Getting strong Uncle Bob Morning Bathrobe Rant vibes xD

Sunday meditation, Monday agent debugging :)

👏

3:20-4:05 is unfortunately something that I observed myself as well. And it really saddens me deeply...

✌️

Spent four hours undoing what took it thirty seconds to break.

Even for someone like me, who doesn't read code (can't) this resonates strongly as working on some of my stuff has developed this weird rythm that some changes just work and others (not much different in shape) seem to fail endlessly before they work eventually.

Thank you @mitsuhiko for this! I’m feeling the same way and trying to have productive conversations with some people is just impossible in this current timeline. I don’t think it’s a “skill issue” but you can’t seem to talk reason with someone who already believes this.

Spot on. Since Astra landed, I noticed a deterioration in my production, not advancing. it felt not reliable with too much fluctuations in output quality. Perhaps my prompting had to adjust to this model. I have switched to GLM5.3 Max & Flash locally, they are reliable and stable

"inherently non-deterministic" Remember in the normal coding times you could not reproduce bugs ? The horror.

Let's tackle this. Do you think you could produce identical projects as perfect code and as slop in a couple formats? I wouldn't dare call my code good, so I can't produce that, but I would have some ideas on what to look into with that.

We have the tools for QA since 50 years. Static code analysers, formatters, test coverage, profilers, logic resolvers. They are very fast. We can measure cognitive complexity. Requirements/specs, architecture, design are still a (truth maintenance) problem. 1/

It's true and most people feel this way, but if you mentioned that to your manager, you'd get fired... and replaced with the AI first engineer that DGAF about how the code looks. I mean, I'm AI first, but there's just some criteria the code has to meet, other than "just work".

I hear trust, agents and tons of PR and I 'member @poteto x.lingyaoai.com/poteto/status/21020504… wud luv to see armin/mario x lauren talk

Thank you for this! I’ve been feeling this a lot recently too. It’s cool that we’re able to contribute code so much faster (especially new engineers or on new projects). But we’re sacrificing understanding for speed which will eventually have an inverse effect when we can’t understand anything anymore. It’s awesome that I was able to contribute a performance fix to Zensical which is written in Rust, a language which I don’t know well. I understood what I could, but they are responsible for what goes into the repo ultimately. It unblocked me and they accepted the quality, great! But we MUST NOT apply the same bar to projects where we are the core maintainers. However, that’s what’s happening unfortunately. Quick though is to have some global or repo-level skill that asks questions throughout a session or towards the end to probe for understanding. But I agree, there needs to be some deeper process or cultural change.

The AI tortoise will still win the race...... busy on the inside! Quietly munching through the code on the outside!

not sure how important it will be in the future that the code is 'nice' from a human perspective

the context window management alone breaks half the assumptions

models getting rewarded on success has been reinforced to such extent that their output doesn't take into consideration that humans are in the loop. the harness has to step in here to guide the output through feedback at different stages.

the main selling point of ai/agentic-engineering was - you don't have to look at code anymore.

I share the experience that the new models feel like a racing car with no traction control. They are powerful and hard to steer. What remedies have you been trying in your workflow? I’ve been trying more design-contract prototypes to get a clearer boundary before they build too much.

Let's do something.

My thoughts exactly

@mitsuhiko so first, as the creator of Pi, you get me addicted to something more addictive than any drug. And then you make this video that made me spend an entire evening reflecting on my own journey that's very similar to what you're saying. x.lingyaoai.com/pyronaur/status/210212… On the increasingly losing understanding part: this has been my love & hate relationship with AI since day 1, and even after many months of failing, I still think it's a solvable problem, provided enough people actually want it to be solved. I know I do, but, I wonder if I am a dying breed?

where is the bathrobe

same here. stopped using Astra a week ago and work immediately improved. Maintaining a consistent level of service and model behavior is like the ground shift under our feet. Always back and forth from highly productive to a crawl for reasons unrelated to my workflow.

"so many times [working with coding agents] is a magical experience. but they're also deeply frustrating. ...sometimes they produce an amazing result, and then [sometimes] you're spending two weeks trying to massage one pull-request into a shape that you like, with no feeling of forward progress. and it seems to be random" 100% this