Fast retweeted
Big update on Pulsar Coffee Meetups. Rome has a new date. ๐ฎ๐น
Saturday, August 1, 11:30 AM to 1 PM.
Hosted by Alex Radu (@Alex__Radu), Pulsar CEO, and Jessy (@13yearoldvc), CSO at @fastxyz.
xitter.cf/PulsarMoneyApp/status/โฆ
Wed July 15, Lisbon ๐ต๐น
Come meet our CMO @kivanarpeev over a coffee (galรฃo, anyone?)
Tell us about your dream money app. Learn why we build with Europe in mind, why we chose @arc and get a teaser into whatโs next.
We start at 11:00AM. RSVP link in comments
Fast retweeted
The agent economy is in full swing ๐ฅโก๏ธ and @PiSquared has once again unveiled a fantastic new skill for agents, namely
๐๐ค๐ข๐ฅ๐ฅ ๐๐ฉ๐ญ๐ข๐ฆ๐ข๐ณ๐๐ซ โ a tool that:
โ ๐๐๐ ๐ก๐ ๐๐๐๐๐ก ๐ ๐๐๐๐๐
โ ๐ถโ๐๐๐๐ โ๐๐ค ๐๐๐๐๐๐๐๐๐ก ๐ฟ๐ฟ๐๐ ๐๐๐ก๐ข๐๐๐๐ฆ ๐ข๐ ๐ ๐กโ๐๐
โ ๐ด๐ข๐ก๐๐๐๐ก๐๐๐๐๐๐ฆ ๐๐๐๐๐๐ฃ๐๐ ๐กโ๐ ๐๐๐ ๐ก๐๐ข๐๐ก๐๐๐๐
๐๐ ๐๐ข๐๐ ๐ ๐ค๐๐๐. ๐๐ ๐ก๐๐๐๐ ๐๐๐ ๐๐๐๐๐.
This is a tool for automatically evaluating and improving agent guidance, ensuring that tools are reliably used by different LLMs. ๐
Now itโs not just about creating an agent, but about teaching it to work correctly.
Now agent skills arenโt just a case of โwriting a prompt and hoping for the bestโ,
but a measurable, optimisable system.
๐งช One skill โ tested across multiple models
๐ Metrics โ instead of subjective impressions
๐ Auto-iteration โ until it gets better
This is the next step:
Agents that donโt just exist,
but improve.
Welcome to optimised agent intelligence โก
#Fast #AI
๐ค Paid partnership
Fast retweeted
The builders I respect in this community have all arrived at the same realization:
Getting great AI output isn't prompting skill. It's staying organized aka "plan mode", working through a structure, not just dumping the problem into chat and hoping for the best.
normalpowers packages that structure. Free, open source, works on Claude (Web, Desktop, Cowork, Code), OpenClaw, Hermes. Adapted from superpowers (the Claude Code plugin) for everyone, not just developers.
Tried it on one of our GTM strategies last week. Came out with something actionable.
If you've spent hours on an AI session and still got garbage, this is why. And it's a 10-minute fix.
"You wasted all this time and tokens on nothing."
every AI user has thought this. the problem isn't prompting, it's that AI answers before it thinks. you get a polished answer to the wrong question. every time.
normalpowers fixes the front of the loop. free, open source.
think superpowers, but for generalists.
Fast retweeted

Grigore Rosu@RosuGrigoreApr 24
๐บ๐ธ
United StatesConnected via United States Android AppAccount-level information, not a live location or per-post device.
Observation after 1 week of running normalpowers on research and planning tasks:
The core failure mode in AI-assisted knowledge work isn't model quality. LLMs are optimized to answer, not to structure inquiry first. Give a model an unstructured problem, get unstructured output.
normalpowers adds the missing structure: 8 skills that make any LLM ask the right questions before generating output. Works on Claude (Web, Desktop, Cowork, Code), OpenClaw, Hermes.
The right frame isn't "better prompting." It's "better epistemics for working with AI."
Free. Open source. Worth 10 minutes of your time.
Fast retweeted

Grigore Rosu@RosuGrigoreApr 24
๐บ๐ธ
United StatesConnected via United States Android AppAccount-level information, not a live location or per-post device.
"Told you to first understand the problem well, and then answer."
Each time I blamed the model ... "next version will be smarter".
Truth is that good communication skills and asking the right questions are often more effective than intelligence alone.
Both for humans and for AI.
"You wasted all this time and tokens on nothing."
every AI user has thought this. the problem isn't prompting, it's that AI answers before it thinks. you get a polished answer to the wrong question. every time.
normalpowers fixes the front of the loop. free, open source.
think superpowers, but for generalists.
Fast retweeted

Grigore Rosu@RosuGrigoreApr 22
๐บ๐ธ
United StatesConnected via United States Android AppAccount-level information, not a live location or per-post device.
we used skill-optimizer to optimize the skill that we're now shipping with skill-optimizer
talk about inception...
also here's how the optimization process went๐
skill-optimizer 1.1.0 is here
โ๏ธAdded prompt surface as a type of tool the skill can be optimized for
๐คAdded Codex auth
๐งโ๐ซBundled SKILL/ guidance to allow AIs to easily use skill-optimizer
๐งนGeneral code cleanup and architecture improvements
github.com/fastxyz/skill-optโฆ
Fast retweeted
Good tool + Bad skill = Bad UX. Optimizing skills over repeated benchmarks is a much better path than prompt folklore indeed.
this tool is a stupidly simple loop:
insert tool + skill, let LLMs use it, benchmark performance, improve, repeat
and it's now open source.
try it, contribute to it, file issues, have fun:
Fast retweeted
This is pretty cool, so thought I'd share!
We took our payment skill from ~60% success on weaker models to 90%+ in around 3 hours for about $5 on OpenRouter
Thatโs a ridiculous trade
Fast retweeted

Grigore Rosu@RosuGrigoreApr 14
๐บ๐ธ
United StatesConnected via United States Android AppAccount-level information, not a live location or per-post device.
A good idea that works becomes a skill.
A skill becomes automatable agent work.
Skill Optimizer is an agent orchestrating other agents to craft the optimal skill for the job together.
Result? A skill that works with all agents.
Fast retweeted
How do you build a network for Agentic Payments? ๐ค
@PiSquared married two breakthroughs: a 2019 method to solve double-spend without a blockchain, and Grigore Rosu's EVM formal verification.
The result: A lightning-fast, mathematically proven execution layer. โก๏ธ
Fast retweeted
A Citizen's Guide to Driving Blockchain Policy by @JeetRaut, Head of Strategy of @PiSquared
2026 will be the most consequential year for U.S. blockchain policy in history. This talk gives attendees practical tools to shape outcomes, including how bills move, who influences them, and what actually motivates policymakers, from jobs and taxes to voter impact.
Full video below ๐๐งต
First day in Denver!
Our VP & Head of BD @lei_lei0904 joined @wrappedxyz's Multichain Day with top builders from @0G_labs , @PerleLabs , and @PiSquared to dive deep into the future of onchain AI infrastructure, from verifiable compute and data provenance to the payment rails powering the agentic economy.
๐กKey highlights:
โข AI x crypto remains strongly bullish despite uncertain market conditions
โข Kiteโs Human-Agent-Session architecture enables secure, low-fee micropayments purpose-built for agents
โข The shift from black-box models to decentralized, verifiable systems
โข Panel insights on storage, data labeling, and formal verification challenges
โข Agents are chain-agnostic and focused on getting tasks done efficiently
โข Spicy takes on training data scarcity and the future of social media
Letโs keep pushing the boundaries. ๐ช
Replying to @wrappedxyz @dgetsylver @theahaco @danielmarinq @NexusLabs @circle @0xCerberus @lifiprotocol @0xJibby @gauntlet_xyz @gttyson @StellarOrg @livingwithwill @realmotioninc @lxjhk @solayer_labs @SilentRhetoric @AlgoFoundation @kanishkkhurana @AcrossProtocol @chasemerlin @brale_xyz @Camohe90 @AleoHQ @chriskim_dev @Aptos @jaycoolh @hedera @thetinaverse @monad @aramxc92 @flipsidecrypto
โฐ Up next at 3:40PM MST โ Panel Discussion on AI Infra Onchain: Compute, Data, and the Coordination Layer!
Featuring:
@gathint from @0g_labs
@lei_lei0904 from @GoKiteAI
@rthurding from @perlelabs
@jeetraut from @PiSquared
@northernapez from @Multichain
๐ด Watch live: xitter.cf/i/broadcasts/1jMJgRRQYโฆ
Fast retweeted
We confuse the destination (Web3) with the vehicle (Blockchains). ๐
@PiSquared drops a spicy take: AI agents will integrate massively with Web3 and Crypto, butย notย necessarily with blockchains.
Blockchains are just a mechanism. The infra has to evolve to survive. ๐ค
Total ordering is killing every blockchain you're invested in.
No one talks about it because most don't even understand it.
Fascinating deep dive with @RosuGrigore on how @pisquared is solving Web3's biggest bottleneck:
๐นInteroperability
๐นScalability
๐นThe ordering problem that breaks at scale
From NASA engineering to 160k+ TPS on FastSet, the math doesn't lie.
Masterclass on what actually comes next: ๐
1:30 โ From NASA to Web3
3:39 โ The Elephant in the Room
4:54 โ What is Pi Squared?
12:06 โ Introducing FastSet
17:42 โ Why Not All Apps Need Ordering
24:46 โ Interoperability by Construction
34:11 โ Language Agnostic Web3
36:24 โ Hitting 150,000+ TPS
42:15 โ Introducing Omniset
50:40 โ The Steam Engine Analogy
Fast retweeted
why we back @PiSquared?
1/ because it creates a new category, not another blockchain, in fact "not blockchainโ at all, but a fully parallelized execution network where transactions are mathematically verifiable and can run without artificial dependencies
2/ on top of that, Pi Squared is shipping a stack we have not seen combined before: fastset for capital efficiency, Omniset for intent-based, bridgeless liquidity across apps, and a payments and markets layer that finally gives builders unified building blocks and a great interface
3/ most networks still force the world into queues. @PiSquared is built different, around how reality actually works - independent actions in parallel. this is the first meaningful step beyond sharding toward how scalable systems should really operate in web3
Fast retweeted

Grigore Rosu@RosuGrigoreJan 3
๐บ๐ธ
United StatesConnected via United States Android AppAccount-level information, not a live location or per-post device.
The so called "blockchain trilemma" incorrectly identifies the three fundamental principles of Crypto/Web3 with blockchains:
1. Decentralization - this is what makes the system unstoppable. Unstoppability is the only reason we talk about decentralization. Censorship-freedom is a corollary. This goes beyond blockchains.
2. Security - this is what gives us true ownership of our assets. Private/public key cryptography, verifiability (formal or not), soundness / consistency of protocols, etc., are just mechanisms to realize it, which all go well beyond blockchains.
3. Scalability - this is what makes the entire thing practical. To be future-proof and scale to the entire planet, including both humans and AI agents, we need infinitely scalable infra. Period. Blockchains are a particular instance of strongly consistent systems, which are known to not scale. Tribalisms and strongly opinionated CT folks and biased investors cannot change facts.
If you want to change the world for the better, look at fundamentals and build your and our and our children's future on solid principles and values. This is what give life meaning, after all.
We cannot have freedom without financial freedom. Or ore generally, without truth freedom. Indeed, _truth_ is main value that we need to express and convey. The rest is details. And for that we need unstoppability, security, and infinite scalability.
This is what FastSet is all about.
Fast retweeted
Your liquidity is fragmented across 10 chains.
@RosuGrigore and @PiSquared built OmniSet to fix that.
One wallet. All your assets unified. Verified in milliseconds.
No bridges. No trust. Just math.
Fast retweeted
โIs it an EVM chain?โ
First, itโs not a chain.
Second, hardwiring a language is stupid.
@RosuGrigore on why FastSet by @PiSquared rejects EVM and lets developers use any language.
Full interview ๐ youtu.be/tzIAJmYzM_o

























