Training Data
“Training Data” – a cartoon showing how giving agents the ability to make payment decisions will require training them to act more responsibly than humans.
The internet was never designed for money. Marc Andreessen called it the “original sin” — the failure to embed seamless, native payments into the web. Instead, we got ads and tracking as the dominant business models. Today, we’re still living with that mistake. ChatGPT has 700 million users. Almost none have spent a cent through it. AI can write code and book flights, but when it hits a paywall, it stops. It can’t transact.
That’s finally starting to change. Coinbase and Cloudflare have revived HTTP 402 with a new protocol called x402. It lets any website ask for payment, and lets machines respond. Other efforts like ATXP and AP2 are working on letting agents authorize and complete payments between each other.
Stablecoins are becoming the default money for this machine-to-machine economy. They’re programmable, auditable, and better suited to fast, small payments than credit cards. But there’s a deeper risk. We’re training AI agents to act like humans. That sounds responsible. But it might just teach them to spam, scam, and manipulate at scale. Fraud becomes frictionless.
We need a better model. Blockchain can help, with transparent audit trails, persistent agent identity, and verifiable intent. Players like Privy (wallet access), Base (settlement layer), and Visa (tokenized credentials) are all working on parts of this stack. As Vicky Bindra, CEO of Trulioo, said, “Agentic commerce has significant potential, but it can only scale with trust built in from the start.” Giving agents the power to pay is a big step. Giving them values is the next one.
198
34
7
“The moment AI starts making payments, it changes the game completely. It’s no longer just software, it becomes an economic participant. So.. the real challenge now isn’t about how it transacts, but about giving those actions meaning and intent.”
“Digital ninjas transacting autonomously with a blockchain backed moral code might just be the plot twist in our high stakes digital saga. Marrying rapid AI operations with built in ethical checkpoints could trim risk and open up a slick new revenue channel.”
“The problem isn’t training agents to act like humans – it’s that we’re training them on human payment behavior from ads and tracking instead of clean transactional models”
“Love the cartoon Ian Foley. Beyond audit trails, do we need enforceable agent identity + spend caps + staking/slashing for abuse, not just post-hoc forensics?”
“yes..the path is clear. Self Custodial Wallets, AI that needs to be defined as a proxy for you, but you personally need to always sign transactions. personal identity and control is the most important thing for AI to function correctly and within bounds defined by humans. AI is the front end, Blockchain is the backend, but it revolves around identity, personal agency and data ownership.”
“Do the fundamentals explain that you need actual money, 26k and a few k margin, to do any active (high-frequency) day trading? The most you may get out of AI is a swing trader in the week-to-week range. If you leave AI in the simulation sandbox, it will likely learn to fail at anything outside a simulation account. I wrote my own strategy scripts and know there is a warehouse of them that these AI can regurgitate that may actually work if the simulation lag matches market lagging.”
“so the robots will get busted too, soon or later )”
“Maybe if you totally grew up around robots then it might like…”
“Algorithm U
However there’s a 100,000 H1B Fee”
“The very premise of this panel is laughable. Economics is the softest of “sciences” that I’m forced to air quote in real life as well”
“I was a part of ETH Global, and there was a team that was able to accelarate the protocol, and it’s on github.”
“Neo classical economics is complete BS they need to make “simplifying assumptions” to make the maths work. Plus the model that’s taught about banking is also just wrong”
“Past positive results of learning economic fundamentals do not guarantee future success. World has changed and is chaning at the pace that fundamentals do not work any more. Otherwise please explain PE of MAG how many now? 7? 2?”
“It’s precisely at the intersection of AI payment systems and blockchain that the acute need for hybrid leaders emerges, a CRO who can build monetization for this new “machine-to-machine” economy, and a CISO who will embed security and trust into its foundation, rather than scaling its risks.”
“This is brilliant and true!”
“Interesting so that becomes another use case for Stablecoins Ian Foley but as you said it creates a new set of payment related risks”
“AI can’t code at all! I think this consensus is from people who don’t know how to engineer at a complex level or people who cant engineer at all! regression <- lm(var1 ~ var2, data = data) …..that is a regression in R. Nearly a line of code. You can YouTube it and learn it in a day. Then you’re a coder!
But the reality to the mathematics under the hood and the complexity of dealing with this at a level thats within the bounds of relevancy and accountability!!!
I am disciplined guy…I study it and math a good 4 to 6 hours a day! And after a solid 9 years of that much effort ….I know enough to know …I have so much to understand about how a computer is dealing with high level computation ..All AI is, is making that processes dynamic and looped. AI is garbage at engineering and coding. Anyone who tells you otherwise, is only interested in how many white cars did ford produce on a Tuesday! Yes chat chat GBT will help ya! But real value! Relevance …you can forget it!!!!!
And trading!!! LOL this is joke! AI to trade is as good as the world is flat!”
“Useful info. Payments that work for agents start with clear identity, audit trails, and limits by design.”
“Payments are the missing native primitive for agents. The trust layer needs three things: persistent identity, programmable risk controls and auditable rails. Open question: how will x402/ATXP handle refunds or chargebacks and AML who’s accountable when an agent misbehaves?”
“Actually, we took this a massive step forward. The robot scientist program (advanced abduction) can provide any LLM a bill that calculates the scientific gains from the dialogue.
We believe that after all the day trading is done, the globe will trade on scientific discovery from dialogue between advanced computers to LLMs.
Think of it as Kolmogorov therapy payments to an ailing and diseased
$8T LLM market.
Since we are busy making biotech discoveries that have causality, we can wait until LLMs are hurting more.”
Sources:
Tony Bradley (Aug 26,, 2025) – Why Simplicity Is The Path To Trust For Agentic AI – Forbes
Tomer Elias (Jul 30, 2025) – AgenticTrust: The Trust Layer for Agentic Commerce in the AI Era – Human Security
Sage Lazzaro (Sep 03, 2025) – Agentic AI is the technology’s new frontier, and CAIOs are toiling to get agents implemented correctly – Fortune
