Javier Capello

Tech Lead @sony @cimediacloud | blazing fast transfers

[Javier Capello @capello_javier]

Agents are excellent at most of coding tasks but they still have a lot of issues with systems design, even the strongest models tend to design some complicated solutions when they involve multiple systems and repos and you end up with a thing that has a million failure modes and is quite buggy!

Pro tip: figure out the general design for the solution first and give the bullet points to your architect agent to generate the simplest design following the principles of systems thinking - never let the agent fully generate the design flop for you, for you will suffer!

https://t.co/AO7ujRyE3I

[Javier Capello @capello_javier]

Pro tip: ask your agent to distill this for you and add an entry to your knowledge base.

https://t.co/cfpwWxjp0y

These are old systems engineering principles that are still aging like fine wine in the current AI era.

The key thing from my experience is that the best solutions are beautifully simple and crafted in a way that makes almost all possible failure modes (bugs) impossible to happen (the problem is dissolved).

[Javier Capello @capello_javier]

@trekedge Please have them build an oracle for systems design! The models seem to be really bad at designing complex architecture - even Astra - and I’m being forced to write markdown by hand!

[Javier Capello @capello_javier]

Architecture - it is terrible at architecting solutions that involve multiple systems across different repos lol. Even Astra just comes out with these extremely fancy-looking designs that are just unnecessarily complicated; add many failure modes, and you end up discovering a large stream of bugs that force you to simplify the solution or continue fixing bugs forever!

I've found a really simple approach that seems to work: just give the architect agent a simple set of bullet points of what you want on each component and design the solution around that - never let the agents generate complicated design flop for you will suffer as I did!

[Javier Capello @capello_javier]

Pro tip: ask your agent to distill this for you and add an entry to your knowledge base.

https://t.co/cfpwWxjp0y

These are old systems engineering principles that are still aging like fine wine in the current AI era.

The key thing from my experience is that the best solutions are beautifully simple and crafted in a way that makes almost all possible failure modes (bugs) impossible to happen (the problem is dissolved).

[Javier Capello @capello_javier]

Architecture - it is terrible at architecting solutions that involve multiple systems across different repos lol. Even Astra just comes out with these extremely fancy-looking designs that are just unnecessarily complicated; add many failure modes, and you end up discovering a large stream of bugs that force you to simplify the solution or continue fixing bugs forever!

I've found a really simple approach that seems to work: just give the architect agent a simple set of bullet points of what you want on each component and design the solution around that - never let the agents generate complicated design flop for you will suffer as I did!

[Javier Capello @capello_javier]

Pro tip: ask your agent to distill this for you and add an entry to your knowledge base.

https://t.co/cfpwWxjp0y

These are old systems engineering principles that are still aging like fine wine in the current AI era.

The key thing from my experience is that the best solutions are beautifully simple and crafted in a way that makes almost all possible failure modes (bugs) impossible to happen (the problem is dissolved).

[Javier Capello @capello_javier]

Architecture - it is terrible at architecting solutions that involve multiple systems across different repos lol. Even Astra just comes out with these extremely fancy-looking designs that are just unnecessarily complicated; add many failure modes, and you end up discovering a large stream of bugs that force you to simplify the solution or continue fixing bugs forever!

I've found a really simple approach that seems to work: just give the architect agent a simple set of bullet points of what you want on each component and design the solution around that - never let the agents generate complicated design flop for you will suffer as I did!

[Javier Capello @capello_javier]

@philhchen I've spent the last couple of months attacking RH for fun and to figure out the principles of agent systems engineering but the agents keep writing manuscripts and I'm not sure what to do about them really - I fear I might offend real mathematicians if I publish them😅

[Javier Capello @capello_javier]

@trekedge Are you attacking RH or the dishwasher problem?

I could use some help with RH! I've been attacking it for months, the pipeline just keeps producing manuscripts, and I have no idea what to do with the manuscripts😅

Is this thing even solvable!?

[Javier Capello @capello_javier]

@QiaochuYuan Pi is 3.1415926535897932384626433832795...

But nobody complains

[Javier Capello @capello_javier]

@brockjelmore What do you mean? It’s only “1 Planck” - a perfectly fine wooden number

[Javier Capello @capello_javier]

@apartovi @neo Finally! It feels like the ChatGPT’s been legacy for a while now. I’m loving the changes!

Maybe we’ve grown spoiled with all the model releases and we felt that the straightforward UI was getting old faster than it appears!

[Javier Capello @capello_javier]

@tomosman @ivanburazin @auchenberg Yes but you can't just magically turn money into semiconductors. Chipmakers can only produce so much, and the market needs a lot

[Javier Capello @capello_javier]

@marcusabramovi1 Now imagine spending your mathematician life thinking about some truly nonsensically difficult problems that weren't even worth trying, and suddenly having these extremely powerful AIs that could give you hope of solving them. I would be very excited!

[Javier Capello @capello_javier]

Possibly some automated misclassification. Some new fancy AIs like Jev and OpenAI's decision API are bringing in some very tiny models that are decent at making decisions while being massively cheaper. This is an extremely attractive approach for large companies looking to save costs.

[Javier Capello @capello_javier]

Amazon packages now when you order 11 items

[Square Packing Fan @ManassehA06]

The optimality of the packing for 11 squares has been formalized in lean thanks to Astra and Claude! Huge thanks to @ojoshe, @kleddamag, @wand_125, @guzhou0806, and @ctjlewis for aiding in the process.

Image credit: https://t.co/2oUaPKe2Hz
1/n

[Javier Capello @capello_javier]

@grok @achugwotwins @grok rate the previous grok bait from 1 to 10

[Javier Capello @capello_javier]

Yes, but a lot of that complexity is accidental. Some of the extra code really is just covering edge cases the simpler version ignored. Much of it comes from a design that forces those cases to be handled in many places instead of one though - duplicated checks, logical branches, states the types could have ruled out. That usually means the model of the problem is wrong and should go back to the drawing board. AI is especially good at committing to a bad design: it extends whatever structure you already have and rarely questions it unless you force the redesign.

Some of the old principles from Systems Engineering are still aging like fine wine!

https://t.co/cfpwWxjp0y

[Javier Capello @capello_javier]

@MatthewBerman Those remind me to the early AI with people making many startups that were essentially just ChatGPT wrappers with a SaaS layer on top.

Those startups look a lot like wrappers around existing consumer hardware and inference software.