brus

human in the loop - AI SW factory architect

[brus @brus_lattice]

You probably mean, mainstream solopreneurs. AI development is everywhere. My take is experience and scale will become limiting factors. Today, 3 million lines of dense code is a comfortable target for a single solopreneur that has domain knowledge. To me, anything that is lower than that and is well known as domain, has zero moat.

[brus @brus_lattice]

@GuoShenCN7 AI Bills, brother.

[brus @brus_lattice]

@yongfook How about build dev tools for normies to build their own stuff?

[brus @brus_lattice]

@thsottiaux an agent without agency is an assistant

[brus @brus_lattice]

Build management is a disaster in today's standards, dependency management is a disaster, encapsulation is a disaster, making context leaking first class citizen, type hiding is build dependent making regex based source code understanding hardest of all languages (minus assembly). Generally context acquisition is bad, reasoning scope is likely much larger then in modern languages not plagued with massive legacy. Verbosity is through the roof. Concurrency is not part of the language. Evolution is shit due to legacy burden. AI also makes mistakes, debugging C is shi..ultra shit.

[brus @brus_lattice]

C is shit as a language geared for AI development. Syntax is nonexpressive high token count disaster. Memory tracking in AI is legacy bandaid not something you want AI to track extra but something that is discretely tracked. AI context window is extremely expensive, relatively, you want higher level abstractions that are learned in weights, C has shit for general purpose standard library which means AI implemented abstractions polute context window. C is legacy shit and should stay legacy and die with AI first, wave of programming languages that I hope will follow after everyone gets over harness obsession.

[brus @brus_lattice]

@thsottiaux remove/replace current analytics with an in depth version that allows us to optimize our usage. We NEED deeper insights on what the model is doing around input, output, which semantic clusters of tokens trigger more usage, which input statements as a whole are hot areas. Please allow the model to self analyze runs of 10-20 minutes, suggest data based improvements on user input, regardless of its form. The global costs here cannot be optimized purely based on trial and error.

[brus @brus_lattice]

Economy of scale always matters and AI is not free. I bet on a resurgence of SaaS but with a different pricing model, one time purchase. If it costs me more in terms of tokens to build something, I'd rather pay one time and get a mature solution.

[brus @brus_lattice]

Absolutely ridiculous, same two tasks, another account, this time 2x Sol high agents. 11 hours later, 7% weekly limit left. I hope this is a bug and it gets fixed soon. @thsottiaux it would be great if you extend analytics with tokens per skill, add another view with tool calls, all stats split per agent.

[brus @brus_lattice]

Thanks Tibo. I switched to codex fully but start to regret this. I can now clearly state that Astra high, is ~4x more token hungry than Fable 5.1 xhigh. It took less then 10 hours, with 2 agents to run through the weekly limit/your last reset. Its even worse because 1 of the agents was in long tool waits and this other was doing heavy implementation, with subagents. With Fable orchestrators I usually could run around 4x agents for 10 active hours.

[brus @brus_lattice]

@zarazhangrui Motivation is net negative sum if group individuals are not intrinsically self motivated. You can push and pull but do this longer term and it drains you. Teams do smooth out motivation levels though, for good and for worse.

[brus @brus_lattice]

@petergyang This thread is about people that talk about AI and don't use AI to build software.

[brus @brus_lattice]

@thsottiaux Perfect but if it's codex, shouldn't this also work with subscriptions? Pretty please!

[brus @brus_lattice]

What a world we live in.. Fable kindly asks Astra (live) to take over some of its work due to dependencies to Astra's subagent's active tasks. Massive scale AI gen sw development is here!

[brus @brus_lattice]

@0xSero Which is the $10k setup you had in mind for first glm 5.3 tier ?

[brus @brus_lattice]

@Da7_Tech @jun_song There was a time when GitHub copilot was message based instead of token based..

[brus @brus_lattice]

@sama I keep my AI relationship businessy, I come, I pay, I leave.

[brus @brus_lattice]

I'm one of these people building an AI SW Factory. SW factories are enterprisey.
Imagine you have hundreds to thousands of people, tens of roles that are working on multi million lines of code, hundreds of components that are pulled in a single product, with strict timelines. In enterprise you get siloes, knowledge distribution, large CI pipelines for tens/hundreds of images.
The holy grail here is having an agent that is able to navigate through this complexity but also keep human in the loop for any relevant reviews/sign offs/guidance.

[brus @brus_lattice]

@GeoffreyHuntley Man if I hear one more time AI transformation, mastering AI, driving AI or any of this corporate shitalese I'm going to .. ohh wait .. 🤮