ManIT retweeted
every team post-training agents has at most a handful of labelled traces, which they also use for eval.
nobody has thousands and the few you have only show the failures you have already seen.
so most synthetic data just remixes them. more of the same.
we went the other way:- extrapolate from your traces and generate the failures that haven't happened yet.
then we trained a model on zero real traces and put it head to head with one trained on real expert data.
the result surprised us.
launching tomorrow. open source + live on the platform 🐒
ManIT retweeted
Introducing Noma: an open decision model.
State in. Calibrated answer out. No generated text.
16 ms per decision
~40x faster than Jev
$0.023 per 1,000 decisions
100% on JevBench easy
Open weights, MPL-2.0.
pip install blackdrome-noma
ManIT retweeted
that viral thread about humans building LLM knowledge bases missed the obvious next step:
what if the agent builds and queries its own knowledge base. no human involved.
every session it runs, it writes to a graph. every session after that, it reads from it before making decisions.
the agent is literally compounding its own intelligence across runs.
vektori inject # agent ingests its own history
vektori recall "what patterns worked in this codebase"
no wiki. no obsidian. no human curating anything.
just an agent that gets measurably better at your codebase every single week.
ManIT retweeted
This is such an important post. We are at a really interesting inflection point with agent memory.
@hwchase17 is right -> memory is just context, collected and injected at the right times.
But what most people miss: even how you represent that context matters architecturally. KV caches, attention states, compressed memory - these aren't implementation details, they are the actual substrate of what your agent "knows."
If that layer is opaque or behind a proprietary API, you lose more than portability. You lose the ability to reason about why your agent behaves the way it does, improve it or take it somewhere else.
Open memory means owning the KV layer. That's what I have been building toward - memory that's inspectable, transferable and not held hostage by any one platform.
ManIT retweeted
Built a complete CPU baseline of the Sumcheck protocol in Rust as part of my coursework.
After discovering ICICLE by @IngoZK, I'm excited to port this to GPU and optimize performance.
Would love to hear from the team!
@yuval_domb @Karthik_Inb @OmerShlomovits
repo link below
ManIT retweeted
So, @suhasasumukh and @heettike are cooking this weekend - $1000 in 48 hours, locked in at the HQ
@localhostIND X @noicedotso wasn’t on my cards for this march.
We’re wrapping Day 1 of the 48 hours and here is the rundown of who’s doing what!
1. @Muskanjain0401 and @aashatwt are building gamified prediction market for friends to dare
2. @10KRotator is building AI agents for expense reports for treasuries onchain. @myStableCorp is a compliance wrapper
3. @notcodesid is building a hyper liquid Algo-trading bot that bets on 5/15 min plays
4. @17Vardani is building a tool that summarises website content visually (think flow charts and mind maps) and auditoryly
5. @7613Perfect and @DsouzaJovian are building an AI video editor with taste.
6. @llmluthor multi agent framework for orchestration and context management
7. @AlexQHunterz Your AI ghostwriter that never sleeps and posts wherever you don’t
8. @vzy010 and @prshbt are building a platform where a token(coin) is minted for a clanker🤖 by typing a vibe and the internet does the pumping - @rotdotrun
9. @Th3lazykid is building a trading bot for @HyperliquidX
10. @itsnishu is building a personalised temogachi assistant
11. @developvthayush is building systems to replace telephones from hotels.
12. @technosh_ does launch vids for start-ups.
13. @PrakharKhatri3 is building a CLI tool to trace & audit autonomous code
14. @vikings_ace and @notacadia are building a no-code node based editor to fine tune AI models
15. @Prem__Mali is pixel perfecting figma to code
16. @probablyaayushi is building an agent to analyse social profiles
17. @hrdktwt is building a chrome extension for bookmarking links and making it intelligent and interactive
18. @anshdixit123 , @Ta_nush & @projektkush are building a creative intelligence tool for D2C/marketing teams
19. @Thevirofficial is building a platform to automate and post AI avatar content for founders.
20. @ArnavvSharmaa & his team are vibe coding and pumping websites.
21. @swarajb is building a Social media for comments to DM
22. @dipanshuhappy is building AI agents that hack for bug bounties (White hat penetration-testing)
23. @AyushSagar31980 mines social media platforms to predict what software people will pay for
24. @whotanish is building an AI pipeline for AI UGC creation
Stay tuned for day 2’s updates!
We’re at T-37 hours.
ManIT retweeted
We built an AI that remembers like a brain, not a database.
Vektori builds a living graph of every conversation you've ever had. Fresh concepts glow bright. Old ones fade, unless you reinforce them.
20+ daily users trusting us with their entire digital memory.
No venture backing. Just pure engineering spite :)
ManIT retweeted
shipped this in 2 weeks. still iterating fast.
using it ourselves daily.
if you juggle multiple llms daily, this is for you.
dm for beta access or reply with how you'd use it
this is an mvp. lots to improve.
@AlexQHunterz @theresidency @enter_delta
ManIT retweeted
they are solving an overlooked problem. Super interesting how this will turn out after sometime.
you shouldn't have to explain yourself to ChatGPT, then Claude, then Gemini all over again
one memory layer.
every AI remembers.
our 30s pitch with @AlexQHunterz
building vektori in public ->
#buildinpublic @theresidency @enter_delta
ManIT retweeted
you shouldn't have to explain yourself to ChatGPT, then Claude, then Gemini all over again
one memory layer.
every AI remembers.
our 30s pitch with @AlexQHunterz
building vektori in public ->
#buildinpublic @theresidency @enter_delta
ManIT retweeted
lfgggg shipped 3 features this week @AlexQHunterz 🚀
balance -> moving fast + not breaking everything
(the patience part is HARD lol)
peep the progress :):):)
@enter_delta @theresidency
