rmonvfer

edit: reading the litestar docs, it even has a built-in event system! I spent a couple weeks building something I could use with FastAPI...

Thank you for writing this, I've been building a large backend with FastAPI for the last year or so and I've gone through all the levels of the purgatory.

I began using the standard "tutorial" style and started cringing when I saw the official template [1] place all CRUD operations in a single file (I've been doing Rails and Spring for a while before) and the way dependencies where managed... let's just say I wasn't feeling very comfortable.

Then came the SQLModel problems. The author pushes it very hard in the FastAPI docs (which imho are terrible because when I'm looking for docs I want that, documentation, not a fancy tutorial) but as an ORM (yes I know its a layer on top of SQLAlchemy) it doesn't even support polymorphic models and the community even has contributed PRs that have gone months without any review (is it even maintained anymore? I honestly can't tell).

I guess I'm the only one to blame for choosing FastAPI to build a big project but after having used it quite a lot (and also read its code because again, docs are extremely poor) I wouldn't recommend it for anything serious. Sure, if you want to build a quick CRUD then go ahead and use SQLModel and FastAPI, but keep in mind that its not built for complex applications (at least not without requiring you to write a framework on top, like I've unfortunately done).

So yeah, a big thank you to the author of this post because I will migrate to Litestar as soon as I wake up tomorrow.

Is Devin still alive?

update: I’ve tried to use lm-studio (like the author) and the tool request kept failing due to a mismatch in the prompt template. I guess they’ll fix it but seems sloppy from lm-studio not having tested this before release.

I think that’s on purpose (the chat history thing), because they actually keep the data (I’m the admin in a Workspace and even though we have Apps Activity turned off, everything still gets logged for compliance and I cannot disable it)

But yeah, it’s Google after all

I’m also very interested to know how well these models handle tool calling as I haven’t been able to make it work after playing with them for a few hours. Looks promising tho.

What a day! Models aside, the Harmony Response Format[1] also seems pretty interesting and I wonder how much of an impact it might have in performance of these models.

[1] https://github.com/openai/harmony

This is the level of epic I aspire to in life

The source code for a pre-release version got leaked a while ago (they forgot to remove the embedded source map) and if you can find it, it’s definitely worth looking into.

I was spending around $800 in Cursor and I’ve switched to Claude Code with a $200 subscription and I couldn’t be happier. The experience is way better (although tbh Claude Code is missing some critical features like being able to rollback changes (or, as Cursor calls them “checkpoints”)) but for 99% of my “Vibe Coding”, it’s just great. I usually run 2 to 4 parallel sessions using git worktrees and the speed is absolutely crazy. Of course not everything is perfect and I still have to check most of the code but if you create a good enough set of “memories” (Claude Code’s version of .cursorrules) it gets stuff right almost all the time.

I was a Plus subscriber and upgraded to Pro just to test Codex, and at least in my experience, it’s been pretty underwhelming.

First, I don’t think they got the UX quite right yet. Having to wait for an undefined amount of time before getting a result is definitely not the best, although the async nature of Codex seems to alleviate this issue (that is, being able to run multiple tasks at once).

Another thing that bugs me is having to define an environment for the tool to be useful. This is very problematic because AFAIK, you can’t spin up containers that might be needed in tests, severely limiting its usefulness. I guess this will eventually change, but the fact that it’s also completely isolated from the internet seems limiting, as one of the reasons o3 is so powerful in ChatGPT is because it can autonomously research using the web to find updated information on whatever you need.

For comparison, I also use Claude a lot, and I’ve found it to work really well to find obscure bugs in a somewhat complex React application by creating a project and adding the GitHub repo as a source. What this allows me is to have a very short wait time, and the difference with Codex is just night and day. Gemini also allows you to do this now, and it works very well because of its massive context window.

All that being said, I do understand where OpenAI is going with this. I guess they want to achieve something like a real coworker (they even say that in their promotional videos for Codex) because you are supposed to give tasks to Codex and wait until it’s done, like a real human, but again, IMHO, it’s too “pull-request-focused”

I guess I’ll be downgrading to Plus again and wait a little to see where this ends up.

Just for the record, I use PostHog in my Europe-based startup. They have a EU region so it’s not a problem for us.

I very much relate to this feeling (and I haven’t finished the rust book either!). In my case, I do use AI a lot (especially o3 and Sonnet 3.7), not to write code but to help me understand things that would’ve taken me a week, in a matter of hours (the conversational aspect is a game changer for me).

Could you provide a little bit of detail about how you did that? Just in case…

As many have noted, not everyone aspires to wealth; a lot of people are happy with having just enough to live comfortably and go on vacation a couple times a year.

The way I see it, (though I could be mistaken), wealth empowers individuals to engage in pursuits they value without external concerns.

Exceptionally talented professionals (e.g., nurses, doctors, scientists, lawyers) are quite frequently rendered incapable of addressing issues they are passionate about due to financial constraints or bureaucratic indifference. Essentially, financial resources enable one to transition from being merely aspirational to being influential (contingent, of course, on the specific situation). Consider activism: many people strive to support their communities, but often, funding for such initiatives is lacking.

How is this any different from LlamaIndex [1]?

[1] https://www.llamaindex.ai

I strongly recommend the Microsoft for Startups [1] program. They’ve been providing us with hundreds of thousands of Azure credits and other Microsoft technologies (even OpenAI credits!) They have a tiered system and you need to apply using LinkedIn but hey, it’s a lot of stuff after all and they’ve been adding stuff consistently for the past year or so.

[1] https://www.microsoft.com/en-us/startups

You should probably look into Guidance [1](previously Microsoft Guidance but looks like it’s been separated from their main organization), which is a language for controlling the output of LLMs (so you can, among many other things, output JSON in a deterministic way)

[1]: https://github.com/guidance-ai/guidance

I've had a similar experience with LangChain. Initially, I was really impressed with it while developing an MVP, but as soon as I needed to add complexity or specific features, things started to go downhill.

A significant chunk of my time was spent navigating through LangChain's codebase, trying to make sense of missing or outdated documentation. Plus, there's a significant disparity between the features of the JS/TS and Python versions. Identifying which version supports which features can be a real challenge.

I don't want to sound overly critical, as I'm currently using it in production. However, I had to modify it so extensively that it barely resembles the original project anymore. Looking back, I can't help but think that adopting this library might have been a misstep. Perhaps we should have taken the time to create our own.

This, 100%. I couldn’t agree more. I’m currently a Software Engineering undergrad in Europe and, at least by my personal experience, the whole education system just focuses on preparing people to pass an exam, not to properly learn a subject.

IMHO, the problem starts when the students get used to memorizing and repeating sample exercises which involve 0 creativity and then graduate and get an actual job.

It may sound hilarious at first but also serve as a good (and extremely simple) learning model.