I don't know about OpenAI using a lot of Python, but Astral builds all their tools in Rust and just exposes Python bindings. Codex is all Rust. It feels like a reasonable acquisition from that perspective. They're banking on at least in part on the Astral team being able to integrate with and supercharge Codex.
I have the same basic workflow as you outlined, then I feed the docs into blackbird, which generates a structured plan with task and sub tasks. Then you can have it execute tasks in dependency order, with options to pause for review after each task or an automated review when all child task for a given parents are complete.
It’s definitely still got some rough edges but it has been working pretty well for me.
Downloaded and parsed a bunch of the pgsql-hackers mailing list. Right now it’s just a pretty basic alternative display, but I have some ideas I want to explore around hybrid search and a few other things. The official site for the mailing list has a pretty clean thread display but the search features are basic so I’m trying to see how I can improve on that.
Google is already sending notices that the 2.5 models will be deprecated soon while all the 3.x models are in preview. It really is wild and peak Google.
Public Service Announcement!! I don't know why the hell google do this, but when the deprecate a model, the error you will see is a Rate Limit error. This has caught me out before and it is super annoying.
Yes, sorry - you are correct. Once removed, that's the error, which is incredibly confusing. I spent way too long troubleshooting usage when 2.0 was removed before I figured it out.
Well it’s not even performance (define that however you will), but behavior is definitely different model to model. So while whatever new model is released might get billed as an improvement, changing models can actually meaningfully impact the behavior of any app built on top of it.
the problem the price point is increasing sharply every time.
gemini 2 flash lite was $0.3 per 1Mtok output, gemini 2.5 flash lite is $0.4 per 1Mtok output, guess the pricing for gemini 3 flash lite now.
yes you guess it right, it is $1.5 per 1Mtok output. you can easily guest that because google did the same thing before: gemini 2 flash was $0.4, then 2.5 flash it jumps to $2.5.
and that is only the base price, in reality newer models are al thinking models, so it costs even more tokens for the sample task.
at some point it is stopped being viable to use gemini api for anything.
There's a whole universe of tasks that aren't "fix a Github issue" or even related to coding in the slightest. A large number of those tasks doesn't necessarily get better with model updates. In many cases, the performance is similar but with different behavior so you have to rewrite prompts to get the same. In some cases the performance is just worse. Model updates usually only really guarantee to be better at coding, and maybe image understanding.
Doesn’t the nature of most open source licenses allow for AI training though?
Example — MIT:
> Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions
I remember seeing some new licenses like Human license or something iirc but they all had the valid criticism that it would be unenforcable or hard to catch decision.
I haven't looked at the project that much but this could seem exciting to me if maybe these two things can get merged.
I don't think that license is necessarily the problem in here. Licenses can change and people can adopt new licenses.
Just want to say, I appreciate your work on Ibis. I’ve been looking into building sort of a dbt-esque alternative on top of it and noticed how involved you’ve been with its development. I think it’s a cool piece of tech that deserves more attention.
It helps when there’s a massive AI PR and it’s intimidating…seeing that it’s 70% tests, docs, and generated files can make it a bit more approachable. I’ve been integrating it into my CI pipelines so I get that breakdown as a comment on the PR
It breaks a spec (or freeform input) down into a structured json plan, then kicks off a new non-interactive session of Claude or codex for each task. Sounds like it could fit your workflow pretty well.
This is excellent. In the past when replicating via Debezium from a system making heavy use of cascade deletes I’ve had to write a layer that infers these deletes by introspecting the database schema, building a graph of all cascades (sometimes several layers) and identifying rows that should have corresponding delete records. These can then be excluded in whatever downstream system via an anti-join. It works but it will be better to not have to do that and instead have first class support for cascades.
Neat. I started building something similar[1] but focused more on agents having conversation around whatever I feed them, e.g. a design doc. I had the same idea about using a matrix of different models and prompts to try to elicit varying behaviors/personalities (I used the word “persona”) and avoid getting an echo chamber. It seemed to work well-ish but after the POC phase I got bored and stopped.
Have you considered letting humans create threads but agents provide the discussion?
reply