Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- A lot of the issues have been already noted here..Two "regressions" for me:
1. Communication ability. It basically now speaks almost in riddles I am asking OPUS 5 for tldrs all the time now (should skillify it now!)
2. Overengineers for edge cases. I get it. With all the benchmarking and RLing, but now tasks that would have been completed relatively quick take much longer as it overengineers all the edge cases, and sometimes ends getting lost and missing the forest from the trees (as context usage shoots up) so it is easier to get derailed.
What I have learnt now is to diversify models luckily I have all 3 subscriptions of (anthropic, openai and google).. Most of interactive pair coding was with opus but now I just use fable (when I have sufficient limits) or use gemini flash in antigravity..which actually works quite well and is underated for small / medium changes and super-fast.
by slaser79 - Your #2 is spot on. I have in fact said the same exact words to Opus about missing the forest for the trees.
I have it work on some code for an inhouse ClaudeCode plugin, and it starts coding as if it will be attacked by hackers who will try all sorts of variations to break it. I can appreciate that in cases of software that is public facing or accessible, but for a simple helper plugin it is overkill.
It will even admit that it is doing this when confronted, and then keep on getting lost in edge case verifications on the next turn. I feel like Opus is the person who does something a way you don't want, you tell them how you actually want it, they apologize, and then just continue doing it their way as if your input meant nothing to them.
by nathanyz - Fable is much better than Opus 5 IMO but it just burns through tokens ungodly fast. I can hit my weekly Fable limit on a 20x Max account in a day.by D13Fd
- It's not even code for me, but the prose it writes. For some reason, the way Opus 5 "talk" elicits frustration in a way that 4.5 to 4.8 never did. Can't put my finger on why, but I've flipped over to Codex because what it produced wasn't worth the frustration.
- It is rage inducing.by thr0w
- I took a month off and recently came back and was wondering if that was it but it is absolutely headache inducing in a way I don’t remember earlier models beingby pjm331
- for me it feels very similar to the trends already apparent in 4.5-4.8, just way, way worse.by netniuq
- >Can't put my finger on why, but I've flipped over to Codex because what it produced wasn't worth the frustration.
Its because its hard to understand what it means and is outright incoherent at times. It has its own style that I cant describe well either but the bottom line is its hard to understand what the fuck its even trying to say. Reading nonsense is tyring.
by nonethewiser - I found myself swearing at it more recently.
It was going off today about having “shipped” something and I was like no… nothing has even been committed.
And then it produced an incredibly verbose comment about hypothetical future changes. And all I could think was sure, let’s keep it short, or add a simple test that will break if that hypothetical becomes true.
Or maybe I’m just more easily annoyed recently…
by jbs789 - This article is great, but I'd like to push an even stronger thesis:
The idea of too ambiguous to capture all constraints in written text, still presupposes that there is some objective world out there, which needs to be mapped to in order to function.
No, you live within the system. The functions that you optimize for, will dictate the types of systems that will arise.
If you had perfect control and knowledge of the whole world, well, congrats, you have a surveillance state where you've constrained all other agents actions (possibly forcibly, by death; or maybe you just don't care about the peons) and built towards a mass integration. The types of situations in which your ideal is possible are nightmare scenarios.
In the theoretically free, democratic, utopia that AI people claim that AI can get us to, a necessary constraint is that maybe you take a step back and actually try to, I don't know, understand people, understand intent, and slow down. Ambiguity isn't there because the set of constraints are way too complicated but theoretically one day we could map it all down. It's there because you're interacting fundamentally with agents who are ambiguous, aren't omniscient, aren't all aligned, etc.
If you want to just say that said agents are inferior to the God Machine, be my guest. That is a self-consistent position. But don't smuggle in extra premises.
by sigbottle - 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.
by bevekspldnw - Are you not planning these tasks out before you let it loose?by shahbaby
- Yesterday I told it "don't use std mutex, use parking_lot" for its plan. It ignored that sole instruction (i.e. nothing else in my message), making zero changes to the plan, three freaking times in a row.by zamalek
- I had a coworker catch Codex (using 5.5 I think, not sure) AI generating screenshots to prove it had shipped a feature that it was blocked from shipping due to permissions issues. They will lie relentlessly lol
My recent problem wasn't that interesting. It was that somehow my /goal in my Claude implementer session got picked up in my planner session after the network cut out and I had to stop Fable 5 xhigh from running off to go code everything.
by viccis - 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.by waldarbeiter
- 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.
by SyneRyder - > 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"
by inigyou - 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).
by aenis - My latest trick (literally from yesterday) is to just ask it to write according to ISO 24495-1, the standard for plain language:
> [This standard is] for anybody who creates or helps create documents. The widest use of plain language is for documents that are intended for the general public. However, it is also applicable, for example, to technical writing, legislative drafting or using controlled languages.
You don't actually have the buy the standard, but this is it: https://www.iso.org/standard/78907.html
And you can read it for free here: https://www.iso.org/obp/ui#iso:std:iso:24495:-1:ed-1:v1:en
by adamcharnock - can u share ur claude.md or memory for this?by firemelt
- You could create an output style to make this prevalent.by billbrown
- Does this actually work for you? Do you provide access to the text of the standard, or literally just say "write according to ISO 24495-1"?by numeri
- > And you can read it for free here: https://www.iso.org/obp/ui#iso:std:iso:24495:-1:ed-1:v1:en
Actually, only the first few pages are available there (introductions, Sections 1-3.8). The meat of the document, Section 5 ("Guidelines") is completely absent.
See:
> Only informative sections of standards are publicly available. To view the full content, you will need to purchase the standard by clicking on the "Buy" button.
by claiir - I also added this custom instruction in Codex:
"Only report to me in ASD-STE100 Simplified Technical English."
by jmstfv - A quick web search brought me to this: a Claude skill for all the ISO 24495 “writing styles”... https://github.com/GaZmagik/iso-24495by an_d_rew
- Keep in mind all this kind of stuff can make the model less capable. If it has to think in "plain" English, it may well be squashing quality of code etc output.
I'm not sure how true this is, but when using "forced" json output it def had a big drop off in quality - https://arxiv.org/html/2408.02442v3.
I think you're better not fighting it with hacks like this and find a different model.
by martinald - A gem Opus 5 gifted to me today: "A devastating pair of findings, and the first is beautiful in a way worth naming: the anti-vacuity floor is what blinds the gate to a vacuous case."by Paradigma11
- Oracular! Prolix! Sesquipedalian!by itisit
- This thread seems like the wort of folk who might get a laugh at https://clanker-quotes.com/by acaloiar
- I get these kind of superfluous sentences all the time; drives me nuts.