"Most expertise gets built through grunt work that AI now lets you skip. Unless you deliberately use AI as a training partner rather than a replacement for thinking, the tool hollows out the judgment it needs you to have."
Do the work first — whenever I see someone jumping into automated workflows without deeply understanding the work, what really moves the needle and all , I do push back hard. Also, totally agree on the friction bit, too. Gave me joy to read.. Great piece! Thank you!!
Devansh, your breakdown of "bounded variation" and the structural inevitability of hallucinations is one of the most accurate diagnoses of the current AI landscape. You completely nailed the concept of the "Oracle." The market premium has absolutely shifted from generating output to verifying truth.
But here is the terrifying reality for enterprises deploying these systems: having an Oracle in your organization is useless if the probabilistic engine can execute an action faster than the Oracle can check it.
If an agentic system is given the power to execute a financial trade, alter procurement, or run an enterprise workflow, and it hits an edge case outside its "bounded variation," it doesn't pause to ask for human judgment. It confidently executes a hallucinated error at machine speed.
You cannot engineer the hallucination out of the software wrapper. The Stanford and JAMA studies you cited prove that.
This is exactly why we engineered Veritas Core. We realized that since you cannot mathematically guarantee a probabilistic model will always be correct, you must physically guarantee it cannot act without the Oracle's consent.
We moved governance to the physics layer. Our bare-metal PCIe circuit breaker physically severs the execution bus at exactly T=0 unless a mathematically bound, offline-verifiable human-authority receipt is presented.
We let the AI compress, forget, and guess in the software layer. But the moment it attempts to impact the real world, it strikes a physical anvil that demands human accountability.
Brilliant article. You have perfectly defined why human judgment is the ultimate scarce resource—Veritas Core is just the hardware required to physically enforce it.
Amazing review and thanks so much @Devansh! The extension to the oracle framework with more specifics is excellent as well. I completely agree.
So glad you liked it.
Love the perspective, and thanks for helping me discover James Wang as well :)
@James Wang is a baller
Brilliant ✨️
"Most expertise gets built through grunt work that AI now lets you skip. Unless you deliberately use AI as a training partner rather than a replacement for thinking, the tool hollows out the judgment it needs you to have."
Glad you liked it
Do the work first — whenever I see someone jumping into automated workflows without deeply understanding the work, what really moves the needle and all , I do push back hard. Also, totally agree on the friction bit, too. Gave me joy to read.. Great piece! Thank you!!
It’s coming…https://skepticalsociopath.substack.com/p/the-end-of-the-world-as-we-know-it?utm_source=share&utm_medium=android&r=wxwe9
Yeah it’s recursive RSI now, at least the process
Devansh, your breakdown of "bounded variation" and the structural inevitability of hallucinations is one of the most accurate diagnoses of the current AI landscape. You completely nailed the concept of the "Oracle." The market premium has absolutely shifted from generating output to verifying truth.
But here is the terrifying reality for enterprises deploying these systems: having an Oracle in your organization is useless if the probabilistic engine can execute an action faster than the Oracle can check it.
If an agentic system is given the power to execute a financial trade, alter procurement, or run an enterprise workflow, and it hits an edge case outside its "bounded variation," it doesn't pause to ask for human judgment. It confidently executes a hallucinated error at machine speed.
You cannot engineer the hallucination out of the software wrapper. The Stanford and JAMA studies you cited prove that.
This is exactly why we engineered Veritas Core. We realized that since you cannot mathematically guarantee a probabilistic model will always be correct, you must physically guarantee it cannot act without the Oracle's consent.
We moved governance to the physics layer. Our bare-metal PCIe circuit breaker physically severs the execution bus at exactly T=0 unless a mathematically bound, offline-verifiable human-authority receipt is presented.
We let the AI compress, forget, and guess in the software layer. But the moment it attempts to impact the real world, it strikes a physical anvil that demands human accountability.
Brilliant article. You have perfectly defined why human judgment is the ultimate scarce resource—Veritas Core is just the hardware required to physically enforce it.
"How to Thrive in the Age of AI" - A Talk on Devansh's Framework
Homage to James Whittaker, MSR 2013
https://substack.com/@dyb5786/note/c-267388413