and understanding what makes good product recommendations. Winner takes all means only the #1 artifact gets deployed to live stores, so it's basically a prompt optimization tournament.
Validators run your yaml file against the benchmark : Amazon RecSys 2023 set
and the winner gets his recommendation configs displayed to the actual shopper via bitsecs widge app that a store downloaded and pays.
Revenue
VERY simple, store owners have 4 pricings ranging from 0-200$/month so it's simple. Market the widget, show improved average order value, and that bitrec's app ACTUALLY drives more sales
increases hype --> store owners buy it in 1 click via shopify or woocommerce --> boom done, integrated.
Website behaviour by shoppers is actually sent back to the miners with detailed reports (handled by the subnet) so miners can have feedback and re-optimize their yaml file before resubmission.
Conclusion
I genuinely believe this subnet is worth arround 0.007 to 0.01. The team has never quit, they're re-iterated their incentive mechanism, do regular code updates and i've personnally had a call with the team. Great people.
I am looking to sell some subnets and buy the dip on this one.
$TAO #bittensor
This is the eval to watch:
Recall@K is an industry standard evaluation which asks given an input X and a target Y, does an evaluation system return Y in less than K recommendations when prompted with X?
We're using this today to guide the future of ecommerce systems. bittensor:native
Bitrecs retweeted
Over 5mo ago, @const_reborn said we should see @Bitrecs become SOTA in shop recommendations.
It looks like that’s already starting to happen.
Bitrecs tested miners on the Amazon All Beauty ecommerce dataset, measured with NDCG@10.
Their bittensor-powered recommendation models are already scoring higher than some published SOTA results on that benchmark:
> gemma-3-27b-it: 0.20
> MiMo-V2-Flash-TEE: 0.13
> published SOTA: 0.06–0.09
The interesting part is that the models are doing this without being specially trained on that exact task first. They’re using LLMs to rank products directly.
They started with a smaller curated set of 100 products. Next is:
100 → 500 → 1,000 → full catalog
Bitrecs could become one of the strongest AI recommendation engines for ecommerce.
Bittensor | $TAO
Like I’ve always said, we need more subnets with real products solving real-world problems for everyday users.
To be honest, I didn't pay much attention to SN122 @Bitrecs until the Stitch brief, but now I’m actually intrigued.
What they’re doing is, helping online stores recommend better products to customers using AI.
Ever wondered why you go on Amazon and see products you actually need being recommended to you?
That’s because Amazon has one of the most powerful recommendation systems in ecommerce, but it’s centralized.
With @Bitrecs, a regular store owner doesn’t need to build their own AI recommendation system. Bitrecs gives them access to enterprise-grade recommendation AI.
And not just any recommendation AI, one powered by a network of miners competing to create better recommendation models and prompts, meaning it keeps improving over time.
With Bitrecs, merchants can:
> add AI product recommendations to their store
> use modern, clean recommendation layouts
> track performance through analytics and dashboards
> choose which AI models power recommendations
> install it easily
Any merchant or business owner can use Bitrecs. They’re already driving value for 100+ merchants and are targeting 1000+ merchants.
Their miners are also scoring high on NDCG@10, which means they’re moving closer to SOTA-level recommendation performance.
Ngl, I love seeing deAI and Bittensor being used by everyday users like this.
Keep cooking, Bitrecs.
Recent RecSys research keeps landing on the same conclusion: prompt design matters more than model size for recommendation quality.
That's the thesis of Bitrecs V2 — a prompt-evolution subnet on @Bittensor.
Bitrecs extensions capture all relevant context to help models build their 'world models' and offer our customers feature complete recommendations, upsell and cross-sell solutions which seamlessly slot into their ecommerce stores.
Our extensions capture user intent, context and signals automatically and feed it to the Bittensor subnet winner's prompt in real time. Updated daily.
122.