curr @Covalent_HQ | ex @encifherio @0xStackr @pushchain | @0xpolygon Fellow’22 | @Arsenal ♦️
Kolkata
Joined January 2021
- Tweets1.7K
- Following743
- Followers1.3K
- Likes2K
Pinned Tweet
I’ve been diving deep into memory engines for LLMs — specifically how far we can reduce token usage while improving long-context accuracy.
So I built a memory layer for LLMs and ran it against the full LongMemEval-S benchmark:
Thanks to @supermemory and @mastra for their benchmarking framework which i used a reference to set up mine
500 questions
gpt-4o actor + gpt-4o judge
Best-of-3 judging passes
Results:
82.22% unweighted accuracy
80.80% micro accuracy
Per-category breakdown:
• SSU: 92.86%
• SSA: 94.64%
• SSP: 70.00%
• KU: 88.46%
• TR: 76.69%
• MS: 70.68%
The system was strongest on single-session factual recall, and weakest on multi-session aggregation + temporal reasoning
archis.eth 🦇🔊 retweeted
Instinct would take 1 hour to automate my AI canvas vid gen by running multiple browsers, which we turned into a 5-minute single-text automation.
My usual process:
1. Doomscroll for a viral UGC video in my niche.
2. Download it and isolate the human performance
3. Transfer the motion to my avatar
We connected the video-generation steps behind a single API that Instinct can call.
Now, once I find a reel, I send the link and describe the avatar I want.
The video comes back in ~5 minutes.
The future is going to be apifying businesses so these agents can operate faster.
RT @Sampriti026: Astra made this launch video with only a couple of iterations.
If you spend hours scrolling for reference videos, you ca…
While everyone’s chasing for virality - you chase for performance
Analyse and find videos that would actually help you onboard customers
Try out today at : ugcs.farm/mcp
We are launching Claude for finding videos that converts.
The first tool that lives inside your Claude and finds the highest converting video in your niche.
Not the most viral one. The one that's actually making people buy.
It goes through comments, engagement, hooks and format to figure out why a video works.
Then it adapts that exact winning format for your product.
Try it here: ugcs.farm/mcp
For free plan, reply "ugc". Or drop your product link and I'll run it and post the results.
archis.eth 🦇🔊 retweeted
I’ve been diving deep into memory engines for LLMs — specifically how far we can reduce token usage while improving long-context accuracy.
So I built a memory layer for LLMs and ran it against the full LongMemEval-S benchmark:
Thanks to @supermemory and @mastra for their benchmarking framework which i used a reference to set up mine
500 questions
gpt-4o actor + gpt-4o judge
Best-of-3 judging passes
Results:
82.22% unweighted accuracy
80.80% micro accuracy
Per-category breakdown:
• SSU: 92.86%
• SSA: 94.64%
• SSP: 70.00%
• KU: 88.46%
• TR: 76.69%
• MS: 70.68%
The system was strongest on single-session factual recall, and weakest on multi-session aggregation + temporal reasoning
The more interesting part was token efficiency.
Compared to the full-history Chain-of-Note baseline, on the same 500-question dataset and prompt template:
• ~3,200 input tokens/question vs ~14,600
• 78% fewer input tokens
• +22pp higher accuracy
The architecture ended up being a mix of retrieval, reranking, and fallback reasoning:
Retrieval side:
• Hybrid dense + BM25 fusion
• Heuristic + cached LLM query expansion
• MMR diversity
• Recency bias for knowledge-update questions
• Time-anchor bias for temporal questions
• Optional cross-encoder + LLM listwise rerank
Answer side:
• Chunk-level hits rolled up into session-level evidence
• Retrieval-confidence scoring
• Low-confidence fallback that injects raw transcript from top sessions
• Pluggable single-shot or self-critique synthesizer
archis.eth 🦇🔊 retweeted
Hyperliquid Markets Weekly Heatmap
DeFi: +11.4% ✅
Stocks: +11.2% ✅
Layer 1: +4.0% ✅
HIP-3: +3.8% ✅
Indices: +3.3% ✅
Meanwhile, AI: -17.5%. RWA: -5.3%. Commodities: -4.5%.
DeFi and Stocks running together while AI gets wiped
The market isn't rotating out of crypto. It's rotating out of the narrative - How are you positioning yourself?
Source: hyperliquid.goldrush.dev
archis.eth 🦇🔊 retweeted
CompliantLLM (@FiddleCubeAI) detects data leaks into any third-party GenAI tools used across your company. It identifies GenAI-specific attacks to catch breaches in both approved and unapproved AI workflows.
ycombinator.com/launches/OBo…
Congrats on the launch, @kaushik_himself & @nupoor_neha
archis.eth 🦇🔊 retweeted
I can literally buy TSLA stock using TRUMP memecoin XD
That too even privately
Using @encifherio Thanks to @JupiterExchange
Played around with @coinbase x402 protocol and here’s what I built recently
Codey ✨ - unlock true pay-as-you-go in your AI powered code editors. Use your wallet as API key, top-up USDC & pay-per-call. Pick your context, let smart agents pick the best model for price & accuracy. You control the budget, no flat subscriptions, no wastage.
Where to next
- Currently wallet is stored in VSCode SecretStorage, can be extended to use CDP wallet with session keys and permissions
- Democratising model choice, anyone can host their service and provide api endpoints and set their fees
Check out the demo, DM to get access and try out
cc @CoinbaseDev
archis.eth 🦇🔊 retweeted
🚀 We've raised $4.2M to modernize financial operations for crypto companies!
Led by @1kxnetwork with @ycombinator, @GoodwaterCap, @DCGco, and @SevenXVentures
Just like Brex transformed traditional business finance, we're making crypto ops seamless.
archis.eth 🦇🔊 retweeted
Introducing: dsheets.new 👩🔧
Decentralized spreadsheets made for the global economy & open Internet
archis.eth 🦇🔊 retweeted
Give account recovery using email guardian a try -> recovery.zk.email
Use passkey for a smooth experience 🚀
We keep pushing Account Recovery features - and we have a new demo to show to you using Passkeys and EIP-7702 🎬
Delivered by: @0xdhruva 💌
