dev striving for something | leaving notes to myself here
Joined May 2021
- Tweets3.2K
- Following671
- Followers581
- Likes8.4K
kshitij retweeted
Finally summarized what I've been sending to friends in DMs View article
secured 1st in weeks 2 and 3 of the qsb challenge :)
week 1 had already made the kernels 6-10x faster, so every improvement was a fight.
thanks @StarkWareLtd @yukonresearch for organising this
kshitij retweeted
said I’d start building more and documenting everything.
so here’s the first one.
I’m building imo (@tryimoo) - a social layer for prediction markets on @solana.
inspired by fomo, pumpfun, and everything i like about consumer crypto.
trying to ship it for @colosseum .
let’s see how far I can take it.
been using opus 5.5 for past 2 days now, it takes me back to the time when opus 4.5 was released
its really good with orchestrating and long running session as the claude team itself says so, it understands what model to use for what case, through my sessions it made choice of using fable 5.1 where it felt like the problem would be solved better by that model and has taken on other work by itself
a lot of it depends on how your harness is set up which i feel like i was able to still get the best out of my subscriptions earlier to this model release, most of those things can still be carried forward like model routing, which model to use for what use cases, which models not to use at all (sonnet and haiku), a lot of it is composition of matt pocock skills such as specs, issues, wayfinder, research and grilling which has optimised my work a lot
its been 2 months since opus 5 was released and half of my subscription was not being utilised to its fullest, mostly i was orchestrating the work through fable by delegating work to opus 4.8, not a big fan of that model but it works if guided right, not the case with opus 5.5 till now, it follows the instructions right, figures out the stuff and i don’t have to particularly monitor it work as i had to do so for other opus models till now
the best part it doesn’t move the session limit that much, i have been running multiple sessions continuously for 24 hours autonomously and my weekly limit moved around 20% and the amount work got done out of it justifies really well (compared to the same amount of work if i’ll try to get it done by fable)
another one
636606729769440499166579950236036751749912014371509557713570027508971809534551913252252094954941974952859310861988904737359709200557919
is a factor of RSA-896
saweis.net/posts/rsa-896.htm…
kshitij retweeted
Submitted Errata for #HackHydra
Errata: an append-only belief graph for agent memory.
→ keeps the evolution of facts instead of overwriting old beliefs
→ lets agents reason about what changed, what superseded it, and what the latest state is
Built on @hydra_db .
simply use pnpm and have nvmrc set up globally to avoid script execution on install
‼️ BREAKING: An active npm supply chain attack has compromised at least 868 packages carrying over 2 billion monthly installs with a credential-stealing worm. Shai-Hulud is back.
It started with the compromise of the GitHub account of the maintainer behind keyv, a library with roughly 127 million weekly npm downloads.
A preinstall hook fires on npm install and drops a stealer that sweeps npm, GitHub, AWS, Kubernetes and Vault secrets, and then spreads to more maintainers.
goal active and fable has landed me 25 PRs till now
utilising it to it’s full potential till it’s here
kshitij retweeted
Update on Metamemory
yesterday, we built a multi-agent coding system
we gave it access to @metamemoryai , then asked it to improve metamemory itself
it spent the day finding bugs, fixing them, and improving our memory engine
yesterday's wins
• Better recall for older memories
• Fixed a memory linking bug at scale
• Improved retrieval ranking
• Killed a feature that looked good in small tests but failed under real load
Recall@10 at 1,000 memories: 0.50 → 0.98
memory has been the core problem for ai agents and we’re solving it @metamemoryai
we are building @metamemoryai , cognitive memory for AI agents
right now every agent forgets everything the moment a chat ends. we fix that, and our memory gets smarter the more it's used
already beating @supermemory on the main benchmark (LongMemEval-S).
now scaling it up + raising a small pre-seed to get there
If you're an investor or just want to nerd out about agent memory, my DMs are open 👇
and also it would be great if you guys can help us or guide us to reach out investors
