M0r13n

> often preferring downstream fixes/hacks/workarounds when upstream fixes would be better.

It sometimes feels like the model is too lazy to do the actual reasonable fix and implements a _hacky_ fix downstream. And to be honest: I feel that!

This weekend I had some time with Opus 5.5 and GLM 5.3. I went over some findings in a library of Claude Mythos 5.1. These were minor, low-severity bug fixes and Mythos provided a PR to address them. Its PR was functionally correct. I still spent ~5 hours looping over them until I was really happy, even though a fully working PR already existed.

Some of the differences were:

- Mythos recommended salvaging damaged labels. I pushed back: data whose structure is broken shouldn’t be trusted

- My simpler more readable test caught a regression that Mythos’s more elaborate test missed

- Mythos allowed invalid values, citing real-world data that I could not reproduce, while I decided to follow the spec

- Mythos kept the old error types. I made every kind of malformed input raise the same error, and accepted a small, documented breaking change for anyone who relied on the old ones

Overall, I much preferred the final version of my PR to the initial draft Mythos provided. However, getting there involved quite a bit of time messing around with the code, as well as some back and forth with Opus (which was valuable).

This is a recent example that matches my overall experience working with current SOTA models (the ones mere mortals like me have access to).

TL;DR: I don’t think AI already has better judgment than (most) senior people. Obviously, that might change with future models. But that’s a hypothetical scenario. It might as well not change anytime soon.

Ask ChatGPT:

> Generate a realistic image of a bad fitting suit without looking obviously fake

Create a new chat and attach the image generated previously:

> Please rate my suit

You will receive an answer that reads something like this:

>7.5/10 — clean, professional, and well coordinated.

Yeah, I don't think superintelligence is imminent.

This seems to assume more code shipped equals more work done. I am still not convinced that this is necessarily the case. Sometimes it is. Sometimes it is not. You mentioned that reviews are now the bottleneck and that your familiarity with the code has decreased. That tradeoff might eat into the parallelization gains over time.

Exactly. I follow a similar workflow. I am still writing code by hand. Not all code. Parts are generated by Opus.

> look at my postprandial response to dosing for the past week and make ratio suggestions for breakfast, lunch, and dinner

I'm not so sure about that. A patient absolutely must critically evaluate the LLM's suggestions. A naive user risks severe complications. A user with that kind of competence, however, doesn't need an LLM for such trivial adjustments - they're obvious

> But for kids or the elderly, I see a use case.

These are exactly the kinds of people who should NOT base their diabetes management on the "suggestions" of LLMs! There is the real risk that such users lack the competence or judgement to critically assess the convincing-sounding output of an LLM.

I don't think that LLMs are trustworthy companions in managing a complex metabolic disease like diabetes - especially if you deviate (ever so slightly) from the norm (very lean, very active, strict diet, etc.)!

I'm a T1D myself and like to experiment with ChatGPT (or Opus). My experiences are mixed

LLMs are overly cautious when it comes to correcting with insulin. They regularly advise against correcting before going to bed, even if this means that my blood glucose remains above 140 mg/dl for the whole night.

I am following a low to medium carb diet (<100g a day). ChatGPT always nudges me to consume more carbohydrates, even though I have a TIR of 90% (70-150 mg/dl). Why would I change my diet if it currently works very well for me? Still, most LLMs seem to favor carbs for some reason.

I am using Fiasp as my fast acting insulin. Typically, I inject around 1 to 4 IUs of Fiasp. Its glucose-lowering effect typically lasts for roughly 2-3 hours. Therefore, I know that it is safe to re-inject after three hours without risking insulin stacking. But ChatGPT regularly advises against that and wants me to wait another 1-2 hours.

I am not against automating diabetes management. In fact, I really appreciate projects that help with that. But I don't consider LLMs to be helpful in this regard. Their combination of training data bias, liability aversion, lack of context, and one-size-fits-all thinking disqualifies them from such tasks.

> Also your in winter are you running the heater constantly?

Nope. But last winter it got really cold (< -10°C). Also the Honda was aggressively heating the battery

I agree. Though I replaced it with a car with a capacity of 40 kW and better overall efficiency. You're right, however, that I could have bought a used VW ID.3 or similar. But I can't stand how these cars look or feel. I'm also not complaining. I really like my car. I just wanted to share my experience :)

yes!

Sadly, I don't own a wall box. So I'm dependent on public infrastructure. If I could charge at home I'd be even happier with my car.

It depends where you are. I live in a somewhat rural area in Northern Germany. There are some chargers, but many of them are out of service like the one next door. I've never stumbled upon a free charger though.

I've had a rough time with the last two EVs I've owned. I bought a Honda e in 2020 because of its retro-future charm, but ended up being disappointed by several things:

- the range was miserable

- the software quality was bad

- no OTA updates ever (despite Honda's promises)

- slow charging

- poor public charging infrastructure in Germany

I should have known that a 35 kW battery wouldn't deliver great range or charging speed. But I didn't fully appreciate how limiting it would be.

Last year, I bought a new Mini Cooper e. Larger battery. Better software. BMW's quality actually delivered this time. The car feels objectively nice. The software is polished. There are updates. Few bugs. But the range still leaves something to be desired. In summer it's okay. During winter 30-40% of the range just melts away.

Public charging in Northern Germany still sucks:

- too few public chargers

- chargers are often broken or out of service

- pricing is intransparent

Municipal utility companies ("Stadtwerke") seem especially bad at maintaining their charger fleets. Every second charger that I want to use is out of service. The one next to my apartment has been labeled as "defective" for a couple of weeks now. Nobody seems to care...

I still like (love during summer) my car. It's a cool car. It feels luxurious. It's comfortable. It's fun to tear around corners. It's still compact enough to maneuver through the city. And it looks cool. But it also costs 40-50k EUR and only has limited range. And public charging really needs to improve.

My comment was not intended to be any criticism or to downplay the performance - quite the opposite :)