Re: AI Generated Applications

Sorry, seems to be I replied myself 😳

Please allow me to continue with this conversation. Your time and replies
are highly appreciated.

I was refining the conversation with two of the most used AI assistants.
Please take a look at the second prompt responses on both:

https://share.gemini.google/LC0W27sQoPAd

https://chatgpt.com/share/6a2c26c5-eab0-83e9-a20b-f4f5807e8e26

This proposes an homoiconic (code as data) RDF/XML stack where the model
representation layers (human, agent, runtime) are serialized as graphs, as
is depicted by the AIs prompted about that themselves.

They both seem to agree that graph layers with those three levels of
abstraction can lead to a semantic-bytecode layer execution level (triples
or quads ASTs encoding). All ontology-driven, deterministically processable
by an AI and, for all, humanly auditable and visualizable by existing tools..

An integration stack needing some capability just needs to semantically
query a deployed application used commonly aligned ontologies. Runtime
(execution) could be just a matter of stream-processing raw RDF / Graph
data (and code) and perform transformations given those capabilities.

Regards,
Sebastián.


On Mon, Jun 15, 2026, 11:22 AM Melvin Carvalho <melvincarvalho@gmail.com>
wrote:

>
>
> pá 12. 6. 2026 v 19:07 odesílatel Chaals Nevile <chaals@fastmail.fm>
> napsal:
>
>>
>>
>> On Friday, 12 June 2026 17:33:47 (+02:00), Sebastian Samaruga wrote:
>>
>> This is what AI thinks about AI:
>>
>> https://chatgpt.com/share/6a2c26c5-eab0-83e9-a20b-f4f5807e8e26
>>
>> https://share.gemini.google/5kqbNH6fps7o
>>
>> On Fri, Jun 12, 2026, 12:17 PM Melvin Carvalho <melvincarvalho@gmail.com>
>> wrote:
>>
>>>
>>>
>>> pá 12. 6. 2026 v 17:13 odesílatel Sebastian Samaruga <ssamarug@gmail.com>
>>> napsal:
>>>
>>>> Question:
>>>>
>>>> Which sense does it make to still asking LLMs coding agents to generate
>>>> code in languages created to be edited by humans in those languages.
>>>>
>>>
>> There is a lot more asking LLMs to do coding than there are times when it
>> makes sense to do that...
>>
>> And because LLMs are so inherently untrustworthy, ensuring intelligent
>> humans can (and *do* - it seems obvious, but the evidence is out there)
>> verify the code is crucial for anything that you actually rely on for
>> anything meaningful, it's likely a *very* bad idea to enable vibe coding in
>> a language not designed for humans.
>>
>> It might be that we look at shifts from languages that are easy to code
>> in, to languages that are easy to inspect and understand. Since the
>> LLM-world is happy for systems to do a lot of work, this means optimising
>> the cost of the most important actions, which is a sensible approach in
>> general.
>>
>
> Largely agree. The risk is not just that LLMs write bad code, but that
> agentic coding systems optimize for getting the change through. Code
> doesn't pass tests, they disable the linter. Merge requires review, they
> force-push or use the GitHub API. I've seen them rewrite build systems just
> to fast-track changes without triggering CI.
>
> That means the language and workflows need to be optimized for reviewing
> code that looks fine on the surface, but may contain subtle bugs that stick
> around for years. And those failures compound.
>
>
>>
>>
>>>> The source of the browser centered web was about pages source files
>>>> easily editable by humans in text editors.. We are still targeting our
>>>> generated "applications" to those languages and what they have evolved to
>>>> for our implementations.
>>>>
>>>> What about an intermediate "agent application language", auditable by
>>>> humans, representation generated by the agents model and then compiled in,
>>>> still browser centric, I think about WASM here.
>>>>
>>>> And what about if this agent "application language" could leverage
>>>> natively AI models inference capabilities providing prompted / generated
>>>> applications with those capabilities.
>>>>
>>>
>> LLMs need a corpus to be useful. Human-oriented languages have a corpus,
>> and how you create code can easily be subject to training. Why make the
>> misery of teaching an AI not to do really stupid things harder?
>>
>> IMHO. Cheers
>>
>> --
>> Charles "Chaals" Nevile
>> Using fastmail.fm because it's worth it
>>
>