> The hackathon winners understood something that most developers do not: the hard part of building useful AI is not the code, it is knowing what the system should do in the first place.
This has always been true of all systems. Not that it isn’t an insight though since I don’t think enough people seem to get it. To build a system, with an LLM or without, you must know what the system needs to do. If you define it in C or in a markdown file it must still be defined. The advantage with LLMs is they bridge the gap between system definition and being able to simulate that system on a processor. The definition of the system is still required and it still must be precise. Even with “AGI” that’s still going to be true just as it’s true today with humans who do the translation between those who deeply understand a system and software.
Agree - domain experts lack the expertise on how things should be built. Developers lack the expertise of what should be built. In each case, one can get into the role of the other, per what you say: "humans who do the translation between those who deeply understand a system and software". LLMs will extrapolate for both (whether it's good or bad)
The only thing missing from the system is the AI GOV that defines the specification of work. Once that is commonplace developers become ephemeral as the code to support the hardened GOV. That is what CANONIC.org is.
This appears to be an advertisement for a (somewhat inscrutable) AI product they're selling called CANONIC, that also has a cryptocoin bolted on to it somehow.
This is partially true. Thanks for the comment. I’m the developer of CANONIC. It’s an AI GOV framework. But crypto it is not. The opposite.
CANONIC is a learning language to fully govern AI. Ask yourself why you are still programming computation with LLMs when they repeatedly outperform humans on such coding tasks. What’s missing is AI GOV. CANONIC is a contract with your AI. COIN becomes an artifact of good AI governence. Hardly an opaque transaction. :)
It’s a tortured metaphor for benefit or value or bounty. So, to work with the cardiologist, she publicly issues a bounty (numerator) in COIN and I privately estimate the effort denominator. Triage the table of bounties by this fraction.
I guess it could also be used to communicate that some problems are too difficult for modest resources, if the reward exceeds 255.
Anyway, the only thing worth spending it on is resource upgrades, right?
Oh boy, if it's powered by AI /AND/ the blockchain, you know it's gonna be legit ;)
Oh boy is right. If BITCOIN is SPECulation to distribute cryptographic transactions on a blockchain pre AI, CANONIC is a SPECification to distribute WORK contracts with your AI on a blockchain. So BITCOIN=SPECulation. While CANONIC COIN=SPECification… of WORK!
It will be interesting to see what happens when more individuals with legitimate expertise in two or more non-software fields build tools that rely upon that intersection of deep, specific knowledge. I wonder how long it will take for AI to really begin encouraging a pursuit of generalist-style learning, as opposed to the specialist inclination that's ingratiated into nearly all formal institutions.
That's maybe not what the author was getting at, but that's what I came away with anyway.
That’s exactly what I’m getting at. LLMs have abstracted computation itself. In fact LLMs are a step back from computation. They are in essence a state machine. A prompt goes in and a response comes out. CANONIC is a language to extend that across a tree of governance towards a complex gated state machine that must be compliant. https://hadleylab.org/blogs/2025-12-29-the-compiler-insight/
Author here. The blog argues that the real story from Anthropic's hackathon isn't that domain experts can build AI (they can) but that hackathon demos and production systems require fundamentally different things. A permit app that works on demo day and a permit system that survives when California revises the code, when the builder leaves, when a municipality asks for an audit trail — those are different problems. We're building a governance framework (CANONIC — CANONIC.org) where every AI capability is declared in a versioned contract. Curious what HN thinks about the gap between "domain expert can build" and "institution can trust what they built."
That's interesting, it reminds me of something we've realized internally at my company, AI coding is best used with strict adherence to requirements and tests (potentially generated by AI), reviewed by a human developer
Indeed. CLAUDE.md is a linear memory for agents. Hardly the structure for multi orchestration requires for agenetic programming today. CANONIC is a learning language to customize agents across your governance tree. Ever internal node is an opportunity to govern intelligence which completely redefines what agents can do, how they communicate and the overall sophistication of fleet orchestration.
Testing
Maybe it’s just personal deja vu, but in current discussions of vibecoding v software engineering I keep seeing parallels from the debates 20 years ago when blogs began proliferating: the bloggers v “real journalism” arguments.
The AI editing of the article makes it a painful read. A shame because the point they were making is a good one, regarding AI coding apps empowering domain experts.
AI editing? It is completely governed by AI. I thought that was obvious. What’s the problem?
The wording in the posted article, which is elevator-pitch. The audience has money, ambitious, mercenary attitude, and proximity to big problem that must be either solved or resold.
Rather, humans involved have looked at this problem and in a few cases succeeded. And in many cases, shelved it and returned to process cardiology, Ugandan infrastructure, etc.
Simplifying here… sounds like essentially the split between (great) product managers and engineers
We need both! Hurrah
Indeed it’s both. Once humans start governing their AI. Once the blockchain community can check the GOV contract for validity. Founder of One becomes a reality. Contracts are institutional memory!
to put it another way, writing resilient software is a domain of its own and software engineers are it's domain experts.
Let the LLM write the software. That’s ephemeral and evolves with time. Humans should govern the entire system to resilience. That is fixed with time. Thou shalt not kill has staying power. Weapons, poison, and other methods of death evolve. More governence deals with them over time.
your LLM prompt is, by definition, underspecified. the code is what describes the actual behaviour of the system, and there are known and understood ways to make that behaviour more robust, correct and resilient, that are independent of the domain the code is modelling, but consistent across different code bases. that's why I say writing code is its own domain.
as an analogy, an art museum couldn't paint their own paintings to hang up (or at least they would not be very good) but neither would monet or picasso have done a particularly good job at designing a space to let millions of people a year view their pictures. both skills are necessary to the overall product.
> The hackathon winners understood something that most developers do not: the hard part of building useful AI is not the code, it is knowing what the system should do in the first place.
This has always been true of all systems. Not that it isn’t an insight though since I don’t think enough people seem to get it. To build a system, with an LLM or without, you must know what the system needs to do. If you define it in C or in a markdown file it must still be defined. The advantage with LLMs is they bridge the gap between system definition and being able to simulate that system on a processor. The definition of the system is still required and it still must be precise. Even with “AGI” that’s still going to be true just as it’s true today with humans who do the translation between those who deeply understand a system and software.
Agree - domain experts lack the expertise on how things should be built. Developers lack the expertise of what should be built. In each case, one can get into the role of the other, per what you say: "humans who do the translation between those who deeply understand a system and software". LLMs will extrapolate for both (whether it's good or bad)
The only thing missing from the system is the AI GOV that defines the specification of work. Once that is commonplace developers become ephemeral as the code to support the hardened GOV. That is what CANONIC.org is.
This appears to be an advertisement for a (somewhat inscrutable) AI product they're selling called CANONIC, that also has a cryptocoin bolted on to it somehow.
This is partially true. Thanks for the comment. I’m the developer of CANONIC. It’s an AI GOV framework. But crypto it is not. The opposite.
CANONIC is a learning language to fully govern AI. Ask yourself why you are still programming computation with LLMs when they repeatedly outperform humans on such coding tasks. What’s missing is AI GOV. CANONIC is a contract with your AI. COIN becomes an artifact of good AI governence. Hardly an opaque transaction. :)
https://hadleylab.org/blogs/2026-02-23-coin-for-humans/
It’s a tortured metaphor for benefit or value or bounty. So, to work with the cardiologist, she publicly issues a bounty (numerator) in COIN and I privately estimate the effort denominator. Triage the table of bounties by this fraction.
I guess it could also be used to communicate that some problems are too difficult for modest resources, if the reward exceeds 255.
Anyway, the only thing worth spending it on is resource upgrades, right?
Oh boy, if it's powered by AI /AND/ the blockchain, you know it's gonna be legit ;)
Oh boy is right. If BITCOIN is SPECulation to distribute cryptographic transactions on a blockchain pre AI, CANONIC is a SPECification to distribute WORK contracts with your AI on a blockchain. So BITCOIN=SPECulation. While CANONIC COIN=SPECification… of WORK!
https://hadleylab.org/blogs/2026-02-23-coin-for-humans/
It will be interesting to see what happens when more individuals with legitimate expertise in two or more non-software fields build tools that rely upon that intersection of deep, specific knowledge. I wonder how long it will take for AI to really begin encouraging a pursuit of generalist-style learning, as opposed to the specialist inclination that's ingratiated into nearly all formal institutions.
That's maybe not what the author was getting at, but that's what I came away with anyway.
That’s exactly what I’m getting at. LLMs have abstracted computation itself. In fact LLMs are a step back from computation. They are in essence a state machine. A prompt goes in and a response comes out. CANONIC is a language to extend that across a tree of governance towards a complex gated state machine that must be compliant. https://hadleylab.org/blogs/2025-12-29-the-compiler-insight/
Author here. The blog argues that the real story from Anthropic's hackathon isn't that domain experts can build AI (they can) but that hackathon demos and production systems require fundamentally different things. A permit app that works on demo day and a permit system that survives when California revises the code, when the builder leaves, when a municipality asks for an audit trail — those are different problems. We're building a governance framework (CANONIC — CANONIC.org) where every AI capability is declared in a versioned contract. Curious what HN thinks about the gap between "domain expert can build" and "institution can trust what they built."
That's interesting, it reminds me of something we've realized internally at my company, AI coding is best used with strict adherence to requirements and tests (potentially generated by AI), reviewed by a human developer
Indeed. CLAUDE.md is a linear memory for agents. Hardly the structure for multi orchestration requires for agenetic programming today. CANONIC is a learning language to customize agents across your governance tree. Ever internal node is an opportunity to govern intelligence which completely redefines what agents can do, how they communicate and the overall sophistication of fleet orchestration.
Testing
Maybe it’s just personal deja vu, but in current discussions of vibecoding v software engineering I keep seeing parallels from the debates 20 years ago when blogs began proliferating: the bloggers v “real journalism” arguments.
The AI editing of the article makes it a painful read. A shame because the point they were making is a good one, regarding AI coding apps empowering domain experts.
AI editing? It is completely governed by AI. I thought that was obvious. What’s the problem?
The wording in the posted article, which is elevator-pitch. The audience has money, ambitious, mercenary attitude, and proximity to big problem that must be either solved or resold.
Rather, humans involved have looked at this problem and in a few cases succeeded. And in many cases, shelved it and returned to process cardiology, Ugandan infrastructure, etc.
Simplifying here… sounds like essentially the split between (great) product managers and engineers
We need both! Hurrah
Indeed it’s both. Once humans start governing their AI. Once the blockchain community can check the GOV contract for validity. Founder of One becomes a reality. Contracts are institutional memory!
to put it another way, writing resilient software is a domain of its own and software engineers are it's domain experts.
Let the LLM write the software. That’s ephemeral and evolves with time. Humans should govern the entire system to resilience. That is fixed with time. Thou shalt not kill has staying power. Weapons, poison, and other methods of death evolve. More governence deals with them over time.
your LLM prompt is, by definition, underspecified. the code is what describes the actual behaviour of the system, and there are known and understood ways to make that behaviour more robust, correct and resilient, that are independent of the domain the code is modelling, but consistent across different code bases. that's why I say writing code is its own domain.
as an analogy, an art museum couldn't paint their own paintings to hang up (or at least they would not be very good) but neither would monet or picasso have done a particularly good job at designing a space to let millions of people a year view their pictures. both skills are necessary to the overall product.