Jonathan Yu

building agents @8090_Factory

[Jonathan Yu @yucjonathan]

RLing a model in a sandbox is basically the Truman Show

[Jonathan Yu @yucjonathan]

@colemurray What are the top providers you’ve found that meet these requirements?

[Jonathan Yu @yucjonathan]

Codex adoption is directly correlated with people learning what "fail open" means

[Jonathan Yu @yucjonathan]

@pidotdev Changing the system prompt when compacting is an interesting idea, especially because the point of compaction is when completion is the most expensive.

How did you decide the performance gains of this were worth breaking cache?

[Jonathan Yu @yucjonathan]

Prescribing unnecessary workflows increases the possibility of giving conflicting instructions (which cause Sonnet 4.6-esque thought loops), and deprives the agent of the ability to innovate and discover potentially better solutions. Less is more.

[Jonathan Yu @yucjonathan]

This time last year, the best approach to building products with language models was to closely instruct them to keep them aligned. Now, the optimal solution is to gtfo of their way.

[Jonathan Yu @yucjonathan]

@IntuitMachine @AndrewCurran_ Per-customer cost aggregation for SaaS platforms is still a relatively unsolved problem. Combine this with stripe meters and you’ve got token-based-billing from a single API

[Jonathan Yu @yucjonathan]

aws down 😔

[Nan Yu @thenanyu]

If you drop every production table does the model get fired or do you get fired.

[Jonathan Yu @yucjonathan]

my new favorite codex mobile use case

[Jonathan Yu @yucjonathan]

@j4ykadam try aerospace, uses virtual workspaces so you don’t have to deal with the animations

[John Calzaretta @jetcalz07]

Most enterprises haven't seen a material productivity gain from AI. Many have simply introduced new forms of chaos and waste.

The problem is bolting AI onto a broken delivery model that requires constant human intervention to stay afloat.

The transformation requires reorganizing software delivery around the flow of information, scaling expertise, and maximizing individual ownership.

Essays like Block's "From Hierarchy to Intelligence" articulate where organizations are roughly headed, but provide little guidance on how others can make the transition.

At 8090, we are building Software Factory to enable this transformation for all. As our own teams built on Software Factory, a new operating model emerged. We call it the Software Factory Production System.

https://t.co/AhG6V9PgzW

[Jonathan Yu @yucjonathan]

I didn't get the tokenmaxxing pull until I started using codex. The urge to maximize your usage is too strong... can't sleep must build

[Jonathan Yu @yucjonathan]

@4484 It'll happen eventually, but strictly after they've gutted the third parties

[Jonathan Yu @yucjonathan]

@8090_Factory Super proud of what the team cooked with this one! Let us know what you think

[Jonathan Yu @yucjonathan]

@RhysSullivan Every time this CTA gets louder I lose faith that cursor 2.0 will have long term support. I wouldn’t be surprised if they force everyone into 3.0 by the end of the year

[Jonathan Yu @yucjonathan]

@kapilansh_twt AWS secrets manager with a quick script to auth and pull locally. Onboarding shouldn’t take more than one shell command

[Jonathan Yu @yucjonathan]

@dejavucoder >would appreciate if you report the observations

What an exhausted token budget does to a mf

[Jonathan Yu @yucjonathan]

@ChineseEmbinUS @XHNews The slopaganda wars have begun

[Arjun Krishna @TheOneAndArjun]

Before Software Factory, my team and I were building an AI SDLC manager for two years to do what 8090 offers. After using Software Factory I shut down our internal effort and am reorienting my team around Software Factory. This has already freed up two engineers.

The Software Factory team gets it and delivers on the core principle: holding software representation in requirements, not in code.
- @jbarseneau on Software Factory