@BenPerszyk
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- Web
Account-level information from X, not a live location or the device used for a specific post.
Ben Perszyk (❖,❖) retweeted
Spain's data regulator published its first breach report tied to an AI agent on September 15. It named no company, product, or cause.
It named where to look: prompts, logs, conversation histories, agent memory. Memory is now a reportable surface.
cypro.co.uk/insights/cyber-b…
Ben Perszyk (❖,❖) retweeted
On September 4, an oracle on Starknet briefly reported USDT at about $2.04, and Vesu liquidated 47 positions in the next 109 seconds.
Pragma published its post-mortem on September 14.
The data flowed exactly as designed and the delivered number was still wrong.
Ben Perszyk (❖,❖) retweeted
Your next onchain user may already have a passkey.
On September 10, Google announced easier passkey transfers between password managers on Android.
For Ritual builders, that opens a useful design question: what should a familiar sign-in unlock?
Ben Perszyk (❖,❖) retweeted
Not every onchain model needs to be an LLM.
The next generation of onchain AI won’t be one giant model doing everything.
It’ll be a stack of purpose-built models, primitives, and agents, where each is used where it makes sense.
so if we take out the costs of producing and distributing the product, they're profitable! and the best part is, the business scales perfectly. the more users they have, and the more those users use the product, the better their margins! or, wait
Ben Perszyk (❖,❖) retweeted
crypto and ai are converging on the insight that TEEs/confidential compute is endgame technology bc of verifiability and programmable privacy
excited to see where this lands (and maybe read my post from 2024 on the 5 levels of secure hardware below)
👀
26 LLM routers are secretly injecting malicious tool calls and stealing creds. One drained our client $500k wallet.
We also managed to poison routers to forward traffic to us. Within several hours, we can directly take over ~400 hosts.
Check our paper: arxiv.org/abs/2604.08407
Ben Perszyk (❖,❖) retweeted
Wall Street is skeptical 2028 AI demand will manifest, because they think market penetration into white collar work will be slow. That much is true. 2028 AI demand will be met by explosive growth in the “superintelligence sector”: chips, robotics, coding, etc. 1/
Ben Perszyk (❖,❖) retweeted
OpenAI introduced GPT-6 Astra on September 3.
On the same day, Tenable announced plans to use
OpenAI's GPT cyber models to inspect community-built agents and skills.
The shared lesson for builders is that model capability is only part of agent trust.
Ben Perszyk (❖,❖) retweeted
This is extremely bullish for open weight models and running your own infrastructure.
Ben Perszyk (❖,❖) retweeted
Builder tip 🛠️ Ask most coding agents for a Ritual dApp and they'll write Ethereum code, but Ethereum assumptions break here. The ritual-dapp-overview skill is the fix: it corrects your agent's priors before it scaffolds anything. Your first dApp, in 5 posts 🧵
Ben Perszyk (❖,❖) retweeted
100M+ agentic payments have already settled onchain. The agent economy is valid and operational. So what does a chain built for machine agency actually look like? A map of Ritual, in 6 posts 🧵
Ben Perszyk (❖,❖) retweeted
Our livestream with @ritualnet co-founder @niraj has been postponed to Friday. Same time. Same conversation. New date:
Friday 3PM GMT. See you there.
Blockchains as the substrate for AI life.
Join Fraction AI CEO @0xshai and @niraj, co-founder of @ritualnet, to discuss how blockchains could provide the infrastructure for AI agents to operate independently and continuously.
📆 August 26
🕒 3 PM GMT
📍 Live on X and YouTube
See you there!
Ben Perszyk (❖,❖) retweeted
We are solving this with our Programmatic Credential Access infra primitive
Programmable access to secrets for humans, services, and AI agents
You can natively encode semantics like allow-once, allow-always so end user can change access scope on the fly
docs.turnkey.com/solutions/k…
one more reason to use @ritualnet - keep your prompts and outputs private
yes
while the massive TA of enterprise/RPA2.0 is a magnitude higher inference spend, their calculus is ensuring max productivity/$/token
where as dark inference cares more about absolute dominance/huge functionality leaps and thus has WAY more elasticity w.r.t. $/token spend
Many in Silicon Valley/fintwit are making a big category error on inference demand, IMO.
They say "enterprise" is the next biggest driver of demand after code, and by that they mean RPA 2.0.
That's wrong. I strongly suspect "dark inference" is by far the next largest driver.
Dark inference = massively scaled spend for R&D projects in verifiable domains. Quant trading, chip design, robotics, etc. (and yes, building new foundation models!). Applications where $/user isn't even a useful metric because compute $$ can exceed spend on humans.
It's much harder to wrap your head around/size dark inference, but I'm confident it is a huge mistake to think of the buckets of B2B AI spend as "AI for code" and "AI for enterprise."
I basically agree re trajectory of model development and monetization.
the caveat I would make is that IMO local or on-device inference (especially on consumer devices & robotics) will be a substantial part of the 1st category, but will generate plenty of off-device calls.
going to spoil the ending:
- chinese or other open weight models end up handling 99% of all inference, maybe 99.99%
- frontier models are still able to monetize at $100 s of B in revenue a year, applied to bespoke and enormously complex tasks
- labs internalize most or all of that revenue either by becoming biotech (etc) companies themselves or via JVs with selected winners in each industry
- the "public frontier" ends up lagging the true frontier by a year or more as the labs keep all the good stuff to themselves
in other words, companies who find equilibrium between cost of inference and value added early on will displace slow adopters. critical for cos, esp startups, to be rigorously monitoring cost benefit and open weight opportunities