sz4kerto

> The ones who have seen obvious degradation in work quality and stability, but who have tried to ignore or hide it. They now have to face down CFO’s and executive management who see zero returns on the current spend and are being asked to approve extreme budget increases.

Hm, there must be businesses like this, but I presume there are many who experience huge returns on current spend - and even 5x costs would be fine, but currently AI prices (given fixed 'intelligence') are dropping like a stone. I can't imagine that the same level of intelligence is going to be more expensive in 12 months.

It's just normal that there's a huge variation of how businesses and individuals adapt to new circumstances, this has been the case forever.

It's not self-evident what we need. We currently ended up with this whole free-market capitalism, which is super distorted of course, but still it kinda says that 'we should let people do whatever they want with their money unless they harm others'. "I want clean water. I want a stable climate. I want bees to survive." Yes, so we all. But we set up a system where people can vote with their wallets.

> If you even review PRs still: when was the last time you didn't just skip over tests?

The opposite. Reviewing PRs is now about just reviewing the tests, as the code is very likely to be fine if tests are relevant and they're passing.

The main enabler of agentic coding is heavy end-to-end/characterisation testing.

We can ask AI to do very large-scale refactors, for example, because if tests pass, it's 99% that everything is fine.

GLM 5.3 Flash has been a relevation for me. It's practically impossible to spend more than $5-$10 per day if you're only working on a single project -- but $10 is a full-day of continuous churn. First I was super sceptical about it, and always used Fable to instruct it, but now I realised that even with complex coding, it's reasonably good.

> I do sort of suspect a fan thermal control curve is a PID response curve written out in long hand but don't really have the math to prove it.

No, a fan response curve is kinda-sorta a P controller. It does not take into account 1) how quickly the temperature is rising or dropping (D) 2) the time passed since the system has drifted from the target temp (I).

"In my previous blog, I revealed that Oura data is not end-to-end encrypted. That means that an Oura user's health data can be unscrambled at certain points as it travels from a person's ring, through their phone app, over the internet, and as it lands on Oura's servers."

Very strange -- it seems to be conflating end-to-end encryption with encryption-in-transit.

We need sandboxing for any agent, so we run it within Docker - so we use CLI.

> And then there's explorative testing, where I always found a good QA invaluable.

Yes, I agree. We do this too. Findings are followed by a post-mortem-like process: - fix the problem - produce an automated test - evaluate why the feature wasn't autotested properly

"Before I weigh in further, I’d like to make sure you’re familiar with the testing pyramid."

The testing pyramid is a par excellance SWE kool-aid. Someone wrote a logically-sounding blogpost about it many years ago and then people started regurgitating it without any empirical evidence behind it.

Many of us have realised that you need a "testing hourglass", not a "testing pyramid". Unit tests are universally considered useful, there's not much debate about it (also they're cheap). Integration tests are expensive and, in most cases, have very limited use. UI and API tests are extremely useful because they are testing whether the system behaves as we expect it to behave.

E.g. for a specific system of ours we have ~30k unit tests and ~10k UI/API tests. UI and API tests are effectively the living, valid documentation of how the system behaves. Those tests are what prevent the system becoming 'legacy'. UI and API tests are what enable large-scale refactors without breaking stuff.

Isolated QA should not exist because anything a QA engineer can do manually can be automated.

So prices will need to increase -- if it makes a senior engineer 10x more productive then coding assistants could easily cost 20x-100x more then what they cost today. Same for video generation.

Recently I've been using AI for some stuff that service providers don't want it to be used for (specifically: medical diagnosis). I found that Grok (4.1) is superior to most of the others when it comes to this, because it doesn't go out of its way to support my own hypotheses.

I believe that syncophancy and guardrails will be major differentiators between LLM services, and the ones with less of those will always have a fan base.