Hi Romain

On Sun, Apr 26, 2026 at 07:40:52PM -0500, Romain Beauxis via ffmpeg-devel wrote:
> Hi Michael,
> 
> Le ven. 24 avr. 2026 à 04:02, Michael Niedermayer via ffmpeg-devel
> <[email protected]> a écrit :
> >
> > Hi
> >
> > Some question about fairy design
> >
> > Fairy is written in python,
> > the main script does
> > * fetches the forges data (separate thread, cache and all)
> > * spawns LLM workers
> > * old style UI with a simple yes/next/skip/redo/quit for each LLM response
> >
> > The LLM workers each call a seperate wraper (aka new process)
> > which then contains the code specific for openai or in the future antropic 
> > and others
> >
> > The problem iam running into with this is, as i want to use multiple agents
> > (ATM teh openai code uses 2, we use a gpt-5.4 nano to setup the container,
> >  well basically untar our git repositories)
> > and then the main gpt-5.4 reviewing the code
> >
> > Now the easy thing is to just add a claude / antropic wraper but this
> > actually feels like bad design
> >
> > what i was thinking is to rather get rid of the process and do these 
> > "wrapers"
> > fully in python.
> > The reason why, is simply, i want to be able to querry multiple LLMs
> > and combine their reviews and these wraper feel a bit wrong for that
> >
> > Its easy to querry openai 3 times and have openai combine 3 results
> > IN the openai wraper. But that will keep bloating up the wraper over time
> > and make them increasingly non interchangeable and also have more logic
> > where it doesnt belong
> >
> > And as my idea was to try to have both GPT and Claude
> > do the review and combine their findings that kind of makes the
> > wrappers at process level feel a bit heavy
> 
> How are you planning on merging the reviews? I'm curious as to how
> this can be done while keeping the review content at a reasonable
> level of verbosity and accuracy..

I have no idea, i intend to figure that out when/if i implement that :)

But the merging LLM would be supposed to "understand" the reviews it
merges, not copy and paste.

thx

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Complexity theory is the science of finding the exact solution to an
approximation. Benchmarking OTOH is finding an approximation of the exact

Attachment: signature.asc
Description: PGP signature

_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to