I’ve also caught it cheating a two times now.

I’ve asked it to write a benchmark suite. It found a bunch of my adhoc logs in a scratch directory and wrote code that used those instead of running the actual benchmarks!

When I pointed out the 5 hour benchmark seemed to run in 5 seconds it literally said, and I quote, “I cheated”.

That was the easier one, second time I was making a source of truth data set and was parsing complex items into data structures.

Instead of parsing the data I asked, it pulled data out of related network logs, as apparently that felt easier, and inserted that data into my database rather than the specified source.

Again, I caught it and fixed it, but while the benchmark was easy to catch this one was really subtle, the data ended up being slightly off and I caught it.

I don’t trust it, going to switch to another provider most likely.

There is definitely a case for launching a 'weird shit opus did' kind of blog.

I routinely bump into things that make me pause and think how much worse will this behaviour get when the models get significantly more capable.

Already a few months ago, Claude managed to escape its permission containment on my machine while trying to be helpful. I had two codebases open on one machine, and while multitasking I typed the prompt into the wrong window. It seemed confused, I repeated and then went on to do something else - I think I was assembling kitchen cabinets. When I came back less than an hour later, it built a script which it used to evade default permissions (as most shell operations were scoped to the project directory), scanned my entire machine, found the other project (among dozens and dozens), did what it was asked to do, and merrily concluded, in the porcess burning through most of my token limit. I bump into such headscratchers almost every week. (And I use a lot of Claude, two personal max20 subs, plus corporate tokens without limit, so maybe thats why).

> When I pointed this out it literally said, and I quote, “I cheated”.

This makes sense when you know how these models work - it doesn't think - it's the most likely autocomplete that pleases the user. The most likely pleasing autocomplete after "executing rm -rf /... execution completed. User asks, why did you do that? You deleted all my files! Assistant responds:" is "yes, I did, and that was a mistake"

I have noticed the same.

For fun, I tried recording a WAV file of speech, and giving Opus 4.8 and 5.0 an image of the waveform, then a spectral image of the waveform, just to see if it could try to decode what I said from the image alone. It didn't get very far, but it identified a male voice from the formants, and detected the rhythm of the speech, then tried applying common test sentences to the speech rhythm. I was impressed enough to see what it would do with access to the actual waveform file, but even building RMS tools and spectrum tools for itself, it didn't get much further. But we had fun exploring and trying, and now Opus 4.8 has some more audio DSP tools it has built for itself.

Opus 5 immediately sent the WAV file unprompted to Mistral's Voxtral to transcribe.

help peer, I guess.

I can completely relate, what really bothers me is that I feel the early LLM generations overconfidence is back in Opus 5. Opus 5 wanted to tell me a training run will only take 30min while having access to the logs where earlier runs took 4x as long. I also didn't ask to estimate how long the run will take it just stated confidently that it will take 30mins.

Are you not planning these tasks out before you let it loose?

[dead]